ABSTRACT. Lecture Overview:

Published literature commonly includes statistical tests to justify the validity of the findings. Recent meta-analyses depict that an alarming proportion of studies cannot be independently replicated. Hence, the ability to understand and interpret statistical methods and results is a critical but frequently overlooked/underdeveloped skill. In this lecture, via the explanation of basic inferential statistics, we will discuss statistical mistakes and how to identify and avoid them. Common problems include incorrect usage of p-values, misinterpreting statistical power, and conflating statistical vs clinical significance. Avoiding these requires an understanding of the domain in which statistical tests are useful and being aware of their limitations. This lecture is derived from an illuminating AAPM 2018 session.

Learning Objectives:

1. Appreciate the gravity of frequent statistical problems in publications.
2. Identify pervasive telltale markers of potential problems in various statistical tests and methodologies.
3. Learn methods for correctly implementing relevant statistical analyses.

About the speaker:

Avishek Chatterjee obtained his undergraduate degree in Physics, Mathematics, and Astronomy from Wesleyan University (CT, USA), graduating in 2006. He moved to Cornell University (NY, USA) for his PhD in particle physics. During his stay at CERN as part of his research, he developed a keen interest in data science, being at the forefront of the Big Data revolution by virtue of working at the Large Hadron Collider, the largest experiment built by man. After finishing his PhD, he was a post-doctoral researcher in the fields of accelerator physics (Cornell), and particle physics (University of Geneva, Switzerland). Since 2016, he has been a researcher at McGill University (Montreal, Canada), where he has leveraged his data analysis experience to work on Medical Physics projects where patient data in the context of radiology and radiation oncology can help improve the patient journey. His current work focuses on traditional radiomics, Deep Learning, and the establishment of a robust theoretical and evidentiary basis for the statistical analyses performed in these areas to ensure this nascent domain can achieve successful clinical translation.

ABSTRACT. Lecture overview:

Patient-reported outcomes (PROs) are outcomes that come directly from the patient without physician interpretation. They are currently a hot topic in clinical oncology and oncology research and are being adopted into radiation oncology practice worldwide. They have been shown to improve clinical care, patient survival (when acted upon in a timely manner) and to inform research data. Despite their promise, however, the routine collection of PROs is complex, requiring validated patient-reported outcome measures, suitable workflows, willing clinicians, and dedicated databases. In this lecture, we will define PROs and associated terms and we will examine the benefits and complexities of PRO adoption into clinical care and research.

Learning objectives:

1. Define and understand PROs, e-PROs, PROMs, PREMs, and PRO instruments.
2. Understand PRO validation as well as the need and difficulty of standardization.
3. Be able to explain the complexities of PRO collection in clinical practice

About the speaker:

John Kildea is a medical physicist at the McGill University Health Centre in Montreal and assistant professor in the Department of Oncology at McGill University.

Dr. Kildea’s background is in astrophysics, having earned his PhD in gamma-ray astronomy at University College Dublin in Ireland in 2003. Before moving into medical physics in 2008, he worked as a postdoc in high-energy astrophysics at McGill University and at the Harvard-Smithsonian Center for Astrophysics. His transition into medical physics started with MSc studies at McGill University in 2008 and finished in 2012 at the end of a two-year residency at the McGill University Health Centre.

As a medical physicist, Dr. Kildea’s research interests lie in neutron dosimetry and quality and safety in radiation oncology, with a particular focus on the use of informatics for quality improvement and comparative-effectiveness research. Dr. Kildea is a strong believer in full patient involvement in healthcare and healthcare research as a catalyst for continuous improvement. At McGill, Dr. Kildea is co-leading the Opal Health Informatics Group with a physician colleague and a patient. The collaboration has led to development of a patient portal smartphone app to provide patients with access to their data and to collect patient-reported outcomes.

ABSTRACT. The overall goal of radiotherapy has remained remarkably consistent for the last 100 years: to eliminate targetable disease (imageable or occult suspected), maximizing the chance of long-term survival, without injury to the patient that causes long term pain, discomfort, lack of mobility, lack of hearing, lack of taste, or any other disability. Another key pillar of optimal, outcomes-driven treatment planning requires methods of planning that respect the physician (or patient) priorities among endpoints, whether explicitly stated or not, that should impact planning tradeoffs. Fortunately, stating priorities for tradeoffs actually makes the planning problem more reliable and automated, reaching a more definitive 'prioritized Pareto' state, where lower priority objectives cannot be improved without sacrificing higher priority objectives. Prioritized planning is also a powerful framework for incorporating robustness guarantees. Thus, progress toward truly outcomes-driven treatment planning has been substantial, and it is possible to see an emerging paradigm where outcomes-driven treatment planning is much more effectively and commonly-practiced on a wide basis. This talk with briefly review what is available and what is missing in dose-volume and genomic prediction models, as well as prioritized planning methods capable of incorporating these models to drive individualized treatment plans.

ABSTRACT. Introduction: In the adaptive dosimetry framework in our institution, the treatment plan is automatically recalculated on each acquired CBCT. Together with automated deformable contour propagation, this gives us the possibility to estimate changes in NTCP during treatment in a fully automated fashion. In this study we investigated the feasibility of using this framework non-small cell lung cancer (NSCLC) patients requiring adaptive interventions. Methods: For 36 NSCLC patients, 14 of which had a plan adaption during treatment, treatment plans were recalculated on all daily CBCTs, and contours were deformed to the CBCT using Mirada Workflowbox®. Subsequently, NTCP values were calculated using previously validated toxicity models for A) dyspneu grade 3, B) 12 months mortality, C) myelopathy, and maximum dose of the mediastinal envelope was used as a dose threshold [1][2]. NTCP values are a weighted sum of all fractions up till the moment of adaptation taking into account the actual delivered dose per fraction. Results and discussion: A realistic clinical trigger level of 2 Gy for mediastinal Dmax, 4% for dyspnea grade 3, 2% for mortality and myelopathy, results in 6% of all fractions being flagged for adaptation (13% in ART group, 2% in non-ART group). On a patient level, patients that had one ore more fractions flagged were: 57% in ART group and 18% in non-ART group. Conclusion: Fully automated NTCP-triggered adaptation has shown to be feasible and may overcome the subjective interpretation of changing anatomy during treatment. For clinical impact, a further prospective evaluation will be necessary to clinically validate these results.

[1] G. Defraene, et al, “P3.14-011 Mean Heart Dose Is an Independent Risk Factor for Early Mortality After Chemoradiotherapy Treatment for Lung Cancer,” J. Thorac. Oncol., 12(11),S2334–S2335 [2] J. P. Kirkpatrick et al, “Radiation Dose-Volume Effects in the Spinal Cord,” Int. J. Radiat. Oncol. Biol. Phys., 76(3)

ABSTRACT. Quantification of local metabolic tumor volume (MTV) changes may allow more accurate evaluation of tumor response to chemotherapy or chemo-radiotherapy (CRT).

61 patients with locally advanced esophageal cancer underwent baseline, post-induction chemotherapy (follow-up) and post-CRT PET/CT. A grayscale blended PET/CT image was generated as a weighted sum of normalized PET and CT images. Follow-up blended PET/CT image was deformably registered to the baseline blended PET/CT image with tumor motion correction and physiologic uptake correction. Jacobian map was computed from the resulting transformation to quantify the local MTV changes. We evaluated registration accuracy by comparing the net MTV changes (∆MTV) calculated by Jacobian integral vs. by semi-automatic MTV segmentation. We then extracted radiomic features from the Jacobian map and PET/CT images to predict pathologic complete response (pCR). A multivariate prediction model was constructed using a Random Forest (RF) classifier coupled with a least absolute shrinkage and selection operator (LASSO) for feature selection. 5-fold cross-validation was repeated ten times to assess the accuracy of the RF-LASSO model.

Tumor motion correction and physiologic uptake correction improved the accuracy of registration and Jacobian map. Qualitatively, Jacobian map by blended PET/CT registration showed smoother local MTV changes than PET-PET and CT-CT registrations. Quantitatively, ∆MTV calculated by Jacobian integral using blended PET/CT registration (-42.0%) was closer to ∆MTV measured by the semi-automatic segmentation (-50.0%) than by PET-PET (-29.4%) and CT-CT (-8.0%) registrations. The RF-LASSO model achieved the highest accuracy of AUC=0.80 in predicting pCR with a single blended PET/CT Jacobian texture feature - Mean of Cluster Shade (MCS). Higher MCS indicated more asymmetric MTV changes and predicted more likely pCR.

Blended PET/CT registration led to more accurate quantification of MTV changes and prediction of pCR than PET-PET and CT-CT registrations.

ABSTRACT. Protocols for carbon ion therapy were developed to: (1) rapidly calculate the ionization density from a combination of condensed history Monte Carlo and track structure simulations of the particle beam, and (2) use the ionization detail (ID) calculated for the clinical beam in this way to simultaneously optimize the radiobiological dose and ID (in this case, the density of large ionization clusters consisting of more than 3 ionizations within a distance of 10 DNA base pairs). These published capabilities were demonstrated experimentally with a clinical carbon beam by planning and irradiation of an anthropomorphic head phantom with a stack of 3 custom-made micro-well plates embedded in a 4 cm wide cubic cavity in the phantom. A set of carbon treatment plans was prepared for the phantom using the open source matRad treatment planning system developed at DKFZ, modified for simultaneous optimization. Phase space data provided by T Tessonnier and K Parodi (see Tessonnier T et al, Front. Oncol. 11:297 and references therein) were used to calculate the dose and fluence in each voxel required for the dose, LET and ID calculation. One plan was developed to deliver a uniform biological dose across the target volume according to the LEM1 biologically equivalent dose model but without ID optimization. A second plan was developed to simultaneously optimize uniform biological dose (LEM1) with uniform ID. The plans were delivered at the Heidelberg Ion Therapy Center. The experimental design will be presented along with a comparison of the biological dose and ID distributions achieved with and without ID optimization and the cell viability measured in the 48 micro-wells in the cavity for both cases.

Acknowledgments: Supported by NIH/NCI P20CA183640 and the HIT, NCT and DKFZ-Heidelberg. Phase space data provided by Thomas Tessonnier and Katia Parodi, LMU, Munich, Germany.

ABSTRACT. Purpose: To investigate spatially varying instantaneous dose rates for different intensity-modulated proton therapy (IMPT) planning strategies and delivery scenarios, and compare these with FLASH irradiations (>30Gy/s). Methods: For a head-and-neck case, 3D dose rate distributions were calculated for the clinically applied (33,855 spots) and spot-reduced (1,006 spots) IMPT plans. For both plans, different delivery scenarios have been simulated: constant beam intensity (PSI-Gantry 2), variable beam intensity per energy-layer or per spot (assuming currently achieved Varian ProBeam intensities on PSI-Gantry 3), and hypothetical, spot-wise varying, unrestricted intensities such that spot times are always kept ≤3ms. The clinical plan, originally generated for PSI-Gantry 2, was not evaluated for PSI-Gantry 3, as many spots were below the minimum deliverable spot weight. For each voxel in the plan, the ‘dose-weighted dose rate’ was calculated, defined as the sum of the spot-wise instantaneous dose rates weighted by their dose contribution to the voxel. All dose rates were calculated assuming a 1.8GyRBE fraction dose. Results: For the clinical plan, average dose-weighted dose rates of 0.7 and 1.7 GyRBE/s were obtained when considering constant and unrestricted beam intensities, respectively. However, for the spot-reduced plan, as higher beam intensities can be utilized, averaged dose-weighted dose rates in healthy tissue increased from 0.7 to 3.4, 5.1 and 14.3 GyRBE/s, for the constant, energy-layer-wise, spot-wise and hypothetical unrestricted beam intensities, respectively. Comparable dose rates were observed for the PTV. Dose rate values were distributed inhomogeneously across the irradiated volume with highest values being found at beam entrance and at the distal end of each field, the latter being potentially beneficial for adjacent/overlapping organs-at-risk. Conclusion: FLASH dose rates were not achieved for conventional planning and current clinical PBS machines. As such, increased energy-layer or spot specific beam intensities, higher fraction doses and/or spot-reduced plans are required to approach FLASH compatible rates.

ABSTRACT. Purpose: To explore the spatial patterns of dose that impact treatment related xerostomia. With machine learning, it is common to explore prediction models with abstract features to find the most accurate predictions. In the context of spatial dose patterns, it is important to understand the features of dose, the standard of care patterns of dose and the existing knowledge of the underlying tissue function when interpreting the findings. Failure in this understanding may lead to incorrect predictions.

Methods: To highlight these concepts, we use as an example, the spatial dose analysis of the salivary glands for xerostomia injury and recovery using both voxel and sub-volume dose features. Ridge logistic regression was used for both feature sets, with additional clinical variables to determine influence patterns across the voxels and the sub-volume dose features to provide insight into the spatial dependence of dose on xerostomia injury and recovery.

Results: The influence patterns suggest potential hypotheses of salivary duct involvement and the importance of low dose regions (mostly superior). The coupling of standard of care dose patterns with spatial analysis will be discussed in the context of hypothesis generation and dose influence on underlying mechanisms.

Conclusions: Decoupling the dose pattern from the underlying tissue function and sensitivity is very challenging using clinical data alone. The challenge comes from the variability (or lack thereof) in the dose patterns. If a single dose feature has the same value for all patients, then there is no information in your dataset about that feature. Since we treat patients very similarly in terms of dose goals to normal tissue, the knowledge about those dose goals is suppressed in clinical data. Therefore, it is important to consider the existing knowledge when using machine learning to advance our understanding as the learned prediction models may not be aware of it.

ABSTRACT. Introduction Xerostomia is common in head and neck cancer patients with radiation therapy. This study aims to evaluate the feasibility of generating outcome-based treatment plans that shape the dose patterns to parotid sub-volumes and improves the probability of xerostomia injury(grade>2 within 3-months) and recovery(reduced to grade<2 by 18-months).

Methods Dose features (D10-D90) of the 9 sub-volumes from parotid glands, submandibular glands, and oral cavities were extracted by a feature generation pipeline. Ridge logistic regression model was previously developed to predict the probability of xerostomia injury and recovery. Partial derivatives were computed and sorted descendingly to get the weight of the dose features. Injury-weighted planning was optimized using weights for lower injury probability, and recovery-weighted planning was optimized using weights for higher recovery probability. After adding objectives for the sub-volume dose features to existing physical dose-volume based objectives, IMRT optimization was applied in Pinnacle. Outcome-based plans were compared to the original plan for comparable target coverage and to ensure dose to the important organ at risks met clinical constraints. For outcome-based plan evaluation, prediction of probability of injury and recovery after final optimization were compared to that in the original plan. Then the optimization plan and the predicted clinical outcomes were reviewed with two independent clinicians to assess the plan quality and what the best plan would be for the patient.

Results The prediction injury and recovery for 3 cases have been improved in both outcome-based plans. Both injury-weighted and recovery-weighted tend to reduce dose to high dose areas, injury-weighted tends to lower dose in the lower dose regions as well.

Conclusions Our study demonstrated the feasibility of outcome-based planning strategy and indicated that it could change the pattern of dose to parotids and improve the overall probability of injury and recovery. Selection of plans must balance injury and recovery potential.

ABSTRACT. Data-driven NTCP modelling is often impeded by high collinearity between explanatory variables (EVs), resulting in unstable variable selection and inflated variance. This problem can lead to models that highly depend on specific correlations between EVs and, therefore, generalise poorly to other populations. We modified a frequently used model development method (MDM) to deal with this problem and tested the performance and generalizability of the resulting models in random simulations.

METHODS

As reference MDM we used stepwise logistic regression with forward variable selection based on the Bayesian Information Criterion (BIC). We modified this MDM by splitting the EVs into overarching groups with mutual correlations ≤0.8. For each group, a prediction model was developed using the reference MDM. Subsequently, models with good performance were recombined into a single logistic model.

We simulated random datasets with 10 EVs drawn from a standard Gaussian distribution with autoregressive correlation structure and correlations of 0.9 between neighbouring EVs. A binary response variable was added according to a logistic response relation with 5 randomly chosen EVs. Datasets of varying size were generated and used to develop prediction models using both MDMs. Using the same response relation we generated multiple validation sets of 1000 observations: one set with the same correlation structure as the training set, to test model performance in the same population, and other sets with different correlations, to test model generalizability.

RESULTS

For data set sizes up to 500 observations, the modified MDM performed and generalized better than the reference MDM, but not in each single simulation. For larger data sets the performance of both MDMs stabilized at a constant high level, with only small differences.

CONCLUSION

The modified models predict and generalize better than the reference models, up to the point where sufficient data is available to reliably estimate all model parameters.

ABSTRACT. Effective irradiation of tumor volumes while sparing surrounding Normal Tissue (NT) demands a trade-off in both spatial conformity and temporal fractionation of the dose. Here we present an approach to incorporate variable RBE models into spatiotemporal optimization of combined photon and carbon ion treatment plans based on the Linear Quadratic Model. A fundamental case of a voxel containing a mixture of healthy and normal tissue was considered. The tumor cells within the voxel are to be irradiated with a prescribed dose of 4 Gy in two fractions (one of photons and one of carbon ions), while maximizing the sparing of healthy cells within the voxel through fractionation. Therefore, a log cell kill based objective function was formulated to compute the optimal spatial and temporal dose distribution. This objective function comprises a squared under-dosage objective on tumor tissue and a squared over-dosage objective on normal tissue for the same voxel. A systematic study of the fundamental implications of a clinically used RBE model for carbon ions on the optimal temporal distribution of dose, was conducted in comparison to the standard photons only strategy. The optimal photons-only solution is equal fractionation, satisfying the coverage objective on the tumor tissue and the sparing objective on the NT at the same time. Due to the strong variability in RBE along the beam, the optimal photon-carbon combination depends on the relative location of the voxel within the carbon ion beam. For the carbon-photon combinations, it is impossible to satisfy the competing objectives at the same time. Furthermore, in regions close to the Bragg peak the ideal dose combination is an only photon dose plan. We advise a structured optimization process to regulate carbon ions when Bragg peaks are located near healthy tissue that could benefit from fractionation.

ABSTRACT. 50-years ago the Data General (DG) Nova mini-computer was released. Hounsfield and colleagues chose this computer to drive the first CT scanner. I recently acquired an incomplete Nova. Here I describe its restoration and programming, which is of historical and educational interest. The Nova is the first commercial 16 bit mini-computer. My machine consists of two large printed circuit boards: the CPU – with 110 small-scale integrated TTL circuits –, and the core memory of 8192 x 16 bit capacity. The latter contains over 100.000 small ferromagnetic rings that store one bit each in the direction of their magnetic field. To make the machine useful I rebuilt the missing front panel that controls it. I took this as an opportunity and added an Arduino processor, a small keypad and a textual display to make it more visual and allow remote control. The Nova now works, and I am converting the XVI cone beam reconstruction code to run on it. The CPU and memory are slow (0.8 MHz), yet each instruction is powerful. A software multiplication routine has only 5 instructions in its inner loop (run 16 times) and takes 120 µs. CT reconstruction requires FFT based filtering of the input data (80 detectors for the EMI), followed by backprojection onto the CT slice (128 x 128 for the EMI), requiring in total about 1 million multiplications for 90 angular samples. I expect my code to take 2 minutes to reconstruct a one slice. The EMI scanner took 7 minutes, likely because of less optimized code and use of higher precision. In conclusion, I successfully restored a Nova 1210 computer. Hopefully this work will help future generations understand more about the workings of computers and image processing - the older hardware is simpler to understand yet quite similar to modern computers.

ABSTRACT. The sensitivity of particle therapy to uncertainties led to the development of robust and probabilistic optimization methods. These methods, however, lack the possibility of a priori constraining the outcome to a desired confidence on dose or plan quality indicators. This work introduces a mathematically rigorous implementation of such confidence constraints on moments of dose-volume histogram (DVH) points based on analytical probabilistic modeling (APM) of range and setup uncertainties. With the resulting closed differential mapping from fluence to a DVH-point’s quantile function, a common quasi-Newton optimizer could be used to optimize treatment plans constrained to not violate a desired DVH-point with predefined confidence. For a C-shaped two-dimensional phantom, treatment plans optimized with (1) a conventional constraint and (2) a confidence constraint are displayed together with their respective nominal and probabilistic DVHs. The feasibility of this novel kind of constrained probabilistic optimization could be successfully demonstrated. After nominal conventional optimization, the applied hard DVH constraint to the target (98% volume coverage at 98% of the prescribed dose) was honored in the nominal scenario while it was violated in a majority of possible error scenarios. The constrained probabilistic optimization succeeded in keeping also a desired part of the probability mass (98%) above the DVH constraint by dynamically trading in organ-at-risk sparing. While the proposed method relies on an analytical propagation of uncertainties through dose calculation, optimization and DVH computation, the proposed optimization paradigm could be extended to work with statistical estimates from scenario samples. Further, this work relied on normal distributions, but also allows parametrization with other probability distributions as long as the cumulative distribution and quantile functions including the respective derivatives can be formed. Future research focuses on application to three-dimensional patient cases and mitigation of computational performance limitations.

ABSTRACT. Objective To demonstrate that a dose distribution for head and neck cancer (HNC) patients can be predicted within minutes, which can then be mimicked to create a deliverable volumetric modulated arc therapy (VMAT) dose distribution with similar quality as the clinical ‘dosimetrist-optimized’ dose distributions, further indicated as reference plans.

Method The machine learning based automated planning (MLAP) involves training of a model, which is used to predict the voxel dose in novel patients. The CT scan, structures and dose distributions of 71 consecutive primary HNC patients, previously treated with dual arc VMAT, were retrieved from our clinical database. In the final step, the predicted dose distribution was input to a mimicking optimization algorithm to generate a deliverable dose distribution. We applied a repeated random subset validation approach; 8 test patients in 4 sets. The dose distributions of the test patients were compared against the reference plans.

Results The predicted dose was in accordance with reference dose for all plans. The mimicked plans had similar target coverage for high risk (66.9 ± 0.5 Gy) and elective volumes (51.9 ± 1.0 Gy) as reference plans, (high risk: 67.1 ± 0.7 Gy, elective: 52.6 ± 2.0 Gy). The NTCP values of mimicked plans (sum NTCP: 71.6 ± 27.0 %) were lower or did not increase with more than 2.0% compared to the reference plans (sum NTCP: 72.3 ± 30.4 %) in 23 of the 32 cases.

Conclusion In this study, we demonstrated that adequate target coverage can be reached using MLAP in the majority of HNC VMAT treatment plans. Similar or lower NTCP values were reached in 23 of the 32 plans This indicates that MLAP can serve as a promising tool for automated treatment planning.

ABSTRACT. Inverse planning for IMRT and VMAT requires huge computation resources because fluence maps have thousands of beamlets needed to be optimized. In this paper we explore how to bypass the fluence map optimization by deep learning.

65 prostate cancer IMRT plans with seven 10 MV beams are collected from our institution to build a deep learning model for fluence map prediction. All patients have a similar seven beam angles distribution which makes the learning possible. A 3D U-net is trained to predict fluence maps. An ordered stack of seven 2D fluence maps is the 3D output of model while Beam Eye View (BEV) of PTV, bladder and rectum mask from corresponding beam angle are three channels of input. Each input channel has a stack of seven 2D BEV images in same way as the output channel does. Mean square error is minimized with an intial learning rate of 1e-4 and a step decay of 0.1 fold every 100 epoches. 65 plans are split into training set(41), validation set(11) and test set(13). Mean square errors for training set, valiadation set and test set are 0.0011±0.0004, 0.0024±0.0009 and 0.0035±0.0024. To evaluate the quality of prediction, 3D doses are calculated from predicted fluence map by CCCS algorithm and compared to doses in TPS by gamma analysis. Gamma passing rates with 3% 3mm criteria for training set, validation set and test set are 0.94±0.05, 0.87±0.06 and 0.83±0.08 respectively. The feasibility of learning fluence map pattern from structure projections is validated and dense prediction of fluence map for VMAT remains to be done in the future, which shortens planning time efficiently.

ABSTRACT. Introduction: Current standard small animal irradiator uses collimators of fixed sizes and shapes, which fails to precisely irradiate the target volume while effectively sparing organs at risk (OARs). To improve treatment quality, we developed a treatment planning system to achieve intensity-modulated radiation therapy (IMRT) using a movable rectangular collimator typically available on current irradiators. Methods: We formulated IMRT treatment planning as a direct aperture optimization problem under the constraint of rectangular aperture shapes. We computed dose deposition matrix using goMC, an in-house developed GPU-based fast Monte Carlo dose calculation package. A column generation algorithm was employed to solve the optimization problem to determine a series of rectangular aperture shapes and corresponding beam-on time. For each beam angle, we further determined the most efficient delivery sequence by solving a shortest-path problem in a graph with vertices being the apertures and links being transition time between apertures. We validated the developed treatment planning system on a water phantom case and a mouse case with a lung tumor. Plan quality and delivery efficiency were compared between the IMRT plan and the plan using cone collimators. Results: In both cases, the IMRT plans generate by the proposed treatment planning system improved the OAR sparing while maintaing the same target coverage. In the mouse case, the delivery time of IMRT plan was ~80 s, approximately doubling the delivery times of cone-based plan due to blocked radiation by apertures and hence reduced beam output. The IMRT case required additional 6.9 s collimator motion time. Conclusion: We have developed an inverse IMRT treatment planning system for small animal irradiation using a movable rectangular collimator. The IMRT plan quality outperforms that of the standard cone-based plan. Treatment delivery time is prolonged mainly due to reduced beam output under the collimation.

ABSTRACT. Background and Purpose: Probabilistic optimization is an alternative to margins for handling geometrical uncertainties in treatment planning of radiotherapy where the uncertainties are explicitly incorporated in the optimization. We present a novel probabilistic method based on the same statistical motivation as those behind conventional margin based planning. This method was applied to treatment planning of cervical cancer patients with focus on the systematic uncertainty caused by organ deformation.

Material and Methods: Percentile Dosage is defined as the dose coverage that a treatment plan meet or exceed to a given probability. The convex measure conditional value at risk (CVaR) was used to constrain the dose coverage of the 10% worst scenarios. The constraint tolerance was gradually tightened until the desired percentile dosage was met. The proposed method was compared to conventional margin based plans where the CTV to PTV margin was adjusted to fulfil the same percentile dosage.

Results: The proposed probabilistic plans and the margin based plans both fulfilled the requested percentile dosage within 1% of the requested. The probabilistic plans showed superior conformance while at the same lowering the rectum max dose with 2.4 Gy.

Conclusions: Probabilistic optimization of radiotherapy treatment plans where a percentile dosage is requested was shown to be feasible. The proposed method makes it possible to prescribe a dose volume histogram value to a chosen confidence. The resulting plans were shown to be superior to margin based plans.

ABSTRACT. Purpose: Inverse radiotherapy treatment planning requires large scale optimizations that must be solved within a clinically feasible time frame. We have developed and tested a quantum-inspired, stochastic algorithm, Quantum Tunnel Annealing (QTA), for intensity-modulated radiation therapy (IMRT). By modeling the likelihood of accepting a higher energy solution after the probability of a particle tunneling through a potential energy barrier, QTA features an additional degree of freedom (in the form of the barrier width, w) not shared by traditional Simulated Annealing (SA). This additional degree of freedom can be utilized to achieve faster convergence rates. Methods: Beamlet-weight optimizations were performed on a stereotactic body radiation therapy (SBRT) liver case that featured complex geometries and overlap between the planning target volume (PTV) and multiple organs at risk (OARs). The convergence rates of QTA were investigated in relation to four barrier-width schedules, one based on the growth rate of gallium arsenide (GaAs) during Metal-Organic Chemical Vapour Deposition (MOCVD) and three designed to model local fluctuations in width by coupling a sinusoidal function with a fractional polynomial function. Results: Three of the four barrier-width schedules tested for QTA converged faster than SA. The highest- and lowest-performing width-schedules both used sinusoid-coupled functions, one being 50% faster than SA and the other being ~66% slower. The MOCVD-based schedule provided ~18% speedup. These results indicate that the barrier-width schedule of QTA can be tuned to achieve faster convergence rates. In addition, it was found that QTA achieved results which were equal to SA in terms of plan quality and were invariant across multiple runs of the algorithm. Conclusion: This proof of concept study suggests that QTA is a robust optimization technique that can achieve faster convergence than SA and can be a useful candidate for future demanding applications in adaptive radiation therapy.

ABSTRACT. Purpose: Manual fine-tuning of an objective function is a common practice in high-dose-rate (HDR) brachytherapy to obtain case-specific plans because most inverse planning optimization algorithms used in clinic are designed to optimize a single treatment plan. To facilitate this process, this study presents a graphics processing unit (GPU)-based implementation of IPSA (Inverse Planning Simulated Annealing) or gSA and of the quasi-Newton optimizer L-BFGS (Limited-memory Broyden-Fletcher-Goldfarb-Shanno) or gL-BFGS that enable up to 1200 parallel plans/s for multi-criteria optimization (MCO) for HDR brachytherapy.

Methods: A 4D MCO problem for HDR prostate brachytherapy was formulated into a weighted sum to form a single joint objective function. gSA and gL-BFGS algorithms were used to calculate multiple plans with various trade-offs simultaneously. 100 previously treated HDR prostate cases were used to benchmark the performance of gSA and gL-BFGS algorithms. To evaluate the fraction of Pareto optimal plans with both algorithms, 1000 plans/case were optimized in parallel for all cases. The computational efficiency of gSA and gL-BFGS algorithms was obtained by timing the optimization process and by varying the number of simultaneous parallel plans from 1 to 10000. Both algorithms were executed on an NVIDIA Titan X (Pascal) GPU.

Results: The mean fraction of Pareto optimal plans in the 4D objective space was 99.6% for gSA and was 99.3% for gL-BFGS. The mean optimization time of both algorithms increased as the number of plans increased. However, the mean optimization time per plan decreased up to 5000 plans for gSA (50 ms/plan or 20 plans/s) and for gL-BFGS (0.8 ms/plan or 1200 plans/s). The Parato Front obtained by gSA and gL-BFGS algorithms are in agreement, showing a convex solution space for the problem at hand.

Conclusion: gSA and gL-BFGS are two fast GPU-based optimizers able to populate the Pareto Front with 1000 parallel plans under one minute.

ABSTRACT. Stochastic optimization allows to account for physical treatment plan uncertainties during inverse optimization by assigning probabilities to individual dose error scenarios. This requires to make assumptions about the magnitude and shape of the corresponding input uncertainty to be modeled. Moreover, input uncertainty assumptions also play an essential role in uncertainty quantification methods such as random sampling or analytical probabilistic modeling (APM). At the moment it is a common standard in particle therapy to model input uncertainties, e.g., range error, with Gaussian probability densities. This study investigates the dosimetric consequences of a conventionally and probabilistically optimized one-dimensional carbon ion SOBP if the actual range error obeys a different probability density functions (PDF). We observed that different PDFs yield different expectation values and standard deviations. Interestingly, the impact of different PDFs is of lower importance after probabilistic optimization as the resulting treatment plan was robust against other PDFs although normally distributed errors were assumed for probabilistic optimization. We could make similar observations for a 1D proton SOBP. A comprehensive analysis with patient cases also including setup uncertainties is the focus of current research.

ABSTRACT. The multi-criterial lexicographic reference point method (LRPM) guarantees fast (1-5 minutes), automated Pareto-optimal treatment plan generation. For carefully tuned parameters, the LRPM has demonstrated capability in consistently generating clinically favourable treatment plans. So far, however, these parameters were tuned manually in a time-consuming, labour-intensive trial-and-error procedure. Instead of manual tuning, we propose a novel procedure to automatically tune the LRPM parameters based on prior treatment plans. To this purpose, plan/patient databases for prostate cancer (287 plans) and head and neck (HN) cancer (105 plans) were collected. All database plans were generated with our clinically applied multi-criterial algorithm (other than LRPM) for automated Pareto-optimal treatment planning. The novel procedure generates a single set of LRPM parameters per treatment site according to user-specified preferences which define acceptable thresholds for differences observed in criteria values between LRPM plans and training plans (part of the database; 36 for prostate, 30 for HN). Quality of the LRPM parameters was tested by comparing differences in criteria values between LRPM plans and evaluation plans (plans not used for training; 251 for prostate, 75 for HN). All LRPM plans had sufficient PTV coverage (V95%≥99% for prostate, V95%≥98% for HN). For prostate cancer, the LRPM plans had overall slightly decreased rectum and bladder dose at the cost of slight increases in anus dose (within constraints), which had a lower clinical priority. Also for HN, all LRPM plans were acceptable with overall slightly higher clinical quality. For a subgroup, large improvements for some criteria were achieved at the cost of slight degradations for other criteria. In conclusion, we have developed a procedure to automatically tune the parameters of the LRPM algorithm for automated Pareto-optimal plan generation. Due to this tuning automation, a major step has been made towards introducing fast and automated generation of Pareto-optimal treatment plans in the clinic.

ABSTRACT. Radiomics is a rapidly evolving field of research combining medical imaging, medical physics and radiation oncology, image processing, image analysis and machine/deep learning. Despite promising results in clinically-relevant datasets, the radiomics approach has yet to be transferred to the clinical practice. Several challenges remain to be addressed before radiomics can realize their full clinical potential. One of the main challenges is the need for large, multicentric, if possible prospective, studies to be carried out, in order to provide a high level of proof. Such studies require standardization and harmonization solutions for radiomics that have to be implemented and validated. In this lecture, we will present the existing methods and currently available initiatives and results, as well as the perspectives regarding standardization and harmonization of the radiomics approach.

ABSTRACT. Radiomics has shown promising results mainly for Computed Tomography (CT) / Positron Emission Tomography (PET) for lung and head and neck cancers. Extending radiomics to MRI (Magnetic Resonance Imaging) is of interest for pelvic malignancies, where derived DWI (Diffusion Weighted Imaging) is used to support staging and evaluation of treatment response. However, DWI and related ADC (apparent diffusion coefficient) maps suffer from reporting inconsistencies beyond that seen for CT-based radiomics assessments. Associated challenges include: a) larger differences in terms of acquisition settings, b) lack of signal normalization, c) more common acquisition artifacts, and d) potentially larger inter-observer variability in tumor delineations. Harmonization and standardization of features extraction from ADC is required, even before investigating the prognostic power of radiomic features. This study evaluates different methods to optimize the extraction of radiomics by looking at 81 ADC maps of Stage IB-IVa cervix cancer patients. Tumors were delineated by two independent observers and most common radiomic features were extracted using an open source package, including different normalizations and quantization techniques. Features reproducibility was evaluated using the Concordance Correlation Coefficient with respect to inter-observer variability for all the configurations considered. Additionally, dependencies between feature values and tumour volume were considered using Spearman analysis. Results show that features reproducibility is strongly affected by image pre-processing (e.g. normalization, quantization) prior to features extraction and by inter-observer variability. Normalizing images based on bladder urine-values and computing features using smaller bin widths lead to the most reproducible configuration. The presented methodology can be extended to other MR sequences and will help pushing forward the standardization and harmonization of radiomics in MRI.

ABSTRACT. Background: Radiomics features have the potential to serve as surrogates for parameters, such as tumor heterogeneity level, pathology, and response to a given therapy. Quantitative image feature analysis is being explored as a noninvasive method to identify diagnostic, prognostic and predictive imaging biomarkers. The aim of this study is to extract multiscale radiomics features from preoperative MRI and identify statistically relevant imaging features to predict the overall survival of Glioblastoma Multiforme (GBM). The predictive power of the selected features is accessed using logistic regression(LR)/support vector machine(SVM) analysis as well as the Kaplan–Meier estimator.

Method: T2-(T2W), T1-(T1W) weighted, T1-contrast enhanced(T1CE), and fluid attenuated inversion recovery (FLAIR) MR brain images, as well as patient age and survival, were obtained for 61 patients with GBM and varying levels of resection from the 2018 Multimodal Brain Tumor Image Segmentation Competition (BRATS) [1,2]. These images were registered and resampled to a standardized voxel size of 1x1x1mm3. Enhancing/whole tumor contours were delineated by 1-3 physicians. Histogram based intensity features, volumetric, morphologic, GLCM, GLRLM, GLSZM, NGTDM, lattice computation, and LBP features were extracted by the Cancer Imaging Phenomics Toolkit (CaPTK) [3]. This resulted in 968 total features per contour per modality. To be defined as statistically relevant, a feature had to show relevance in the univariate Cox Proportional Hazards (CPH) model (p-value<0.05), be highly ranked by 10 cross-fold SVM recursive feature elimination (SVM-RFE), and possess non-collinearity. K-M and 10 cross-fold LR/SVM were used to quantify statistically relevant feature performance. Survival time was placed into three categories based on dataset patient survival time for regression analysis: <10, 10-15, >15 months.

Results: Eight image features from enhancing tumor delineation, and three image features from whole tumor contour were found to be statistically relevant with regards to overall survival. Logistic Regression (0.69±0.19) outperformed SVM(0.52±0.19).

ABSTRACT. In this study, we analyzed baseline MR-based image features of parotid and submandibular glands to predict radiation-induced xerostomia in head and neck cancer (HNC). A retrospective analysis was performed on 92 HNC patients who were treated using radiotherapy at a single institution. T1 post-contrast MR images along with NCI-CTCAE xerostomia grade (3-month follow-up) were prospectively collected. Ipsilateral/contralateral parotid and submandibular glands (iPG, cPG, iSG, cSG) were contoured by the attending radiation oncologist. MR images were normalized using two methods: 1) relative intensity based on the mean subcutaneous fat at the level of the salivary glands, and 2) discretization of the intensities from 0 to 500 based on the minimum and maximum of the region of interest. Gray-level-co-occurrence-matrix (GLCM), gray-level-run-length-matrix (GLRLM), gray-level-size-zone-matrix (GLSZM), and gray-level-dependence-matrix (GLDM) with bin widths of 10, 15, 20, 25, and 50 were computed along with shape and first order features for ipsilateral/contralateral parotid and submandibular glands relative to the location of the primary. Features that were correlated with xerostomia (p<0.05) were further reduced using a LASSO logistic regression. Generalized Linear Model was used to predict xerostomia under five for original and normalized imaging (bin width 10, 15, 20, 25, and 50) using a ten-fold cross validation. The prediction performance was determined using the area under the receiver operator characteristic curve (ROC-AUC). For the original images, the model resulting in the greatest AUC contained features with bin width 10 (AUC/Sensitivity/Specificity=0.73/0.68/0.73). The greatest AUC contained fat normalized features with bin width 20 (AUC/Sensitivity/Specificity=0.77/0.70/0.78). The selected features using bin width 20, regardless of normalization, described the local heterogeneity of the submandibular glands such as the GLRLM run-variance and GLSZM gray-level-non-uniformity-normalized. Taken together, fat normalized images resulted in the greatest prediction performance. Future studies are required to validate these findings in a larger cohort.

ABSTRACT. Purpose: The aim of this study was to investigate whether CT image-derived radiomics are able to predict Human papillomavirus (HPV) status of patients diagnosed with primary oropharyngeal squamous cell carcinoma (OPSCC). in a non-invasive approach.

Material and methods: Six independent cohorts, 255 patients in total were collected in this study, in which patients were treated with radiation only or chemo-radiation therapy as part of their treatment. HPV positive was defined as expression of the p16 gene variant. CT scans with visible artifacts within the GTV were excluded. The data was randomly split into training (n=142), tuning (n=48) and test (n=65) sets. A total of 1105 radiomic features were extracted from the GTV via PyRadiomics. For the model, a parsimonious yet optimally predictive set of features from the remaining candidates after selection were determined by recursive feature elimination. Multivariable logistic regression, SVM, and random forest, were applied for training classifier. The classification performance of HPV status was assessed by the area under the receiver operator curve (AUC). The Wilcoxon test was used to assess significance between AUCs and random (AUC=0.5).

Results: Out of 255 patients, 174 (68.2%) patients were with HPV positive and the rest 81 (31.8%) patients were with negative. After feature selection, 3 radiomic features, (i) original_shape_Sphericity, (ii) original_firstorder_ Entropy, and (iii) log-sigma-3-0-mm-3D_glszm_SizeZoneNonUniformityNormalized, were selected to develop models. The multivariable logistic regression classifier achieved the highest AUC on both training and test sets, yielding mean AUCs of 0.79 (95% CI: 0.78 - 0.79, p-value < 10-4 Wilcoxon test) and 0.72 (95% CI: 0.71 - 0.72, p-value < 10-4 Wilcoxon test), respectively.

Conclusion: It is possible to classify HPV status for HNSCC patients using CT image-derived features, which could lead to better precision treatment decision. However, the results should be further validated on larger and external datasets.

ABSTRACT. Purpose: In treatment management of renal cell carcinomas (RCC), distinguishing indolent tumors from those that will progress would enable early triage to observation versus invasive treatment. In this work, our main objective was to evaluate the feasibility of using a decision-support system based on radiomics analysis of magnetic resonance imaging (MRI) for RCC management.

Methods: A dataset of 569 patients with renal masses coming from five different institutions and with pre-treatment MRI scans (T1-weighted contrast enhanced and T2-weighted sequences) was analyzed. A total of 3087 radiomics features were computed by extracting textures with different parameterizations. The whole dataset was randomly divided into 10 teaching and testing sets. The teaching sets were again randomly sub-divided into 10 training and validation sets to perform radiomics feature set reduction and optimize random forests (RF) models for three types of initial feature sets: i) radiomics; ii) clinical; and iii) radiomics+clinical features. The modeled clinical endpoints were: A) HighGrade (n=398; high/low=0.6); B) Malignancy (n=569; malignant/benign=4.6); and C) Subtype (n=412; clear cell/papillary RCC=5.2). Adjustments for class imbalance were applied during modeling. Prediction performance of final RF models over the 10 independent testing sets was then calculated using receiver operating characteristic metrics: the area under the curve (AUC), sensitivity (SEN) and specificity (SPE).

Results: For each clinical endpoint, independent testing set results are presented only for the type of initial feature set estimated to possess the highest predictive power (i.e. estimated from the validation sets): A) for HighGrade, radiomics models reached AUC=0.74+-0.05, SEN=0.62+-0.09, SPE=0.74+-0.06; B) for Malignancy, clinical models reached AUC=0.89+-0.08, SEN=0.90+-0.08, SPE=0.68+-0.13; and C) for Subtype radiomics+clinical models reached AUC=0.83+-0.04, SEN=0.81+-0.06, SPE=0.70+-0.09.

Conclusion: Using combinations of clinical and radiomics-based models, it could be possible to construct a decision-support system for the early management of renal cell carcinomas.

ABSTRACT. Introduction: Radiomics has focused on extracting pixel values directly from patient images. In this work, we propose a novel methodology to extract information using radial histograms. This samples pixel values as a function of distance from a defined structure. Here, we apply this approach to stage-1 non-small cell lung cancer (NSCLC) patients treated with SABR to investigate if pixel values outside the GTV predict for survival.

Methods: 404 stage-1 NSCLC patients with radiotherapy planning data, tumour size, age, and performance status were analysed. The GTV was reconstructed and 2-dimensional cross-histograms of pixel intensity versus distance from GTV calculated for all phases of the 4DCT. Histograms were limited to inside the lung contour to avoid becoming a proxy for tumour location. Average cross-histograms were created for patients alive and dead, with global significance of any differences tested with permutation testing. Average pixel intensities from the significant region were extracted and tested in a cox-proportional hazards model with Kaplan-Meier curves plotted.

Results: A significant difference in mean pixel intensity was found between patients who survived 12 months post radiotherapy. The significant region was found across the tumour boundary, -5mm to 5mm, with permutation testing showing a significance of p=0.03. The cox-proportional hazards model showed pixel intesities significant (p=0.015) but tumour volume, age and performance status were not significant. the Kaplan-Meier curve, split at the 3rd quartile (811HU), showed a log-rank p=0.016. Median survival of 3.2 versus 2.1 years was found for low and high density.

Discussion & Conclusions: This work utilised a novel technique for extracting radial pixel intensity information as function of distance from a defined structure. We have applied this technique to stage 1 NSCLC patients treated with SABR and found a significant difference in pixel values across the tumour boundary, where patients with higher pixel values do worse.

ABSTRACT. The combination of quantitative information extracted from patients scans and artificial intelligence (radiomics) to develop clinical prediction models has recently increased in radiation oncology. Radiomics has demonstrated promising prognostic and predictive results for overall survival, distant metastases and cancer biology. However, a recent number of publications pointed out possible vulnerabilities in radiomic signature development processes, such as larger dependencies between features and with respect to accepted clinical predictors (e.g. tumour extension), or the risk of spurious associations due to limited datasets. Refinement of radiomic results and methodologies is required to ensure progression of the field, by introducing safeguards that can improve the quality of radiomic studies. Machine learning (ML), often used in the final step of the radiomics pipeline for features aggregation, represents a powerful tool to strengthen radiomics methodologies, prior to modelling. In this study we propose machine learning-based methods (supervised and unsupervised learning) to tackle the above-mentioned vulnerabilities and a support for safeguards. 841 radiomic features were extracted using an open source software from two retrospective publicly available datasets of lung and head neck cancers. Spearman analysis combined with unsupervised hierarchical clustering and principal component analysis (PCA) were used to identify relations between radiomics and clinical outcomes (overall survival). Bootstrapping techniques with logistic regression were used to verify features prognostic power and robustness. Results show that over 80% of the features had large pairwise correlations. Nearly 30% of the features presented strong correlations with tumor volume. Using volume-independent features for clustering and PCA did not allow risk stratification of patients. This methodology has the power to be extended to any radiomics study as test of robustness for signatures development.

ABSTRACT. Continued advances in digital technologies have dramatically lowered the costs of measurement, communication, computation and storage of data.  The combined effect of these advances is having a profound effect on how humans interact with each other and with machines. In the context of healthcare, it is well-understood that there is a significant need to make the care we deliver safer, of higher quality, and more personalized.  Furthermore, it is now being recognized that the process of delivering care should also yield the insights and evidence for the next generation of care. This ‘learning health system’ paradigm is not new. However, the development of machine learning (ML) and data governance (DG) technologies are driving new architectural forms for the digital systems and new mindsets for the participants – including the patients themselves. In this presentation, the speaker will share experiences both successful and otherwise in an effort to position Canada’s largest hospital network (University Health Network) to be the world’s fastest learning health science centre.

ABSTRACT. Observational studies are increasingly accepted as a source of evidence in radiotherapy research where randomized trials are not appropriate or data does not exist. Rapid learning offers a potential methodology to address the persistent criticism of the quality of such studies that impedes their clinical application. The approach analyses routine data in a cyclical improvement process. By iteratively making small changes to practice, treatments could be optimized in a safe, continuously monitored environment.

An informatics platform has been developed at The Christie NHS Foundation Trust with the aim of integrating big data analyses into routine clinical practice. A sensitivity analysis framework was developed to enable rapid learning optimization of the many small changes to radiotherapy practice currently made within the professional competence of departments. The framework was used to retrospectively analyse a change made to a lung cancer IGRT protocol. We find that uncensored 12 month follow up of NSCLC patients accrued for 11 months (N=363, total time=23 month), or 13 months censored accrual and follow up (N=429), can detect a discrepancy in survival (HR=1.6±0.2 at the α=0.05/2, β=0.8 level) resulting from residual set-up errors shifting patients’ hearts towards or away from the high dose region. Examination of the data post protocol change shows no hazard ratio detectable at this magnitude.

The framework is event driven and thus depends on patient accrual rate and event probability. Our model of rapid learning relies on sufficient cycle speed to limit the chance of other confounding changes occurring in the same timeframe. At institutes with smaller patient throughput, or in diseases with fewer events, therefore, alternative quickly accumulating improvement metrics, such as patient reported toxicities, must be used alongside survival. Current knowledge suggests heart dose constraints should be revised and this work shows that the resulting effects could be detected using this approach.

ABSTRACT. Ensuring an accurate and consistent delivery of high quality radiotherapy treatment (RT) at institutional level is extremely challenging due to the absence of data standardization/ontologies, adherence to an uniform treatment workflow, institution-specific factors such as practice culture, workflow set-up, clinician preferences, etc. These reasons contribute significantly towards variations in treatment care. Additionally, the aggregation of RT treatment data for the purposes of overall quality and outcome assessment is arduous since the clinical workflow is interwoven with numerous steps performed by multiple clinicians across varying clinical systems; thereby rendering RT data scattered across disparate databases. The Veterans Affairs (VA) designed an initiative to assess the quality of RT care, identify variations and outcomes, and gauge the performances of all VA RT clinics, through an electronic IT infrastructure - HINGE. HINGE is designed to facilitate an automated, accurate and seamless aggregation of RT data within and across every VA clinic at institutional level. HINGE employs design strategies that focus on capturing specific data elements that quantify quality of RT care by establishing inter-connectivity between VA's EMR - VistA, treatment management and planning systems in the clinic. Additionally, HINGE acts as one-stop integrated data-entry portal with customizable workflow setup and auto-populated discrete data-entry templates for every step in the sequential RT workflow. This allows for generation of high-quality, standardized and consistent RT treatment data gathered at institutional level which can be assessed and compared uniformly with established quality metrics (QM). The HINGE dashboard has visual analyses of every clinic's performance of each QM, national expectations and system-wide evaluation of treatments delivered. Additionally, HINGE is also a platform for a vast array of research analytics including advanced computational algorithms for enterprise-wide analyses, DVH-based analyses and more.

ABSTRACT. The objective of this study was to demonstrate the proof of concept FAIR ((Findable, Accessible, Interoperable and Reusable) Radiomics infrastructure on the top of DICOM data, which can make radiomics more FAIR by generating a standardized DICOM representation of the annotation results. Furthermore, the FAIR-Radiomics infrastructure provides I/O for standard radiomics analysis by PyRadiomics, and deep neural networks (DNNs) by Docker application. Four CT and PET/CT image datasets (632 patients totally), MultipleDelineation, RIDER, HeadandNeck1 and LUNG1, were included in this study. The FAIR-Radiomics infrastructure was developed and implemented using Amazon Service Web (ASW) image with the version of Ubuntu 18.04, 4 virtual CPU, 16 GB memory and 200 GB volume space. A bash script was generated to install all relevant dependencies and packages and download DICOM data of 4 datasets from XNAT repository automatically. The results of each patient comprised volume of interest (VOI) binary masks, volume image, DICOM segmentation (SEG) of all VOIs, DNNs results, 1152 radiomic features, and DICOM structured report (SR) coded by ontologies. We proposed an infrastructure to make the publicly available datasets more FAIR for quantitative imaging research, by generating DICOM SEG, reporting DICOM SR as well as providing conversion into other common standards. Furthermore, our program supports universal radiomics extraction using PyRadiomic and DNNs implementation via Docker container. Finally, the AWS image will be ready to share and the source code will be published soon, which allows other researchers to reuse the data, DNN model, radiomic features, DICOM SEG and DICOM SR.

ABSTRACT. Radiation oncology information architecture has been largely neglected in conventional health care operations resulting in independent data islands organized around application support rather than information accuracy and access. Monolithic application models of the prior century continue in this highly demanding computational environment, each application managing their own data instances and functionality added to the same core source code. We present an alternate information architecture now implemented at MGH, currently in use for (but not limited to) the workflow management of the proton treatment facilities. CorTx-RT is a service-oriented information architecture utilizing an enterprise system bus (ESB) to orchestrate workflow activities concerning all aspects of radiotherapy patient care. Using modelled workflows based on IHE-RO definitions and standard communication interfaces (e.g. HL7, DICOM, REST) the ESB provides clinical context for all data as workflows are executed. We deconstruct the functional environment of the radiation oncology computational environment such that the individual functions may be presented as web services in this service-oriented environment. Utilizing this service model, the user may define workflows mapping execution along with data flow. The data identified by context is stored by reference in a DICOM-RT 2nd generation data model where deeper data relationships may now be modelled through the use of the RT Course framework. This context-based data repository is stored in a non-SQL key-value pair transactional database that allows mapping of data ontologies through an object interface programming model. As such, the system may be used as a platform for big data research efforts where the data lake is exposed through defined interfaces.

ABSTRACT. Introduction Distributed learning can only be done if different sets of data in different centers are stored in the same format. Mapping of the local data into the desired format is not trivial. Using the right tools is important for cost effective mapping. The aim of this study was to evaluate the tools used in the implementation of the ontologies for two similar distributed learning projects. Materials & Methods In both projects data was sourced from multiple clinical databases. The data preparation step both anonymized, collated and stored the data in a database designed to efficiently fit the target ontology. The two methods differed mainly in preparation of the data and the RDF mapper. One used Pentaho and D2RQ whereas the other used Python and Ontop. Results Both methods resulted in a working database accessible for SPARQL requests. The greatest difference was in the ease of development. D2RQ has a complicated build-process which slowed down development of Method 1. The D2RQ syntax is difficult to read which impedes development and maintenance. The configuration files for Ontop are simpler to edit and test manually, compared to D2RQ. The configuration code is split into files and the mapping files are split into sections corresponding directly to an entity in the resulting knowledge graph. This division makes it intuitive to both develop and maintain the system. Discussion & Conclusions Both methods have resulted in usable mappings but using Ontop greatly simplified the implementation of mapping of data between the interim database and RDF, while the data preparation step, in Python, used a more familiar technology. In this study the use of Python, PostgreSQL and Ontop was more straight forward and intuitive, compared to the other set of tools. Therefore these are the tools recommended in this study.

ABSTRACT. The Veterans Health Affairs has designed a Health Information Gateway and Exchange (HINGE) to facilitate big data analytics in radiation oncology at an enterprise level. Despite the availability of copious amounts of structured radiotherapy planning and delivery data, complimentary clinical data is often unstructured and not conducive to data aggregation. The existing IT systems in radiotherapy workflow such as electronic medical records (EMR), radiotherapy treatment management (RTMS) and planning systems (TPS) are rarely interoperable. Furthermore, lack of standardization in clinical workflow, radiotherapy data taxonomy/ontologies, and variability in the structure of clinical notes makes data parsing difficult; thereby limiting the scope of big data analytics. HINGE is a web-based electronic data capture system which is connected with EMR, RTMS and TPS with a specific focus to engender big data analytics in radiation oncology. It has following features; (i) data farming: disease-specific discrete clinical data entry templates designed to capture data elements in sequential radiotherapy workflow from initial consult to follow-up care, (ii) portable architecture and RESTful services: the integrated portal is oblivious to the underlying specificity of the clinical systems and follows service-oriented architectural design, (iii) JSON-based APIs and FHIR extensibility: all the treatment data is stored and transported as documents in JSON format within the HINGE application via HTTP protocol. This also allows for authorized third-party application to plug in seamlessly (iv) built-in DICOM RT module: processing of DICOM files adhering to TG-263 nomenclature that allows DVH analysis (v) customizable workflow setup: since every clinic has its own preferred workflow set-up, HINGE design allows for dynamic UI rendering allowing multiple data entry template choices and customization. The goal of HINGE is to prospectively aggregate a comprehensive set of radiotherapy data for all patients treated in VHA.

ABSTRACT. With advancement in technology and our understanding of oncology, treatment options and medications are ever evolving. With this, the amount of data generated within radiation oncology is enormously increasing. Artificial Intelligence and machine learning models train from these data and act as an aid to the clinicians in making better treatment choice thereby increasing the chance of survival of the patient. The current data management and representation seems inadequate mainly because of lack of semantic and syntactic interoperability. Traditional EMR based and csv based data storage are stuck within the organizational boundary where it has been generated. Such data cannot be shared directly to other parties both for clinical and research purposes. As an alternative and sustainable choice, we demonstrated the possibility of FHIR as a globally accepted standard for representing data in radiation oncology and oncology in general. The rich information model coupled with the simple implementation of FHIR makes it suitable for research and clinical practice both. FHIR represents data as resources, identified by URI’s and communicated using HTTP based REST-ful protocols. However, there are no dedicated resources for representing oncology data in FHIR yet. We used a public radiotherapy data set as a reference data set to convert into FHIR. https://github.com/AnanyaCN/FHIR_Radiotherapy. A command line tool is developed to convert the 548 patient records from csv to fhir resource instances compatible to the latest version of FHIR. The resources are hosted in a hapi-fhir dstu3 server hosted in the locally and test queries were executed. The implementation is a proof of concept and future improvements are required for achieving maturity of the profiles. In general, we think we can achieve the same interoperability level for structured data by using FHIR, as with DICOM: we know the standard is not perfect, but it works for most common use cases.

ABSTRACT. The increasing digitization of patient information in electronic health records, coupled with sophisticated predictive algorithms based on machine learning methods are revolutionizing healthcare delivery via adverse event prediction and clinical decision support tools.

However, many complex decision problems in healthcare settings are best phrased as a choice between competing hypothetical actions, based on retrospective data where past actions were not chosen randomly, and outcomes are systematically censored.  Addressing problems of this type requires methodological tools from causal inference and missing data.

I will describe some of my recent work on using causal inference tools in healthcare settings.  In particular, I will show that informative predictors of an outcome (in our case ICU readmission of surgical patients) do not necessarily correspond to significant causes, discuss how analysis of causal pathways can be used to understand the impact of treatment adherence in outcomes of HIV patients, and outline how causal dimension reduction strategies can help understand complex high dimensional treatments that arise in radiation oncology.

ABSTRACT. An ART method based on relative gamma analysis and patient fraction classification capable, after a few fractions, to predict if a patient is likely to have frequent gas pockets or not during the course of treatment, in order to adjust the PTV margin. This retrospective study includes 47 patients treated for prostate cancer for which daily EPID images were acquired and analyzed using the γ-index analysis relative to the first fraction. The first step was to correlate gas pockets detection with γ-parameters using ROC curves. The second step uses γ-parameters, namely the largest connected pixel area with a γ > 1 (Reg1), mean γ (avg-g), standard deviation, and the Top 1% γ, to define patient states using statistics from a k-means clustering analysis. The patients were then categorized into “favorable” and “problematic” depending on the number of fractions in which they had gases. Finally, a program able to appropriately categorize a patient as problematic or favorable using the state information of the first few fractions was written and tested with ROC curves. Correlation between gas pockets detection and γ-parameters shows a good sensitivity and specificity for the Top 1% γ and for Reg1. Three states, which are tied to a level of gas volume in the patient for a given fraction, were defined from 1 for no gas pockets to 3 for major gas pockets. Using the prediction algorithm with the information of the first few treatment fractions allows one to classify a patient into one of the categories and adapt treatment when appropriate. The correlation between gas pockets and γ-parameters confirmed the performance of the method to detect gas. Analysis of daily EPID images provides a method to identify prostate patients that would have little or no gas pockets and allows for margin adaptation when deemed necessary.

ABSTRACT. Contours are used in radiotherapy treatment planning to identify regions to be irradiated with high dose and regions to be spared. Therefore, any contouring uncertainty influences the whole treatment. Even though this is the biggest remaining source of uncertainty when daily IGRT or adaptation is used, it has not been accounted for quantitatively in treatment planning. Using probabilistic planning allows to directly account for contouring uncertainties in plan optimisation. The first step is to create an algorithm that can generate many realistic contours with variation matching actual inter-observer variation.

We propose a methodology to generate random contours, based on measured spatial inter-observer variation, IOV, and a single parameter that controls its geometrical dependency: σ, the width of the 3D Gaussian used as point spread function (PSF). We used a level set formulation of the median shape, with the level set function defined as the signed distance transform. To create a new contour, we added the median level set and a noise map which was weighted with the IOV map and then convolved with the PSF. Thresholding the level set function reconstructs the newly generated contour.

We used data from 18 patients from the golden atlas, consisting of five prostate delineations on T2-w MRI scans. To evaluate the similarity between the contours, we calculated the maximum distance to agreement to the median shape (maxDTA), and the minimum dose of the contours using an ideal dose distribution. We used the two-sample Kolmogorov-Smirnov test to compare the distributions for maxDTA and minDose between the generated and manually delineated contours.

Only σ=0.75 cm produced maxDTA and minDose distributions that were not significantly different from the manually delineated structures. Accounting for the PSF is essential to correctly simulate inter-observer variation. The first step to incorporate contour variations directly into treatment optimisation has been taken.

ABSTRACT. The aim of this work is to use a simulation environment to evaluate the potential benefits of using photon-counting CT (PCCT) against dual-energy CT (DECT) in the context of quantitative contrast-enhanced CT for radiotherapy. An adaptation of Bayesian eigentissue decomposition (a tissue characterization method formulated in a maximum a posteriori scheme for noisy spectral data) which incorporates the estimation of contrast agent fractions and virtual non-contrast (VNC) parameters is proposed and validated.

PCCT and DECT are compared using two simulation frameworks: one including ideal CT numbers with image-based Gaussian noise and another defined as a virtual patient with projection-based noise and beam hardening artifacts, both scenarios considering spectral distortion for PCCT. Modalities are compared for their accuracy in estimating four key physical parameters: 1) the contrast agent fraction, and VNC parameters relevant to radiotherapy such as 2) the electron density, 3) the proton stopping power and 4) the photon linear attenuation coefficient. Considering both simulation frameworks, a reduction of root mean square errors is observed for all parameters with PCCT, with the exception of the accuracy on the contrast agent fraction being about constant through modalities in the virtual patient. Notably, root mean square errors on VNC electron density and stopping power are respectively reduced from 2.2 to 1.4% and 3.1 to 1.4% when going from DECT to PCCT with 4 energy bins. This increase in accuracy is comparable to the differences between contrast-enhanced and non-contrast DECT.

Overall, this study suggests that in a realistic simulation framework, using a maximum a posteriori material decomposition technique, the overall accuracy of radiotherapy-related parameters can be increased when using PCCT instead of DECT. This confirms the potential of PCCT to provide robust and quantitative tissue parameters for contrast-enhanced CT required in radiotherapy applications.

ABSTRACT. Objective: We present a novel stopping criterion to terminate iterative Maximum-Likelihood Expectation-Maximization (MLEM) unfolding of secondary neutron fluence spectra in radiotherapy.

Background: Secondary neutrons produced in radiotherapy pose a carcinogenic risk to patients. Specialized detectors like the Nested Neutron Spectrometer (NNS) measure a set of neutron CPS that must be unfolded to determine the underlying neutron fluence spectrum. Using MLEM to unfold such spectra requires a stopping criterion that terminates unfolding after sufficient solution convergence but prior to accumulation of noise in the unfolded spectrum.

Methodology: First, we implemented the published MLEM-STOP method used in PET image reconstruction. In this method, unfolding is terminated when subsequent iterations would only reconstruct the Poisson uncertainty inherent to the measurements. This is determined by identifying the iteration at which an indicator function, J, falls below a constant J-threshold. When applied to our radiotherapy NNS measurements, we found that unfolding was terminated too early for low CPS measurements and was never terminated for high CPS measurements. We developed a modified approach using four NNS datasets with known ground-truth spectra (Monte Carlo simulated spectra) to determine the optimal measured CPS at which the published MLEM-STOP method applied best to unfolding neutron spectra in radiotherapy. Our modified method dynamically scales the J-threshold for every dataset by the ratio of the average measured CPS to the optimal CPS.

Results: We used our modified MLEM-STOP method to unfold spectra for the four datasets with known ground-truth. All four spectra compared favourably to the corresponding ground-truths and ideal unfolded spectra in terms of noise content and solution convergence. Promising results were also obtained for other datasets without a known ground-truth.

Conclusions: Use of our modified MLEM-STOP increases confidence in unfolded neutron spectra in radiotherapy, which is important for accurate assessment of risk posed by neutrons to patients.

ABSTRACT. We develop and clinically implement a fully automated approach to intensity modulated radiation therapy (IMRT) treatment planning. The in-house-developed optimization algorithm is integrated with a commercial treatment planning system using API scripting, allowing the proposed algorithm to become part of our clinical workflow. The algorithm is based on a well-known hierarchical constrained optimization technique and is referred to internally at our institution as ECHO. It consists of three constrained optimization steps to maximize target coverage, minimize OAR doses, and smooth out the fluence map, respectively. A novel scatter-dose correction technique is also introduced which speeds up the optimization process significantly (about 12 times) and enables the clinical implementation of ECHO. We show that the scatter-dose obtained from open-field dose calculation multiplied by the average beamlet intensities is a very good approximation of the scatter-dose from optimal beamlet intensities. Therefore, insignificant scattering contributions can be neglected in the influence matrix and compensated with the open-field scatter correction term, resulting in a sparse and computationally efficient truncated influence matrix for optimization process. ECHO has been used in our clinic to deliver more than 700 SBRT paraspinal plans with all the plans meeting all the institutional clinical criteria. The median time to generate one ECHO plan was 63.5 minutes, including the influence matrix calculation and optimization.

ABSTRACT. The 1.5T MR-linac has recently become clinically available. This system, which combines a 1.5T MRI scanner and a 7MV linear accelerator, is able to provide MR-images for online RT guidance. This allows for multiple strategies to perform online plan adaptation. Our aim was to report on the different plan adaptation strategies for different treatment sites to give a concise overview of the available options. Performing plan adaptation is possible using the adapt to position (ATP) workflow, which allows for correction for the new patient position, or the adapt to shape (ATS) workflow, which allows for plan adaptation based on the new patient anatomy. Using the ATS workflow, the plan is optimized on the daily MRI and adapted contours. Four and six different plan adaptation methods are available for ATP and ATS, respectively. To explore the different methods, we have simulated online plan adaptation for five different cases with varying characteristics: prostate, rectum, oesophagus and lymph node oligometastases (single and multiple target). Reference plans were created on a pre-treatment CT which were used for adapting the plan to the daily anatomy, as visible on MR-imaging. For the ATS workflow, the patient contours were manually adapted by a radiation oncologist. The plans were evaluated based on the clinical dose constraints and the optimization time was measured. Our results show that he time needed for plan adaptation ranged between the 17 and 485 seconds. More advanced plan adaptation methods generally resulted in more plans that met the clinical dose criteria. The most advanced method in the ATS workflow, in which a plan is fully re-optimized based on the daily contour information, is feasible within a timeframe acceptable for online treatment and dosimetrically yields the best plans. Consequently, for all of the above treatment sites, we have applied this method for clinical MR-linac treatment.

ABSTRACT. Purpose: To assess the overall change of plan quality in terms of dosimetric parameters in patient cohorts before and after the clinical introduction of knowledge-based planning (KBP) across multiple disease sites.

Materials/Methods: Human planners created the clinical plans in the “pre-KBP” group (n=181); retrospective re-planning with KBP was employed for this group to evaluate any unrealized plan quality, typically manifesting as under-spared organs-at-risk (OAR). In the “post-KBP” group (n=503), a KBP plan was created first, and subsequently manually refined at planner’s discretion to create the clinical plan. Prostate, prostatic fossa, left/right lung SBRT and head-and-neck were investigated. Plan quality in “pre-KBP” and “post-KBP” groups was assessed using site-specific DVH parameters (normalized to target dose), comparing OAR values for clinical and KBP plans, i.e. ∆Dx,pre vs ∆Dx,post with ∆Dx=Dx,clinical-Dx,KBP. The direct comparison of both plans mitigates the variability in patient anatomy and is interpreted as unrealized plan quality. Thus, the average of ∆Dx among all patients yields a robust measure of aggregated plan quality and variability in a cohort. PTV DVH parameters were measured as absolute average of each group. Significance testing between “pre-KBP” and “post-KBP” utilized unpaired two-sided t-tests (p<0.001).

Results: Excess OAR dose (∆Dx,pre>0) was significantly reduced post-KBP in prostate plans for bladder V40Gy (2.2%/-0.7% for “pre-KBP” and “post-KBP” respectively), penile bulb Dmean (6.5%/1.2%) and rectum V40Gy (4.7%/-0.8%), V65Gy (1.4%/0.3%) and V75Gy (1.0%/-0.1%). Prostatic fossa showed significant improvements in the same organs. No significant OAR changes were observed in lung SBRT. Head-and-neck showed significant decrease of ∆Dx in 7 organs, with largest differences in Cricopharyngeus Dmean (21.5%/0.7%), Larynx Dmean (8.9%/0.5%), and Esophagus Dmean (7.0%/0.4%). Changes in PTV DVH parameters were minimal, except an increase D1% for lung SBRT at explicit physicians’ request.

Conclusion: The widespread adoption of KBP decreased excess organ-at-risk dose and variability in most disease sites.

ABSTRACT. Introduction It is well known that the inverse planning problem in radiation therapy can be expressed as a linear optimization problem, i.e., as the minimum of a linear function subject to linear inequality constraints. Although stable algorithms that solve such problems to proven optimality exist, they are not used for radiation therapy because they are too slow and use too much memory. This talk is about an innovative and fast implementation of a proven algorithm, intended for clinical use at the proton center of Massachusetts General Hospital (MGH).

Materials & Methods An Interior point method (IPM) is an acknowledged algorithm for linear optimization. The core of an IPM is a square linear system whose dimension depends on the number of voxels. This dependence is the result of mean overdose constraints, where overdose is defined as the dose above either a predefined threshold or a dose-volume histogram point. We show that the linear system has a block diagonal plus low rank structure. A state-of-the-art IPM which exploits that structure is implemented in C++ and integrated in the in-house treatment planning system of MGH. It supports all known linear inverse planning formulations as well as robust optimization. During its execution it shows how much improvement is still possible, so it can be terminated early when the remaining potential is less than, e.g., 0.1 Gy.

Results The new IPM has successfully optimized over two hundred test cases. It is several times faster per iteration than the commercial alternative CPLEX, which does not recognize the special structure. The early stopping feature offers an additional 30% speedup. Early clinical tests show favorable results.

ABSTRACT. Background and Purpose: In radiation therapy treatment modality, the complete system demands both machine and patient specific quality assurance (QA) exclusively. This study has discussed only patient specific QA, keeping the objective of optimum utilization of therapeutic equipment’s with minimum error. The purpose of the study is to evaluate the deviation of calculated dose in TPS, and actual dose delivered to the target volume in IMRT technique. Method and Materials: In this study 6 MV photon beams of Clinac iX (Varian) has been used for the irradiation of cervical cancer. Treatment Planning System (TPS) has been performed through Eclipse -13.7 and the corresponding dose was calculated accordingly. The TPS calculated dose has been verified with EPID and Matrix with Multicube Phantom individually. Results: In most of the cases the prescribed dose is 50 Gy for PTV intermediate and 62.5 Gy to PTV High. It is observed that 95% of the PTV (High) volume received 95% to 97% doses with better dose conformity to the target and sparing of the surrounding structures. The average area gamma variation for EPID and Matrix are 2.017% and 1.533% respectively which is significantly lower than the recommended reference values of IAEA (3%) reports. The structure of tumor varies with the progress of cure of disease and individual organ acquires its original shape. In course of treatment delivery period OAR become closer to the PTV. Thus due attention has to be paid for 2nd planning within treatment delivery period to ensure QA. Conclusion: The comparison system of each treatment modality must have some arrangement which can notice the deviation of radiation delivery in same dimension. Three dimensional dolphin can be used instead of two dimensional EPID and Matrix as comparative means. The patient specific QA is mandatory before clinical implementation of IMRT treatment delivery.

ABSTRACT. This study investigates the role of radiomics for prediction of extracapsular nodal extension (ECE) from diagnostic CT images of the involved lymph nodes (LNs) in head/neck cancer patients. Twenty patients (10 ECE+ and 10 ECE-) with head/neck cancer were studied. Oral cavities of all the patients were surgically resected and 1-3 stage of involvment for their LN was pathologically confirmed. Pre-surgical contrast-enhanced computed tomography (CE-CT) images of positive LNs (within 2-months prior to surgery) were contoured to include the resected LN with a margin of 4 mm. One-hundred-seventy-two radiomics features from 9 different feature categories were extracted from CE-CT images of the LNs. Levene and Kolmogorov-Smirnov’s tests with absolute-biserial-correlation (>0.45) were used to reveal significant associations between ECE status and radiomics features. Ultimately, either ANOVA or Welch’s test was used as the test of significance between the two groups. Among 172 radiomics features only three features (1 Energy and 2 Entropy-based features) from 2 feature categories (LAW’s and 2D-Wavelet-Transform) were significantly (p-value<0.05) different between the two groups. The percentage of relative mean difference between the radiomics features of the ECE+ and ECE- groups (100*(µn- µp)/ µn) for the LAW’s textural energy and the two entropy features of 2DWT were: - 65.37%, + 319.05%, and 659.87% respectively. Results of this pilot study suggest that the CE-CT images of LNs with more spotty patterns can be associated with ECE+ and therefore a higher chance for cell metastases beyond the capsule of involved LNs. Results also imply that higher levels of disorder (Entropy) of high frequency information of vertical-components in CE-CT images may be associated with increasing chance of ECE. This pilot study, albeit subject to confirmation in a larger patient population, indicate a potential role for the use of radiomics-based signatures in models developed for predicting ECE+ in patients with H/N-Squamous-Cell-Carcinoma.

ABSTRACT. The purpose of this work is to investigate the efficacy and capability of a deep convolutional neural network (DCNN) in the classification of eight categories of dermoscopy skin lesion images. ResNet 152 deep learning algorithm which exceeded the human performance on ImageNet challenge in 2015, is applied on total 10135 dermoscopy skin images include 10015 images from HAM10000 database, and 120 images from PH2 database. This experiment includes eight major diagnostic skin disease categories - melanoma, melanocytic nevi, basal cell carcinoma, benign keratosis, actinic keratosis and intraepithelial carcinoma, dermatofibroma, vascular lesions, and atypical nevi. The results of the ResNet 152 are compared with the performance of highly trained and experienced dermatologists. Our results show that, DCNN model outperformed dermatologists with at least an 11% margin. The best ROC AUC scores are achieved 94.40% and 99.10% for melanoma and basal cell carcinoma, respectively. Whereas dermatologists achieved 82.26% for melanoma and 88.82% for basal cell carcinoma. Also, the DCNN model can achieve very high AUC scores for each individual diagnostic category.

ABSTRACT. Picture Archiving and Communication Systems (PACS) used in clinical workflows are rich with data in the form of images as well as associated annotations, segmentations and finding reports by specialists in radiology, nuclear medicine or radiation oncology. From a research perspective, annotated images constitute the raw material for radiomic, machine learning and similar AI-based initiatives. It is not straightforward however to harness the potential of the tremendous amount of data collected in PACS on a daily basis. Research infrastructures must be designed to interfere minimally with clinical workflows and to provide mechanisms to prevent unauthorized access to images. Furthermore, these infrastructure must provide user-friendly interfaces for data operations since raw images and annotations are rarely suitable for direct ingestion by downstream analysis pipelines. At our institution, an imaging research infrastructure has been deployed that addresses these performance, confidentiality and conviviality issues. The solution is based on Orthanc, an open-source and lightweight DICOM server acting as a controlled gateway to clinical PACS. Functionalities are exposed to users through a REST API, who can perform several operations (query, upload, receive, transfer) in the programming language of their choice. The REST API is particularly useful to easily perform queries, browse metadata and parse objects such as DICOM Structured Reports (SR). In this paper, the architecture of the solution is presented in terms of hardware, software and network topology. Three use cases are also detailed, each addressing different needs: a simple retrieve operation to physically access images, a data pipeline solution for automatic calculation of image-derived metrics, and an automated parser used to find occurence of specific terms in radiology finding reports. Our experience shows that relying on a DICOM-compliant software enforces good practices and fosters standard awareness, with benefits in terms of interoperability and long-term usability of data.

ABSTRACT. Purpose- Iodine signal enhancement in tissues, expressed with respect to relative electron density RED, is relatively small due to the inherent low sensitivity of single energy CT (SECT) in differentiating between iodine and other materials. Dual energy CT (DECT) has the ability to better differentiate between materials with similar RED values, but different effective atomic numbers (Zeff). This work aims to combine DECT material differentiation, iodine parameterization, and CT perfusion to develop a technique to overcome the current limitations of CT perfusion and improve the accuracy of imaging-derived pharmacokinetic parameters. Materials and Methods- DECT scans of phantoms containing 0-15 mgI/mL of iodinated contrast agent were acquired (Siemens Somatom Definition Flash) at various energy combinations. Values of RED and Zeff were calculated in each voxel using the spectral method and stoichiometric parameterization. Response of the DECT scanner to iodine was calibrated through parametrization of linear attenuation coefficient. The calibrated response was used to create a map of Zeff for quantification of iodine concentrations and compared to SECT quantification. Results- DECT iodine quantification using the Zeff calibration spectral method resulted in an average error of 3.6% across the concentrations in phantoms, compared to 16.2% in SECT. Calibration using both Zeff and RED led to an average quantification error of 4.4%. Stoichiometric calibration did not yield any improvement over other methods of calibration. Detection accuracy of iodine did not vary significantly with tube potential, but was reduced with decreasing tube current. Conclusions- Using DECT and calibration with respect to Zeff leads to improved accuracy in iodine quantification, especially in the low concentration range. This will be used in clinical perfusion scans to redefine contrast enhancement curves with respect to iodine concentration and validated in a head and neck cancer clinical trial to assess their efficacy in improving quantitative analysis of perfusion parametric maps.

ABSTRACT. Purpose: Synthetic CT (sCT) was required in MR-only treatment planning for dose calculation. Materials and Methods: In training phase, six prostate patients’ planning CT and simulation MR images were acquired at same day, and CT images were deformed to MR. The body region was categorized into two regions of the bone (R1) and other organs (R2), while bone contour was drawn manually. The exponential relationship between MR and deformed CT (dCT) in R1 and R2 with high homogeneity was obtained. For validation purpose, three prostate patients’ CT and MR images were used. The synthetic CT was generated by applying the exponential curve across R1 and R2 in MR. Treatment planning was established by prescribing 70 Gy with 28 fractions by using MRIdian treatment planning system with step-and-shoot intensity-modulation radiation therapy (IMRT) technique. Dose distribution was recalculated by replacing dCT with sCT. Image quality was assessed in terms of noise and smoothing level, and dose volumetric parameters with dCT and sCT were compared in planning target volume, rectum, and bladder. Results: The sCT showed comparable image quality with dCT showing similar smoothing level, and the significant noise reduction (137% on average). In dose-volumetric comparison between dCT and sCT, the dose received by at least 95% (D95%), D5%, the mean dose (Dmean) of the planning target volume (PTV) increased by 0.36, 0.02, 0.37 Gy, and no change was observed in HI, on average. The percent volume receiving 65 Gy (V65Gy), and V58Gy of the rectum, and V65Gy of the bladder increased by 0.04%, 0.33%, and 0.22%, respectively. Conclusions: While several misrepresentations in air cavity and bone was observed, the overall image quality was moderately preserved. Furthermore, the dose-volumetric parameters in PTV, rectum, and bladder showed no distinct differences between dCT and sCT.

ABSTRACT. Purpose: Accurate dose calculation is of vital importance for proton therapy. Monte Carlo (MC) dose calculation is recognized as the most accurate method. Deep learning methods, due to their superior speed, can now improve dose calculation accuracy by converting the pencil beam dose computed by treatment planning system (TPS) to mimicked MC dose. This work aims to achieve this goal by developing a deep learning model that can precisely predict MC dose from patient CT image and proton pencil beam dose for different types of cancer. Methods: The proposed model is based on our newly developed hierarchically densely connected U-Net (HD U-Net) deep learning architecture. The model includes two input channels: the pencil beam dose and the patient CT image. Both the pencil beam dose and the patient CT image were normalized before used for training. A total number of 202 patients (76 prostate patients, 94 liver patients and 32 lung patients) were randomly chose for training/validation and testing. The accuracy of the model was evaluated by comparing mean square error (MSE) and gamma index between the predict and the true MC dose distributions. Results: Using one NVIDIA Tesla K80 card, the trained model takes less than 5 seconds to predict the MC dose. Compared with pencil beam doses, MSE between the predict and the true MC doses are improved by at least 64% in most prostate test cases. The gamma index results show good agreements between the predict and the true MC dose. The 3D gamma passing rates of all prostate test set with 1%/1 mm criterion are over 89%. Conclusions: The developed model can accurately predict MC dose distributions form pencil beam dose distributions and patient CT image. This model can be an efficient and practical tool to improve the accuracy of proton dose calculation of TPS.

ABSTRACT. Multi-centre data collection outside of clinical trials is an increasingly important part of radiotherapy as institutions work to audit each other and compare outcomes on a routine basis. When these systems are automated it is necessary to be able to compare structure outlines and dose volume metrics based on those outlines without human intervention. Variation in outlining nomenclature means that this can be challenging as small differences in naming convention can cause text matching to fail or give incorrect answers. In this work we consider using properties of the structure shape and voxel value distribution to identify the label rather than its name. A multiplayer perceptron neural network with a single hidden layer was constructed to predict the true label of structures using their shape and textural properties. The model was trained using publicly available datasets. We found that using this relatively simple topology it was possible to achieve efficiencies and purities of better than 90%. The network was trained using 85% of the data and then tested with the remaining 15%, this was repeated by random resampling of the data. This suggests that it may be feasible to identify structures reliably from their shape and intensity values. Work is continuing to apply the model to clinical data sets from other sources to validate the model away from the training data and tune the model further.

ABSTRACT. To implement an independent 3D dose engine developed within the exit detector Delivery Quality Assurance(DQA) tool for helical Tomotherapy(HT). Previous DQA tool used standalone HT dose engine (Method 1) for dose reconstruction on patient CT data. The 3D independent dose calculation engine (Method 2) was implemented within DQA tool verification in place of standalone dose engine. The dose reconstruction accuracy was verified with ion chamber measurements at several points inside the Cheese phantom for clinical representative plan. A 3-D dose evaluation is then conducted using 3%/3 mm Gamma index (global) for dose comparison between Method1 and Method 2. We also compared the ratio of dose volume histogram (DVH) metrics computed for the planned to delivered dose for each of the dose reconstruction methods. Finally, the total time required for performing DQA using each dose calculation methods was also compared. Agreement within 2% was seen between the ion chamber measured dose and the planned dose calculated by the tool using both methods at several points inside the Cheese phantom. The average gamma pass rate (3%/3mm) from 30 clinical plans DQA measurements was 99.6±0.97 and 99.8±0.2 performed using dose reconstruction method 1 and method 2 respectively. The average ratio DVH metrics between planned to delivered dose from both dose reconstruction methods was 1±0.02% and 0.99±0.03% for targets and OARs structures respectively for both methods. The average time required for dose calculation was 5±1.5 min and 12.5±2.5 min for method 1 and method 2 respectively. Comprehensive clinical validation to implement an independent 3D dose calculation dose engine within our in-house exit detector-based dose reconstruction tool for HT DQA has been undertaken, demonstrating equivalence to the stand alone dose calculation engine. This will extend the utilisation of this tool to other clinics that do not have an access to standalone HT dose engine

ABSTRACT. The evolution of radiotherapy treatments sees the increasing use of intensity modulated conformal radiotherapy (IMRT) techniques, via their classic form of delivery ("step and shoot" and "sliding window" mode) and dynamic arc therapy. These complex method are demanding in terms of quality assurance and preparation time because they require systematic dosimetric checks for each treatment plan before it is delivered to ensure that it is correctly carried out. We propose here a solution to optimize the practices of a radiotherapy department using metrics to quantify the complexity of a treatment plan in order to assess the relevance of the need for systematic quality control upstream. The goal of this study was then the development of a decision support software to provide information on the chances of a dynamic treatment plan to pass the quality controls. According to the result of this prediction, the physicist would decide the relevance of a QA for a given plan (currently in France each plan is systematically controlled before treatment). We have developed a software prototype based on the one hand on the analysis of DICOM RTPlan files for complexity indices calculation, and on the other hand on the QA results of these same plans, provided by clinical centers, and acting as label for the learning stage. The outcome is a model that in turn provides QA prediction. First results are promising.

ABSTRACT. Artificial neural networks (ANN) methods applied to external beam radiation therapy, can be of great interest, especially to improve the efficiency of quality assurance (QA) based on electronic portal imaging device (EPID). In this work, the application of an ANN algorithm is proposed to predict 2D reconstructed absorbed dose distributions based on EPID images, acquired during treatment delivery on Varian and Elekta linacs. A supervised ANN algorithm consisting of a non-recurrent feed-forward multilayer model was employed. The algorithm was trained (learning phase) with input and output data sets taken from EPID images, and treatment planning system (TPS) absorbed dose distributions, respectively. New EPID images were then used to predict the delivered absorbed dose distributions. EPID images were taken from an aSi-1000 on a Varian Clinac 23iX® and an iViewGT on Elekta Synergy® during 6MV treatment delivery. 2D absorbed dose distributions of intensity modulated radiation therapy (IMRT) plans were calculated in EclipseTM and PinnacleTM, at the maximum depth dose in a water phantom. The setting and the behavior of EPID complexity for the QA step were correctly learned by the ANN algorithm creating a pattern by minimizing the error on a maximum sample of data. Global gamma passing rates (2%/2mm) > 98% were obtained for the evaluated cases from Varian and Elekta instances, which showed the ANN capability to predict the reconstructed absorbed dose distributions for IMRT verification, based on EPID information and for different linacs

ABSTRACT. The aim of this paper is to present the lessons learned from developing PlanChecker, a tool for automated quality control of plan transfer from the treatment planning system (TPS) and record & verification system. The plan checker was started in 2011 as a single-physicist project with the initial intention of replacing time-consuming and error-prone manual checking of the plan parameter transfer, especially for advanced radiotherapy plan. During cycling of testing-bug-debugging, it was found hard to trace up and fix the bugs due to the bad design of code and program structure. In 2013, the program was redesigned and re-programmed using Python using class-based modular design. It was restructured into the following modules: input-module, a comparison module, a logical module and output module and graphic user interface. From 2014, version control of this tool was introduced using Github, which enables to switch the developing mode from single physicist to multi-physicists. The bug reporting was using JIRA, which provides a quick way of reporting and fixing the issues. Quality control of the software was initiated by performing the unit-test each class, module and function and end-user the testing. PlanChecker has been implemented into our clinical workflow and deployed as a Citrix application across two centres since 2014. On average, about 20 new patients’ plans are checked on a daily basis with this tool, not only reducing the risk of patient mistreatment but also improving the efficiency of our clinical workflow. Good design prior to starting to write the code, version control, good bug-reporting tool and quality control of the software is critical for making the in-house tools sustainable, extensible and maintainable, especially for the developers without formal training in computer science and software engineering.

ABSTRACT. The division of radiation oncology at Peter MacCallum Cancer Centre operates 16 linear accelerators across 5 physical sites. In 2018 a new cloud based planning system (Varian Eclipse 15.5) was commissioned with the view to standardise and update treatment planning procedures across the whole organisation. Planning for the project began mid 2015 with a decision reached after a tender process in 2017. A project manager with IT background was appointed and two medical physicists and two radiation therapists were dedicated full time to the project. A multidisciplinary steering group was formed and several working groups created. The project consisted of several components including: • Assessment, acquisition (where needed) and standardisation of all physics parameters for all imaging equipment and treatment units • Compilation, updating and documentation of planning procedures for 12 tumour streams • Training of staff across five campuses by the vendor and development of in-house training resources • Merging of the two largest data bases and consolidation of a total of 9 databases • Creation of test plans for some 100 patients using many treatment scenarios and a variety of techniques • Audit and quality assurance procedures for the system Several new features were introduced including modulated treatments for flattening filter free beams and the use of Acuros with dose to medium specification. The planning system went live as scheduled in three campuses on Sep 24, 2018 with the other two campuses following in the next weeks. An external audit was performed and patient specific measurements done for all patients with intensity modulated plans in the first three months after introduction (> 700). Implementation of a single cloud based treatment planning system across five campuses has proved to be a complex and rewarding task the success of which heavily relied the full engagement of the multidisciplinary team.

ABSTRACT. Multileaf collimator (MLC) leaf position accuracy is crucial during radiotherapy. Small leaf position offsets can result in large dosimetric errors for patients. Picket fences are a quality assurance procedure used to assess leaf position accuracy. Monthly validation of picket fences is specifically recommended in the latest Canadian Partnership for Quality Radiotherapy technical quality control guidelines, and a tolerance of 0.5mm is recommended. However, quantifying picket fences is challenging due to changes in calibration, the wide variety of possible dose profiles across leaf gaps, stray pixels, and geometrical variance across linacs.

Recently, while commissioning a new linac with refined leaf gap control, the picket fence dose profile was found to exhibit significant peak-trough variability. Existing open source analysis tools were unable to quantify the gap and we were left unable to assess leaf position accuracy. To remediate this shortcoming, we created a novel non-parametric algorithm that can handle arbitrary peak-trough shapes, stray pixels, and detector-MLC rotation. It quantifies leaf gap separations using a statistically robust procedure based on multi-sampling of the dose profile without requiring open-field measurements.

An implementation suitable for analyzing digital picket fences was validated using an intentionally-bad picket fence and found to be sensitive to leaf position errors <0.25mm. Sensitivity to detector-MLC rotations is high and appears to be limited by the width of a single detector pixel. Three years of historical picket fences were analyzed retrospectively to characterize individual machines; we found that the age of the linac correlates with the magnitude of detector-MLC rotation correction needed, leaf gap tends to grow over time, and leaf position variations seem to coincide with MLC interventions.

In this work we give an overview of our algorithm, describe results of our historical investigation, and demonstrate the web application.

ABSTRACT. Introduction The accuracy of dose calculation algorithm (DCA) is highly regarded in the radiotherapy sequences. This study aims at assessing the accuracy of five dose calculation algorithms in tissue inhomogeneity corrections, based on the International Atomic Energy Agency TEC-DOC 1583. Materials & Methods A heterogeneous phantom was scanned using computed tomography and tests were planned on three-dimensional treatment planning systems (3D TPSs) based on IAEA TEC-DOC 1583. Doses were measured with 6 and 18 MV photon beams and then ion chambers and the deviation between measured and calculated TPS doses were reported. Evaluated five DCAs include: Monte Carlo (MC) algorithm employed by Monaco, pencil beam convolution (PBC) and anisotropic analytical (AAA) algorithms employed by Eclipse and Superposition (SP) and Clarkson algorithms employed by PCRT3D TPSs. Results In Clarkson algorithm, low and high energy photons indicated 7.1 % and 14.8 % deviations out of agreement criteria, respectively. SP, AAA, and PBC algorithms indicated 0.9 %, 7.4 % and 13.8 % for low energy photon and 9.5%, 21.3% and 23.2% for high energy photon deviations out of agreement criteria, respectively. While MC algorithm showed 1.8 % deviations at high-energy photons. Discussion & Conclusions The accuracy of DCAs in TPSs is different. Some simple DCAs such as Clarkson exhibits large deviations in some cases. Therefore the transition to more advanced algorithms such as MC would be desirable, particularly for calculation in the presence of inhomogeneity or high-energy beams. References:

2. TecDoc I. 1583: commissioning of radiotherapy treatment planning systems: testing for typical external beam treatment techniques. Vienna: International Atomic Energy Agency. 2008.

Acknowledgements This research was supported by Tehran University of Medical Sciences and Health grant number Services with grant number [33446].

ABSTRACT. For photon radiotherapy planning, analytical algorithms are used instead of the slow gold-standard (Monte-Carlo). These algorithms, i.e. Pencil beam, Anisotropic Analytical Algorithm (AAA), Collapsed Cone Convolution (CCC), yield good performances in most cases. Nevertheless, in case of highly modulated beams or many heterogeneous interfaces, the dose distribution computed with these algorithms can lead to a deviation from measurements or from Monte-Carlo simulations.

We propose a new approach using artificial neural networks to compute the dose distribution resulting from a clinical treatment plan. Our goal is to provide an alternative to the widely used analytical algorithms, which is accurate and fast enough for treatment planning quality control in Modulated Radiation Therapy (IMRT and Arctherapy). Several configurations of data in training sets and of number of neurons on hidden layer are evaluated. To create the learning sets, Monte-Carlo simulations are computed using BEAMnrc/EGSnrc. To select the best network, a comparison with Monte-Carlo dose distributions is performed using gamma index with 3%/3mm margins.

The obtained results for a square beam in a prostate area shows a good agreement with the Monte-Carlo simulation. Moreover, for a dose grid of 2.5 mm x 2.5 mm x 2.5 mm voxels, the computation time on a traditional computer (Intel i7) for a clinical IMRT case is only 2 minutes.

ABSTRACT. To facilitate VMAT utilisation across Genesis Care radiotherapy centres in Australia and UK, a working group was formed. It has been agreed that the current native Pinnacle scripts will be complemented with Python code stored on a separate server. The Pinnacle syntax has been subdivided into sections called "blocks" that were converted using Jinja templating language. The use of Jinja allows for a variable within the Pinnacle syntax to be substituted with a Python variable and only compiled when needed. The user designing a script would handle ROI, beam, prescription, optimisation objectives Jinja blocks. Then, the Python compiler would access the block's definitions stored as Pinnacle native language and populate with necessary variables. The compiled syntax is sent from the scripting server back to the Pinnacle server and executed. For debugging purposes, the user has access to the "Sentry" automatic Python emailing system and to a copy of the compiled Pinnacle syntax. The scripting working group has documented a quality assurance process of code release. A flowchart has been produced that allows users to determine additional tests before clinical code distribution e.g. plan deliverability on a linear accelerator The GIT platform has been implemented for scripting files version control. The Python compiler produces code that is linked to an approved static commit ID. Three access levels for users have been defined for additional security. A staged scripting rollout was performed across 32 Genesis Care Centres. It facilitated increase of VMAT treatments by 60%, ~1 day decrease in planning time and reduced planning errors once a centre has fully implemented the scripting process. The development of class solutions for several anatomical treatment sites has been translated further into a de-centralized planning room across three Australian states, where the time zone difference could be used as an advantage.

ABSTRACT. The Eclipse Scripting Application Programming Interface (API) allows bespoke functionality to be added to the Treatment Planning System (TPS) by creating software Scripts. We report on the use of Scripts to: 1. improve the quality of treatment plans by automating part of the Physics second check , 2. improve the Treatment Planning workflow efficiency by generating Dose Volume Histogram (DVH) analysis reports and also modifying Treatment Plan Reports to display data not available in standard reports, 3. automate parts of the TPS Quality Assurance program by generating depth dose and transverse profiles to compare with reference data, and, 4. carry out clinical research projects using data mining techniques to analyse treatment plan characteristics across the entire clinical database. We have found the Eclipse Scripting API to be an invaluable tool for tailoring the TPS to suit our clinical needs. We have found it can be used to improve treatment plan quality and planning workflow efficiency. It is extremely popular with all staff including Dosimetrists, Radiation Oncologists and Physicists.

ABSTRACT. Transrectal ultrasound (US) imaging is used during prostate high dose rate (HDR) radiation treatment to guide the insertion of catheters through the perineum into the prostate. Following a 3D US acquisition, the catheters are manually reconstructed in the treatment planning system defining potential dwell positions for the HDR source. We propose a novel approach to catheter reconstruction from transrectal US images through the use of a deep convolutional neural network combined with pre- and post-processing algorithms, collectively referred to as “DeepDwell”. Manually reconstructed catheters and transrectal US images were used to train, validate, and assess the performance of the proposed method. A total of 14 prostate patient datasets (12 training, 1 validation, 1 test) were used. The architecture of the neural network in DeepDwell was based on the ‘U-Net’ model with several modifications influenced by other developments in the field of image segmentation. This was followed with processing to allow for clustering, cleaning, and delineation of dwell positions to define the predicted catheters. An analysis comparing predicted to ground truth catheters was performed over all slices, evaluating centre-to-centre distance and similarity. DeepDwell was shown to quickly segment a majority of the catheters (13/16) in the test dataset. Of those detected, the average distance is within 0.3 ± 0.1 mm from the ground truth suggesting that the use of machine learning can be of clinical benefit in HDR brachytherapy. Future work will focus on improvements of detection and clinical exploration.

ABSTRACT. The Quality Check List (QCL) feature within MOSAIQ provides functionality to track a patient’s treatment planning progress. As it can be difficult to obtain an overview of outstanding and upcoming QCL items within MOSAIQ, we designed a customizable web-based dashboard providing such an overview. Our dashboard is designed to display appropriate views for all staff roles involved in treatment planning as well as on a shared screen in the planning room.

QCL data is updated from MOSAIQ using a set of SQL queries which run at regular intervals. This is displayed in a web frontend allowing users to customize their view based on their role or current needs. The dashboard application was validated by implementing automated unit and functional tests as well as performing manual test cases. A soft rollout of the dashboard within our clinic has further validated the application. Query times are logged to ensure minimal impact on the live MOSAIQ database.

Following the soft rollout, the QCL dashboard has shown promise in aiding staff members with planning their clinical tasks and to reduce delays in the treatment planning process. Determining the appropriate start date of a patient’s treatment is not always straightforward in scenarios such as replans or second phase treatments. An adjustment to the clinical process of booking patient treatments may be required, to ensure appropriate status codes are entered for bookings where a certain type of treatment will begin. Once the dashboard has been fully rolled out within the clinic, QCL timing data will be analysed to determine clinical impact of the application.

ABSTRACT. Intensity-modulated radiation therapy, rotational or helical delivery technologies were developed to focus on maximizing does to target volume while sparing healthy tissue. To achieve this, the dose distribution to surrounding healthy tissues is determined by organ type, physician constraints, and published normal tissue tolerance criteria. With defined, constraints, dosimetrists optimize each patient’s plan to meet these objectives. However, for unidentified organs or healthy tissue for which no explicit dosimetric constraints are defined optimization procedures might omit them, allowing high levels of dose to be unnecessarily deposited. To enhance plan development and reduce dose to healthy tissue an application was developed using Matlab that provides dosimetrists and physicians a means of outlining and working with an ideal patient plan. An initial dose is created from the prescription criteria and uniformly calculates as a homogeneous dose distribution to the target volume with uniform dose falloffs in all directions. Correction factors for heterogeneity are applied with the ability to manipulate dose in the vicinity that is correlated to OARs to obtain the most desirable plan. This technique maintains integral dose within the treatment volume when redistributing the dose and maintaining uniformity within the tumor volume. Current dose calculations take approximately 2 minutes while calculating a high resolution dose (CT voxel size), while dose OAR modification requires 1-4 seconds depending on the calculation region. Initial historical plan comparisons have shown the the capability of computing model based treatment plans that could be used to guide the optimization of the intensity modulated plans and potentially reach a better solution in less time than the original ones.

ABSTRACT. Stereotactic radiation body therapy (SBRT) treatments require the implementation of robust in-vivo dosimetry techniques, to which electronic portal imaging device (EPID) offer an attractive solution. However the use of EPIDs in such conditions is still challenging and is not completely supported for some EPID models and by commercial solutions. Therefore, the purpose of this work was to implement an automatic and independent characterisation of EPID continuous imaging, for further 3D verification purposes of SBRT treatments. Different image data sets were acquired with an aS1000 EPID installed on a Varian TrueBeam STx system, in continuous acquisition mode (CM) to fully characterise different dosimetric properties of EPID continuous acquisition when exposed to 6 and 10 MV FFF photon beams with maximum dose rates (1400 and 2400 MU/min). Images were also acquired in integrated mode (IM) to allow comparison of both modes. For an automated analysis and characterisation of EPID continuous imaging, in-house software was developed in Matlab allowing processing of all acquired images in different formats and obtaining direct comparisons of derived integrated images, from both acquisition modes. EPID dose response linearity repeatability and reproducibility, ghosting effect and field-size dependence were evaluated. The implementation of an automated characterisation process allowed preforming complete comparisons of both modes, for the different evaluated dosimetric properties. Comparable dose response linearity was obtained for static and arc fields. Field size dependence of EPID response in both modes agreed within 1%. Response repeatability and reproducibility were within 0.6% in al cases. Ghosting effect was found to be comparable to signal variations between continuous acquisition frames. An automated characterisation of EPID continuous imaging of FFF beams has been implemented and allowed a direct evaluation of relevant dosimetric properties, showing feasible results for further 3D dosimetry of SBRT.

ABSTRACT. Introduction Research has shown that pain is a significant problem for up to half of all cancer patients [1]. Pain is recognized to be the sixth vital sign by Canadian Partnership Against Cancer [2]. Various reports have shown that patients who have indicated their pain outcomes in real-time and have them acted upon have a better chance of survival compared to patients who did not [3].

While distress screening can help facilitate immediate improvements in current clinical care, new and emerging Artificial Intelligence (AI) technologies are starting to show promise for predicting in advance patient outcomes such as pain, allowing the outcomes to be treated prophylactically. Natural Language Processing (NLP) techniques, which extract measurable information from clinical texts, is showing promising results to predict outcomes and improve treatment techniques [4].

Materials & Methods We developed and tested a NLP pipeline (Figure. 1) to extract and evaluate pain scores from clinical notes. For this study we used a publicly-available deidentified clinical notes dataset provided by the i2b2 National Center for Biomedical Computing [5]. Our NLP algorithm was developed in Python using the NLTK and Spacy NLP toolkits. Biomedical concepts are extracted from the text of consultation notes using Unified Medical Language System (UMLS) Metamap toolkit. A set of 42 unique pain-related terms was identified in the UMLS concepts and extracted from consultation notes using rule-based algorithms. Associated pain scores were obtained using a negation detection algorithm.

Results Our NLP software was tested on 1240 consultation notes from the i2b2 database. 4488 unique indications of pain were detected. Pain was found in 91% of the consultation notes. Based on an audit of 400 of these records, our NLP algorithm showed over 95% accuracy in identifying and quantifying pain documented in patients' consultation notes. Pain score was detected with 89% accuracy.

ABSTRACT. Motivation: Interfraction motion of pelvic organs over the course of prostate radiotherapy can lead to induction of radiation toxicities. Dose reconstruction and monitoring are means to calculate the dose delivered to patients and can be performed using daily CBCT images taken during image-guided patient setup. However, consensus has not yet been reached on the best way to compare planned and daily-delivered doses.

Methods: Daily CBCTs of 20 patients who underwent hypofractionated prostate radiotherapy (60 Gy/20 fx) were used to calculate dose delivered to the rectum. Rectums were contoured on each daily image for all patients, and original treatment fields were copied to daily isocenter positions to calculate daily DVHs. DVHs were sampled and rectum volumes, lengths, and dose points compared to planning values in both relative and absolute volume through a variety of statistical tests.

Results: Significant changes in rectal dose between planning and delivery were detected in multiple patients. Z-scores were determined to be the most effective means of comparing planning and delivered doses in an individual patient. The use of relative units in DVH presentation was found to be vulnerable to interobserver contouring differences, as well as changes in volume outside of the main radiation field. In contrast, absolute volumes were much more robust in the presence of both complications and more easily comparable across the course of treatment than their relative counterparts.

Conclusions: We recommend the use of absolute DVHs when comparing planned and delivered doses in order to avoid volume effects from confounding underlying daily variations.

ABSTRACT. Introduction: Non-inferiority (NI) trials are becoming more common as recognition grows that reduced toxicity or cost may offset reduced efficacy in some situations. Patients have differing preferences, however, and as machine learning approaches are increasingly successful, probability predictions for survival and toxicity are increasingly personalized. The amount of efficacy to give up is decided for the cohort as a whole during the trial design and is called the NI-margin, although many trials use a qualitative method to select this value. One such example in radiotherapy is dose de-escalation for HPV+ oropharyngeal cancer patients. Methods & Results: We modeled the decision of selecting the optimal prescription dose at the patient-level using an Influence Diagram, and calculate a distribution of quality-adjusted life expediencies (QALYs) with a Markov Model for a cohort of patients generated using a Monte Carlo simulation, and selected the NI-margin as the loss in overall survival which is balanced by the expected increase in quality for the cohort. The calculated distribution of QALYs demonstrates that this choice is sub-optimal for some. We use Value of Information analysis to calculate the cost, in terms of QALYs, of the uncertainty in modeled parameters which may lead to the incorrect choice. The Expected Value of Perfect Information for Parameters provides a ranking of the importance of more precise (i.e. less uncertain) information for various parameters in the decision model. Conclusion: A decision model can be used to calculate the NI-margin by calculating the impact of different therapies on a modeled cohort. Value of Information analysis can inform research directions towards developing clinical tests with sufficient performance in order for practitioners and patients to select the option with the greatest benefit.

ABSTRACT. Decision-making is a very central aspect of healthcare, rapid-learning models could be applied in a clinical context to aid in the decision-making process. Using R and Shiny, we developed a rapid learning model to predict patients overall survival probability at a given time point and the treatment option. The shiny web application makes use of four clinical variables: the International Federation of Gynecologists and Obstetricians (FIGO) score, Histological classification, Age of the patient and the proposed treatment to calculate the probability of a patient's overall survival within a specified period. Its features are flexible with a friendly user interface and several drops down options for detailed personalisation. Additionally, the user is presented with the details of the model's results and specifications to make results transparent and explainable. The modular structure of the shiny application can accept a wide range of age values and prediction time points, making the application fully adaptive.

ABSTRACT. Traditional radiotherapy for prostate cancer (PCa) applies a uniform radiation dose to the entire prostate. Focal therapy provides an opportunity to spare healthy tissues while maintaining treatment efficacy. Our group has proposed a specific form of focal therapy, termed “bio-focused radiotherapy” (BiRT), where the radiation dose delivered to the tumour volume depends on the biological characteristics of the tumour (cell density, tumour aggressiveness and hypoxia). This study aimed to develop predictive models from multiparametric MRI (mpMRI) for the implementation of BiRT.

In vivo mpMRI scans were acquired from 30 patients prior to radical prostatectomy. Whole-mount pathology sections acquired at 5 mm intervals through the excised prostate proveded the “ground truth” information of tumour location and characteristics for predictive model development. Predictive models were developed using mpMRI data as features and ground truth from histology as labels. A multi-purpose scalable machine learning workflow (SMLW) was first developed and tested in predicting tumour location at a voxel level. The following investigations focused on the prediction of prostate cell density and tumour aggressiveness using regression techniques. For hypoxia, a radiogenomics approach was applied to investigate the relationship between mpMRI data and corresponding genomic profiles.

Tumour location achieved an area under the curve (AUC) ranging from 0.81 to 0.94. The predictive model for cell density achieved a relative error of 13.25%. Results for tumour aggressiveness achieved an AUC of 0.91, with two high-performance texture features identified as promising biomarkers. Radiogenomic analysis of prostate hypoxia revealed a selection of 16 texture features which showed weak but significant correlations with hypoxia-related gene expression levels.

We have presented a machine learning approach to detect and biologically characterise prostate cancer from mpMRI data with promising results. This approach can be potentially applied to other cancer types to develop decision support systems using non-invasive imaging data.

ABSTRACT. A comprehensive clinical data model in oncology is required to support reliable, consistent capture and exchange of medical information. In an initiative to enable full interoperability across our platform, we have developed a FHIR-based (Fast Healthcare Interoperability Resources) representation of standardized oncology data. Using NCCN Clinical Practice Guidelines in Oncology as our primary source allowed us to not only delineate the scope and range of knowledge representation formalisms required for the execution of a decision support system, but also to assemble a core set of structured data elements that would eventually serve as a base framework for standardized data exchange across Varian’s 360 Oncology platform. To enable the domain experts to define data components, have developed a custom tool that allows and generate their formal structure definitions in JSON format in a simple and quick way. We've taken meaningful steps towards creating a central data model and implementing and standardizing enterprise-wide knowledge management. This development is essential for improving the overall quality of cancer data and its aggregation, and it’s a crucial step to ensuring the quality of long-term care for cancer patients.

ABSTRACT. Structure delineation variability (DV) in radiation therapy (RT) affects planned treatment doses that conformally target and avoid delineated volumes. Prior studies have quantified DV dosimetric impacts in isolation from other uncertainties that exist in RT such as patient setup variability (SV). This study develops a methodology for quantifying the impact of DV on clinically meaningful dosimetric plan quality metrics (PQM) with and without confounding SV with the claim that DV is only clinically relevant if it changes the probability of achieving clinical objectives in the context of SV.

Variable OAR delineations from N=14 independent observers for a single head-and-neck case were used. VMAT plans were auto-generated to meet clinical PQMs for each observer’s OARs set with a consistent target volume. Each plan was then cross-evaluated with all 14 OAR sets without and with simulated SV that included per-fraction random (σ) and per-treatment-course systematic (Σ) setup errors (σ=Σ=2, 4, and 10mm). PQM distributions represented possible outcomes in the presence of DV and/or SV. DV dosimetric impact was assessed by examining changes in the probabilities of achieving each PQM (PPQM).

DV’s dosimetric impact (ΔPPQM) decreased with increasing SV. Without SV, DV alone reduced PPQM by an average of ΔPPQM=-26.3% (range -79.0%, 0.0) across the six OARs. With SV=2mm, ΔPPQM reduced to -25.2% (range -75.7%, 0.0%), to -16.9% (range -52.3%, +6.8%) with SV=4mm, and to -11.3% (range -30.4%, -1.9%) with SV=10mm.

We developed a method to evaluate the clinical impact of OAR DV while considering inherent SV and applied it to a head-and-neck case with 14 independent OAR delineators. The expected trend of decreasing DV impact with increasing SV was observed, however the effect of DV was not washed-out by 10mm SV, indicating a potential clinical impact for this observer set. Other observer sets must be evaluated to generalize conclusions.

ABSTRACT. Radiotherapy is an essential tool in the arsenal of cancer treatment. Basically, it consists of irradiating the tumor while sparing healthy tissues. In the case of mobile tumors, the position uncertainty leads to large margins. A tumor tracking method can reduce dramatically these margins. A classical approach is to acquire X-ray images in the treatment room that are used to estimate the tumor position. Unfortunately, X-ray imaging irradiates the patient. There is a trade-off on the number of such X-ray images: a large amount of them permits a precise tracking of the tumor but irradiates the healthy tissues; on the contrary, reducing the amount of X-ray images acquisition would better preserve healthy tissues from X-rays but would also require an increasing of the margins. Nowaday, the number of these images is fixed by a clinician on the basis of an acceptable induced X-ray dose. In clinical applications, these images are taken periodically, i.e. at constant rate. In [1], Aspeel et al. propose to relax this constraint by allowing variable rate imaging. Their optimal intermittent Kalman predictor for tumor tracking is a prediction method that takes the X-ray images, when they are the most useful. They present a combinatorial optimization problem to select the best measurement times and solve it with a genetic algorithm (GA). In this paper, different implementations of GAs are studied to efficiently solve this problem. Numerical evidences look to show that a GA that implements count preserving crossover outperforms a GA with classic crossover. In addition, it has better convergence properties. This permits a better use of each X-ray image in case of tumor tracking with intermittent measurements.

ABSTRACT. Deformable image registration (DIR) is a core component in the workflow of adaptive radiotherapy. Various commercial vendors now offer DIR algorithms in their software. However a current significant challenge for radiotherapy departments is in the safe implementation of, and confidence in using, DIR algorithms. This work surveys the current use, commissioning & validation and ongoing quality assurance (QA) of DIR in the UK. The purpose of this survey is to use the results to inform as to whether there is a need to develop additional UK specific guidance and QA tools. The survey was developed in SurveyMonkey and sent to all radiotherapy centres in the UK, a total of 71 centres. Centres that do not use DIR clinically were asked if they have future plans to do so and in what timescale. We report on data received from the first 34 responding centres, one month after the opening of the survey. 32 respondents had access to software the could perform DIR. Of these, 37.5% indicated that they use DIR clinically. The most common application of DIR was to propagate contours from one scan to another (100%) and patient-specific QA was the most common form of ongoing QA. In centres that do not currently use DIR in clinical practice, 47% have plans to implement an application of DIR within 1 year. Initial results of the survey indicated that two out of three centre were not using DIR in routine clinical practice. Key barriers reported included determining when a DIR was satisfactory. The results of the survey highlight that there is a need for additional UK guidelines, better tools for commissioning DIR software and better tools for the QA of registration results, which should include developing or recommending which quantitative metrics to use.

ABSTRACT. With the recent developments of hybrid Photon-MRI solutions in standard radiotherapy, continuous real time imaging during treatment is now possible and more research are going towards online adaptive treatment and online tumor tracking. In proton therapy, the hardware solution is not yet available but is under serious consideration. MRI is the ideal imaging modality for this application. It can give a good soft tissue contrast and it is not irradiant, which is crucial for a continuous use during the treatment. However, CT type images are still necessary to extract stopping power information to compute the proton or photon energy loss with a sufficient accuracy. Also, complete 3D image acquisitions are not yet fast enough, only 2D frames can be acquired in real-time during treatment.

To go towards a real time adapted treatment in which the protons are continuously shot towards the target based on a precise tumor and surrounding tissues tracking, it would be useful to have a real-time verification method to evaluate the quality of the treatment delivery.

We propose a workflow which aim is to reproduce in real time and continuously, the breathing induced anatomic motion tracked on 2D dynamic MRI into continuous virtual 4DCT sequences, which can be used to observe during the treatment the dose deposition of protons and be used as an adaptive treatment guiding tool or as a verification tool of the conformance of the delivery.

The proposed method is fast enough to ensure a real-time guidance of the treatment and is designed in such a way that is reliable through an easy interaction with the practitioner.

ABSTRACT. The evaluation of radiation dose to cardiac structures during breast cancer radiation therapy is important for understanding the role radiation may play in inducing adverse cardiac effects. The goal of this work is to develop an automatic segmentation method of cardiac structures based on a library of 30 detailed cardiac atlases coupled with a B-spline 3D deformation method. A set of 30 atlases of cardiac structures was manually drawn using contrast-enhanced computed tomography image sets from 30 adult female patients. For a given patient for contouring using the developed method, it determines a most-similar atlas from agreement of heart-to-heart and performs B-spline deformation of outer contours of the whole heart of the selected atlas by comparing with those of an individual patient. To test the performance of the auto-contouring method, we conducted a leave-one-out cross validation with the 30 atlas models. By computing the Dice similarity coefficients and average surface distances between manual and automatic contours, the agreement is comparable with inter-observer variability of manual contouring. This automated segmentation method will be able to rapidly define heart chambers and coronary arteries. If combined with an improved technique to evaluate both in-field and out-field radiation dose, it will be useful for providing accurate cardiac structure doses for epidemiologic studies of adverse cardiovascular outcomes.

ABSTRACT. Radiation pneumonitis (RP) is one of the most important dose-limiting factors in the radiation therapy of the lung. However, there is a large inter-patient heterogeneity in the lung’s response to radiation, which is most probably due to the radiobiological differences among these patients. In order to explore these differences, a lot of efforts have recently been focused on stratifying the patient population based on some biomarker measurements taken before the treatment. This approach, however, ignores valuable information that can be collected during the treatment course. In this work, we propose a novel machine learning approach by combining Random Forest (RF) and Bayesian Networks (BN) to predict the RP risk based on mid-treatment biomarker information. 27 late-stage non-small cell lung cancer (NSCLC) patients was retrospectively analysed. Mid-treatment [18]F-flurodeoxyglucose (FDG) positron emission tomography (PET) images were used as the biomarker. The clinically symptomatic acute RP was defined as grade 2 or higher. Upon training, the final RF+BN model was able to predict the RP risk with high degree of accuracy (area under the curve [AUC] of receiver-operating characteristic [ROC] = 0.84). In particular, three groups of features were deemed crucial to the model’s predictive power: radiobiological (represented by mid-treatment FDG standard uptake value [SUV] metrics); dosimetric (represented by the mean lung dose [MLD], V20, and V60); and anatomical (represented by GTV volume). When compared to the pre-treatment model, it was found that adding mid-treatment information was crucial in improving the accuracy of the model. These findings both highlight the limitations of pre-treatment stratification methods and hint at a great potential for online assessment of lung tissue radiosensitivity during the course of the treatment.

ABSTRACT. Spiculations are spikes on the surface of pulmonary nodule and are important predictors of malignancy in lung cancer. In this work, we introduced an interpretable, parameter-free technique for quantifying this critical feature using the area distortion metric from the angle-preserving spherical conformal parameterization. Firstly, for a tumor mesh we computed an angle-preserving spherical parameterization and the normalized area distortion ϵ_i for each vertex. Secondly baselines were detected where ϵ_i is zero. Thirdly, we recursively traversed closed curves from baselines toward the negative ϵ_i values until a vertex with the most negative ϵ_i is reached. This detected the apex of a spike. Fourthly, we computed the spike height and a spiculation score. Finally, we construct a malignancy classification model using SVM-LASSO techniques. Our spiculation score was found to be highly correlated (Spearman’s rank correlation coefficient ρ=0.48) with the radiologist’s spiculation score. Adding our spiculation scores to shape and texture features in the SVM-LASSO model improved the malignancy prediction from 84.6% to 88.9% on a public LIDC-IDRI CT image dataset.

ABSTRACT. Purpose A surrogate marker for a clinical endpoint has purported benefits including potential for early assessment, being inexpensive, and being non-invasive. We aimed to show how conclusions derived from surrogate markers can be misleading.

Materials & Methods A Monte Carlo (MC) experiment was used to simulate five patient cohorts, each of size 1000: group 1 (control group), group 2 (radiation therapy, RT, only), group 3 (RT+placebo), group 4 (RT+trial drug), and group 5 (RT+herbal supplement). The simulated endpoint was radiation toxicity (scale of 0-5). Groups 1, 2, and 4 were made statistically different; groups 2, 3, and 5 were not. The surrogate marker was the true toxicity grade plus a noise term (random number drawn from a standard normal distribution, then multiplied by a fixed scale factor). Scale factors of 0.5, 1,... 2.5 were used. The experiment was repeated 1000 times. We counted the number of instances where groups 2 and 4 were statistically different according to the Wilcoxon rank-sum test (p<0.001).

Results For the five noise level scale factors (0.5-2.5), the median Spearman correlation (ρ) between the surrogate marker and the toxicity grade were 0.91, 0.78, 0.67, 0.57, and 0.49, respectively. The rate of reduction in ρ slowed as the noise level scale factor increased. When using the surrogate marker, groups 2 and 4 differed 996, 865, 466, 228, and 125 times (scale factor going from 0.5 to 2.5), even though they were truly different 1000 times.

Conclusions This paper was motivated by the potential near-future use of surrogate radiomic markers, but since the analysis performed is purely statistical, the implications are applicable to markers from other domains. When designing a clinical trial that uses a surrogate marker, such an analysis needs to be performed to ensure that the trial does not miss a statistically significant difference.

ABSTRACT. To perform radiomics analysis of primary tumors extracted from pre-treatment contrast-enhanced computed-tomography (CE-CT) images for patients with oropharyngeal cancers to identify discriminant features, and construct an optimal classifier for prediction of Human-Papilloma-Virus (HPV) status. One-hundred-eight-seven patients with oropharyngeal cancers with known HPV status (confirmed by immunohistochemistry-P16-protein testing) were retrospectively studied: Group-A: 95 patients (19HPV- and 76HPV+) from the MICAII-grand-challenge. Group-B: 92 patients (52HPV- and 40HPV+) from our institution. 172 radiomics features were extracted from diagnostic CE-CT images of the gross-tumor-volume (GTV). Levene and Kolmogorov-Smirnov’s tests with absolute-biserial-correlation were used to identify discriminant features between the HPV+ and HPV- groups. Discriminant features were used to train and test eight different classifiers: Generalized-Linear-Model, Support-Vector-Machine, Artificial-Neural-Network, Discriminant-Analysis, Naive-Bayes, Random-Forest, Decision-TREE, and k-Nearest-Neighbors. Area-Under-Receiver-Operating-Characteristic (AUROC), Positive-Predictive and Negative-Predictive values (PPV and NPV, respectively) were used to identify the optimal discriminant features and evaluate the classifiers. Ultimately, both datasets were combined and randomly split (20,000 iterations) into training and test cohorts using a bootstrapping-permutation sampling technique and the performance of the optimal classifier for prediction of the unseen cohorts was evaluated. Among 172 radiomics features only 12 features (from 3 different categories) were significantly different (p<0.05, |BSC|>0.48) between the HPV+ and - groups. Among the 8 classifiers trained and applied for prediction of HPV status, the generalized-linear-model (GLM) showed the highest performance for each discriminant feature and the combined 12 features: AUROC/PPV/NPV=0.88/0.83/0.81. The GLM also showed a high prediction power (AUROC/PPV/NPV= 0.85/0.73/0.79 and AUROC/PPV/NPV= 0.87/0.81/0.87 for group A and B respectively) against unseen test sets. Results imply that GTV’s for HPV+ patients exhibit higher intensities, smaller lesion, more sphericity, and higher spatial intensity-variation/heterogeneity. Results suggest that radiomics features primarily associated with the intensity arrangement and morphological appearance of the tumor on diagnostic CT may be potentially useful for classification of HPV status.

ABSTRACT. Purpose In a Machine Learning (ML) classification problem, it is common practice to use a balanced training set. Two common balancing approaches are: (1) undersample the majority class, or (2) generate synthetic samples of the minority class. Both these approaches have limitations. We demonstrate via Monte Carlo the feasibility of a promising alternative approach: to use the imbalanced training set and alter the classification threshold of the ML classifier instead.

Materials & Methods The predictive feature was modeled as the sum of two Gaussians (unit standard deviation), one centered at 0 (associated with negative outcome), the other centered at 1 (associated with positive outcome). The training sets had the negative class fraction varied from 0.1-0.9, in steps of 0.1; the testing set was balanced. Three simple ML classifiers were investigated: logistic regression, Naive Bayes, and linear discriminant. The optimized classification threshold (OCT) was determined from the Receiver Operating Characteristic (ROC) curve, and defined as the threshold corresponding to the point on the curve that has the minimum distance from the point (0,1) (i.e., false-positive-rate = 0, true-positive-rate = 1).

Results Upon applying the OCT instead of the default threshold, accuracy (0.691±0.001), sensitivity (0.693±0.002), specificity (0.691±0.002), and AUC (0.760±0.001) are essentially the same for the nine training sets for all three classifiers. By contrast, F-score and Matthews correlation coefficient are not stable across the nine training sets, despite using the OCT. Thus, they are unreliable metrics for imbalanced sets.

Conclusions We have demonstrated that it is possible to train ML classifiers on imbalanced sets by altering the classification threshold. However, the proof of concept was done on sets of size 10^6. When dealing with clinical datasets of size ~10^2, this approach must be used with caution since the OCT derived from such a small set may be affected by statistical fluctuations.

ABSTRACT. In this paper, we demonstrate the use of various automation techniques and artificial intelligence based implementations that can be very effective in delivering optimal care in a busy cancer center. The proposed technique makes use of Radiomics based feature extraction, merging imaging based features with non-imaging based clinical features to generate the data, conversion of these features into Semantic ontology compliant format and then run the clinical outcome predictive models for the patient under consideration. Further, we also demonstrate the use of automatic tumor segmentation algorithm built using convoluted neural network that can further optimize the care and reduce the burden on the clinicians. The early results obtained are very encouraging and the plan is to extend this workflow to address other cancers. The current deployment is specifically addressing non-small cell lung cancer (NSCLC) using PET CT images in a busy cancer center in India