ABSTRACT. Progress in deep learning has considerably improved the progress for deploying artificial intelligence solutions for healthcare. Modern deep neural networks go beyond classical ones in many ways, including the ability to train deeper networks which generalize better, the ability to analyze free-form text, the ability to manipulate complex data structures (such as the graphs of molecules), or the ability to synthesize new data according to specifications. We review a few applications of deep learning in healthcare, including in particular in analyzing medical images, in sensor fusion and managing missing input modalities, in processing doctors' reports, in drug design, as well as in analyzing time-series of patient records. We argue that what we are currently seeing is only the tip of the iceberg, and that healthcare will be profoundly transformed by progress in artificial intelligence and greater ability to take advantage of large quantities of medical data.

ABSTRACT. Deep learning methods have recently emerged as the most accurate machine learning methods and have shown tremendous potential in many applications in computer vision in both real-world and medical image analysis. With the ability to automatically extract and identify relevant features at varying resolutions specific to tasks, these methods have surpassed standard segmentation methods used in radiation oncology. However, deep learning requires a large number of well curated, expert-segmented datasets for training, which is difficult to obtain for several medical image applications in radiation oncology and radiology, with problems ranging from dataset differences due to varying imaging acquisitions, inherent variability in the expert “ground truth” delineations due to inter-rater variability to difficulty in even obtaining datasets for novel applications such as MR-guided radiotherapy. This talk will present some of techniques from deep learning to address this issue. Presented techniques will include cross-modality educed learning to leverage more informative MRI to enhance interpretation of CT scans, as well as cross-domain adaptation through unsupervised and semi-supervised learning for learning from small datasets applied to tumor and normal organ segmentations in lung and head and neck structures in CT and MRI.

ABSTRACT. Stereotactic body radiation therapy (SBRT), alternatively termed stereotactic ablative radiotherapy (SABR) or Stereotactic Radiosurgery (SRS), delivers high dose with a sub-millimeter accuracy and requires meticulous precautions on positioning, as sharp dose gradients near critical neighboring structures (spinal cord for spine tumor treatment) are an important clinical objective, avoiding complications such as radiation myelopathy, compression fracture, or radiculopathy. To allow for a dose escalation per fraction, proper immobilization needs to be combined with (internal) motion monitoring. Metallic fiducials, as applied in prostate or pancreas treatments, are not suitable in clinical practice for spine SBRT. However, the latest advances in deep learning (DL) allow for fast localization of the vertebrae as landmarks. Acquiring projection images during treatment delivery allows for instant 2D position verification as well as sequential (delayed) 3D position verification when incorporated in a DTS or CBCT. Upgrading to an instant 3D position verification system could be envisioned with a dual kV imaging setup. This paper evaluates a fast DL model for vertebra detection (developed in-house) on its accuracy to detect motion of the vertebrae on projection images acquired during treatment. The introduced motion consists of both translational and rotational variations resulting in a (sub-)millimeter positional change.

ABSTRACT. In medical computer vision tasks such as segmentation and diagnosis, it is important to ensure that inference is done on a model trained on similar datasets. Furthermore, training datasets need to avoid misclassified data. We develop a method to classify CT datasets by anatomical region using an autoencoder and dimensionality reduction. The method is self-supervised, in that training datasets are not labelled. Latent variables are extracted from the each CT dataset using a modified U-net without skip connections. The binary cross entropy between the 3D input image and the decoded (output) image is minimized. The dimensionality of the latent variables is reduced to 2 using principal component analysis and t-SNE. The U-net bottleneck is 12x12x8x64 = 73728 latent variables, <2% of the number of voxels in the resized input image. We train on 712 publicly available CT images and perform inference on a total of 1745 images. The resulting t-SNE plot showed images grouped by anatomical region. Head and neck, thorax, abdomen, and extremity scans were separated into distinct clusters. Lung datasets were in two clusters, possibly based on either the superior edge of the image or the type of couch (diagnostic or radiotherapy). Images of legs were in a cluster. Abdominal images spanned a large area of the t-SNE plot. The analysis also separates datasets with different patient orientations, couch surfaces), and imaging modality (CT or MR), which can an important quality assurance step in training and inference for segmentation models. This work has shown that a self-supervised algorithm can cluster images based on anatomical region. Future work will be to optimize the network and number of latent variables and to investigate if other features such as age and sex and be determined.

ABSTRACT. Breast ultrasound (US) is an effective imaging modality for breast cancer detection and diagnosis. US computer-aided diagnosis (CAD) system has been developed for decades either using conventional handcrafted features or modern automatic deep-learned features, where the former relies on human accumulated clinic experience and the later demands amounts of datasets. In this paper, we developed a novel BIRADS-SDL network, semi-supervised deep learning (SDL) with the integration of clinical-approved breast lesion characteristics (BIRADS features), to achieve accurate breast US diagnosis with a small training dataset. Breast US images are converted to BIRADS-oriented feature maps (BFMs) using a distance-transformation coupled with a Gaussian filter. Then the converted BFMs are used as the input of SDL network, which is characterized as a multi-task learning by integrating unsupervised stacked convolutional auto-encode (SCAE)-based image feature extraction and diagnosis-oriented supervised lesion classification. Such integrated multi-task learning allows SCAE to extract image features with the constraints from the lesion classification task, while the lesion classification is achieved by utilizing the SCAE encoder features with a convolutional network. The developed BIRADS-SDL network is trained with an alternative learning strategy by balancing reconstruction error and classification label prediction error. The performance of BIRADS-SDL network is compared to the conventional SCAE methods and SDL methods using original image input, as well as SCAE with BFMs input. Experimental results on a public breast US dataset show that BIRADS-SDL ranked the best among four networks, with classification accuracy around 92.00±2.38%, which indicate that BIRADS-SDL could be promising for effective breast US lesion CAD using small dataset.

ABSTRACT. Introduction Extracting information from pathology and surgical notes constitutes a major bottleneck for their use in clinical research. Manually performing this task is expensive and time consuming. Rule-based approaches have been developed to facilitate the extraction, but they do not scale to different types of reports. Deep Learning (DL) techniques have therefore been recently applied to the task. In this study, we investigate whether a state-of-the-art Deep Learning algorithm can outperform a rule-based method in the extraction of tumour information from pathology and surgical notes of chondrosarcoma and enchondroma patients. Method We develop a DL approach to automatically extract fourteen binary attributes from chondrosarcoma and enchondroma pathology and surgical notes. The attributes are meant to capture crucial information about the tumour, such as histology type, origination and grade. Our approach is based on a Word-Level Convolutional Neural Network (CNN), which has shown impressive performance in previous applications. We train and evaluate it on a dataset of 960 pathologically confirmed chondrosarcoma or enchondroma patients. The performance of our approach is compared to an existing rule-based system, developed in two months by a full-time clinician. Results The DL algorithm obtains an across-attribute average accuracy of 95%, compared to 84% of the rule-based approach. We noticed that skewed attributes (i.e. attributes in which one label appears more than 85% of the times) are more difficult to model for the DL approach. For the five attributes having a balanced label distribution, the DL algorithm always outperforms the rule-based system. In the remaining nine attributes, DL performs better on 7 and equal on 2. Conclusion Our DL algorithm is capable of extracting information from two different types of reports. This method outperforms the rule-based approach by 11%, obtaining higher scores in almost all attributes. The extracted information will be used to support retrospective research.

ABSTRACT. Purpose: We hypothesize that convolutional neural networks could enhance the predictive power of traditional radiomics, by detecting image patterns that may not be covered by a traditional radiomic framework.

Materials & Methods: We test this hypothesis by training a CNN to predict treatment outcomes of patients with head & neck squamous cell carcinoma, based solely on their pre-treatment computed tomography image. The training (194 patients) and validation sets (106 patients), which are mutually independent and include 4 institutions, come from The Cancer Imaging Archive. This study is specifically benchmarked against a previous study on the same data, which correlated several radiomic features from CT images with the outcome of H&N cancer patients. Our methodology represents a novel approach in that we use a single end-to-end CNN trained de novo (with no secondary machine learning algorithms) to predict outcomes. Furthermore, we introduce several visualization tools to help gain intuition into the network’s behavior.

Results: Our method results in an AUC of 0.88 in predicting distant metastasis, comparable to the benchmark result of 0.86. When combining our model with the previous model, the AUC improves to 0.92. Among other tools, a class activation map was used to visualize the network’s behavior.

Conclusion: This study showed the potential of using a CNN built de novo to perform outcome predictions on the pre-treatment CT image of head & neck cancer patients. Our framework overcomes many of the typical issues encountered when building a traditional radiomics-based model. Our model is capable of being interpreted in an intuitive fashion and eschews the need for feature engineering.

ABSTRACT. Introduction: Manual digitization of interstitial needles in high dose-rate brachytherapy (HDRBT) is a tedious and time-consuming task. The accuracy is affected by human experience and available time for this task. It also becomes particularly challenging when needles are close, or even touching each other. Hence, an accurate and efficient auto-digitization tool is desired. In this work, we develop a deep-learning based method for robust and efficient auto-digitization of interstitial needles.

Methods: The proposed method consists of two steps. The first step is to segment all needles from 2D CT slices using a deep U-net. We train the U-net using 95,000 CT slices and corresponding applicator mask images. CT images are collected from 15 patient cases and augmented with translation and rotation to enlarge data size. The second step is to determine the trajectory of each needle. We develop a novel iterative scheme integrating nearest-neighbor clustering and polynomial curve fitting to extract each needle’s central line based on the segmentation results from the first step. Applicator tip position is determined as the point with the highest CT-number gradient along the needle. We evaluate our method using three interstitial needle cases that are not used in the U-net training.

Results: For the segmentation step, the average Dice similarity coefficient reaches 0.95 on training data, 0.94 on validation data, and 0.93 on testing data. After the digitization step, the Hausdorff distance between central lines digitized using the proposed method and manual digitization is ~0.6 mm on average while the mean tip position distance is ~0.4 mm.

Conclusion: We have developed a deep-learning based auto-digitization method for interstitial needles in HDRBT. The achieved accuracy makes our tool clinically attractive.

ABSTRACT. Purpose: To develop a generative adversarial network (GAN) method to produce synthetic CT (sCT) images for proton therapy MRI-only treatment planning with accurate geometric and electron density information from MRI. Methods: A deep GAN model was designed to learn a direct mapping function to convert a MRI slice to its corresponding CT slice. The model was trained by collecting all MRI slices with corresponding CT slices from each training subject’s MRI/CT pair. One hundred brain tumor patients with both CT and T2-weighted MRI images was used for this study. DVH analysis was performed on CT and sCT images for plans generated using intensity modulated proton therapy (IMPT). Results: The average of mean absolute error (MAE) ± SD values for all test data using 5 fold cross-validation were 41.8 ± 10.0, 48.8 ± 13.0, 48.2 ± 12.4, 48.2 ± 12.2, 48.3 ± 12.0. This leads to an average MAE over all cross validation sets equal to 47.2 ± 11.0. The results extracted from RayStation showed an excellent agreement for most DVH metrics computed on the CT and sCT, with a mean absolute difference below 0.5% (0.3 Gy) of the prescription dose for the CTV and below 2% (1.2%) for the considered OARs. Conclusions: A GAN model method was developed, and shown to be able to produce highly accurate sCT estimations from conventional, single-sequence MRI images in near real time. Quantitative results also showed that the proposed method can generate sCT images with improved accuracy and faster computation speed compared to conventional methods. This work demonstrated the feasibility of using sCT generated with a deep learning method based on GANs, for MRI-only treatment planning in intensity modulated proton therapy.

ABSTRACT. Efficient flow of radiotherapy patients in cancer treatment centers is tightly related to the availability of the physicians and the use of the linear accelerators. Variability is one of the most complicating factors to consider into the planning activities: patients arrival and priorities are not known in advance, and duration of treatments is rarely constant. In practice, centers use grids of constant length for patients’ appointments, and relatively intuitive strategies to choose the right slot for each patient. In this presentation, we discuss more sophisticated methods based on operations research for physician scheduling and patient appointment booking. Patient appointments are based on their priority, target wait times and duration of treatment. Before starting their treatment, patients first consult with the physician, then undergo a set of sets. First, we propose to base the physician schedule on patients activities. This ensures that physicians are never the bottleneck in patient flow. Second, we develop a hybrid method combining stochastic optimization and online optimization to schedule patients. We use information from future arrivals of patients to model as accurately as possible the expected use of resources. We also use predictive models to better assess the patient's treatment duration (preparation, entry into the room, treatment, etc.). We show that it is possible to create appointments schedules of different durations according to a set of criteria. This work is the result of a collaboration with the Center Intégré de Cancérologie de Laval.

ABSTRACT. Introduction

The purpose of this work was to develop an artificial intelligence (AI)-guided workflow, that enables clinical incident reporting, tracking, and risk analysis, which are critical components of quality assurance (QA) in radiation oncology. Text information used in clinical reports has unique terminology and requires special domain knowledge to be processed correctly. We have developed and trained an AI-guided incident learning application utilizing recurrent neural networks (RNN) and a word-to-vector, natural-language processing-(NLP)-based model with oncology domain knowledge, to automatically determine categories and severity level of incidents and identify risks in the clinical workflow.

Methods

A total of 3210 existing incident tickets from 16 categories entered by clinical staff over 7 years were used for development of the classification model. Ninety percent of the events were used for training and 10% for validation of the RNN-based algorithm. In order to apply oncology domain knowledge to the embedding layer, we trained a word-to-vector model on a total of 286 million words from biomedical research articles obtained from OpenI with keyword “Oncology.” The AI model were deployed as REST API services which served as an engine to analyze incidents entered through a web application.

Results

After 20 epochs of training, the RNN model reached accuracy of 89.5%, 85.0%, 75.2% for classification of the top three, two and one categories, respectively. RMS error of the severity level prediction reached 0.79 after 74 epochs. A 100-dimensional word-to-vector model was trained using corpus from biomedical articles.

Conclusions

A RNN-based incident learning framework for automatic multi-label classification and severity level estimation has been developed. Visualization of word embedding layer shows that domain knowledge has been applied to the model. The trained word-to-vector model addressed the word-level label space ambiguity by identifying highly polysemous words that are unique in oncology diagnosis and treatment terminology.

ABSTRACT. Our group has developed a software performing an automatized analysis of daily acquired images for linac quality controls. Full test reports are generated as web pages and summary reports are sent to physicists by email. Source code of the main application is written in C++, called by a python script that also generates the reports. Portal images and measurement informations are loaded from the record and communication server through DICOM and SQL requests respectively. The application is automatically launched after morning test measurements made by therapists to perform daily analysis. A manual mode can also be used for monthly and annual quality controls. Implemented analysis goes beyond AAPM TG-142 and includes MLC tests (static fields and VMAT plans), jaw-junctional fields and jaw positioning, dose and profiles for open and EDW fields as well as isocenter geometry checks. Automatized data handling and analysis by QaTest has been running smoothly since early 2015, providing online results on a local server that anyone on-site can consult. The QaTest application could easily be interfaced with more general platforms like QATrack+ through python scripting. Using QATrack+ as a front-end control test lists and result storage with QaTest launched as a background analysis tool is one of the expected developments. In the future, QaTest schema of automatized portal image handling could be expanded to daily patient in-vivo measurement analysis, and study of specific variables could point out patients who might need some treatment adjustments. Such a knowledge-based predictive tool would help both the decision and planification of treatment plan adjustments.

ABSTRACT. Introduction: The overall accuracy of prescribed delivered dose in radiotherapy has to be within 5%. Uncertainties in the dose calculations performed by the treatment planning system (TPS), delivery errors or data corruption might cause significant differences between predicted and delivered doses. Therefore, in treatment beam patient specific quality assurance (QA) is currently the gold standard. This measurement-based approach, however, is work intensive and impose a bottleneck to treatment efficiency and deployment of daily adaptive radiotherapy. In this work, we investigated the potential to replace measurements with a simulation-based patient specific QA using a Monte Carlo (MC) code as independent dose calculation engine in combination with patient treatments delivery log files. Materials and Methods: An automated server-based QA platform was developed. It consists of a web interface, servers and computation scripts, and is capable to autonomously launch simulations, identify and report dosimetric inconsistencies. MCsquare is employed as dose calculation engine. For validation purposes, in-water MC simulations for a wide range of Spread of Bragg Peak (SOBP) plans with varying range and modulation were performed. Furthermore, clinical cases were submitted to automated Log-based QA, which consists of MC simulations of dose distributions based on plans reconstructed from delivery logs. Results: MC simulated dose distributions for 30 SOBP fields show great consistency when compared to the TPS calculated dose maps, with gamma (2 mm, 2%) pass ratios of 99% ± 0.5%. The comparison of measurement- and Log-based QA results for clinical cases - including craniospinal axis, intracranial and head and neck cases - showed that delivery spot positions recorded in log files marginally affected dose distributions. Conclusion: A retrospective application of the proposed automated Log-based QA workflow for a wide range of clinical plans shows similar passing rates to measurement-based QA, supporting the replacement of measurements with MC-based treatment delivery files QA.

ABSTRACT. Broader clinical adoption of adaptive radiotherapy is hindered by several factors. Key among these is the human effort required to manage the plethora of data acquired and processed over a course of radiotherapy. This is especially true when these data; such as Tx directives, planning images, daily Tx records and on-line images as well as the processing and evaluation tools reside in more than a one system. In addition to resource concerns, manual processing and transfer of data, perhaps multiple times, increases the risk for errors and limits the ability to maintain a robust audit trail and clear “source of truth.”

To address these barriers, commercial software providers are implementing some degree of automation to support parts or all of “some adaptive radiotherapy workflow.” While these developments are helping to reduce the manual processing required, the resulting workflows are typically “one size fits all” and inherently restrictive. Basically, the current solutions require users; physicians, dosimetrists, physicists and therapists to adapt their clinical workflow to the software rather than vice versa.

Fortunately, commercial software providers are embracing technologies which expose internal data models and algorithms as well as standards which support greater and more fine-grained interoperability. By leveraging these technologies, it is possible to construct solutions that “adapt the software” to support specific clinical needs and workflows rather than vice versa. In this presentation, we describe the development of a completely automated software solution to manage an adaptive workflow that integrates and tracks the flow of data from the original Tx directive to a dose-to-date calculation, potentially triggered by a “session complete” event. This work builds on concepts and work presented at the two previous ICCR conferences in Melbourne and London.

ABSTRACT. Purpose: High-dose-rate brachytherapy (HDRBT) treatment planning is manually performed on the day of treatment. After that, every plan is subject to a quality assurance check and documentation prior to treatment delivery. This complex process contains many tedious steps performed under a time pressure, which is highly inefficient and prone to human errors. We have created the AutoBrachy system that automates the treatment planning, quality assurance (QA), and documentation steps for cylinder and tandem-and-ovoid (T/O) applicators for HDRBT of gynecological cancer.

Methods: AutoBrachy is an interactive web-based system. After inputing treatment specific information, such as prescription dose, the planning module accesses patient CT images stored on a data server and digitizes applicator using image processing techniques, place source locations and determines dwell time via inverse optimization. The generated plan is imported to the treatment planning system, and is further adjusted by a physicist as needed. After physician approval, the treatment plan is exported from the TPS and checked by the QA module that performs a secondary dose check and other checks on geometric and dosimetric accuracy. Potential problems are highlighted in a QA report. Upon QA completion, over 50 dosimetric variables are automatically stored in the searchable AutoBrachy database per requirements of ICRU report 89.

Results: Applicators were accurately digitized with less than 1mm error compared to manual digitization. Generated plans were clinically acceptable. Compared to average manual planning time of 15 minutes for cylinder and 30 minutes for T/O cases, AutoBracy significantly reduced planning time to less than 3 minutes. The QA and documentation module allowed for the creation of faster and comprehensive documentations.

Conclusion: AutoBrachy has successfully automated treatment planning, QA, and documentation for cylinder and T/O applicators. Further improvements to support other applicators are in progress.

ABSTRACT. Introduction: Dose escalation for prostate cancers can reduce clinical failure but increases risk of genitourinary toxicities. Modern image-guided radiotherapy provides rectum and bladder sparing for prostate cancers with narrow target margins, thereby maintaining acceptable toxicity risks. Daily bladder fill variations causes dose-volume data obtained from planning CTs to poorly estimate actually delivered doses. While the problem of adapting treatments to counter bladder variations receives considerable attention, the complementary problem of quantifying actual delivered bladder dose does not. In this work we describe a technique for quantifying actual bladder dose using 2D dose surface maps (DSMs). The association with genitourinary toxicity is explored via a case-control study.

Methods: 28 prostate cancer patients treated with radiotherapy were studied. Seven developed toxicity; each was paired with three control patients presenting without toxicity. Daily CBCTs were registered to planning CTs and bladders were manually contoured. DSMs were constructed by marching digital rays through virtualized patient geometries, detecting where rays intersected bladder surfaces, and interpolating dose at the intersection points. DSM dose distributions were compared (cases vs. controls) using population-averages. Dose histograms for normalized differences maps (positive and negative) were tested for significance by comparing the 5th percentile to an expected 2σ threshold.

Results & Discussion: Bladder volumes varied insignificantly throughout radiotherapy (p=0.16). Patients who developed toxicity had 10 Gy lower bladder surface dose compared to control patients. The positive dose-difference histogram was significantly beyond the expected theoretical variation (p=0.006) suggesting that bladder surface doses can vary significantly during radiotherapy.

Conclusion: The proposed technique can be used to quantify actual delivered dose distributions to the bladder over the course of radiotherapy for prostate cancers, and could therefore be used to more accurately predict toxicity. This technique could in principle be combined with adaptive radiotherapy to avoid toxicity.

ABSTRACT. Inter-fractional tissue deformations pose a serious problem to accurate dose delivery in adaptive radiation therapy. Deformable image registration approaches to this problem are widely used, but the quality of results varies and resulting transformations often lead to unrealistic deformations. Here, we present a deformable-model registration approach which guarantees the biofidelity of resulting tissue deformations. It is based on our previously published, articulated skeleton model and is meant for inter-fractional treatment of head and neck cancer patients. The model is built up by manual delineation of bones on the patient’s planning CT. Following this, joints are placed automatically between bones, and articulation is achieved using the Simbody library. For optimization, we use a hierarchical scheme to reduce the dimensionality of the optimization space. In this scheme, individual bones are optimized sequentially using a Nelder-Mead type algorithm while already optimized bones are held in place. The optimizer uses a similarity metric based on counting overlapping voxels between the model’s bones and thresholded (120 to 2000 HU) fraction CTs. We have assessed the quality of our registration approach by using landmarks which where manually defined on bones. We compare its accuracy to the inter-observer variation of landmarks positioned by multiple human observers. The median deviation of our registration approach was found to be similar to the inter-observer variation. The accuracy achieved was between 1-2mm and acceptable for clinical use. Biofidelic deformations were guaranteed by the confinement of movements to the skeletons range of motion. Currently, the time needed for convergence is about 30 minutes and our future work will focus on increasing performance while preserving accuracy.

ABSTRACT. In this work we explore the combination of 2.5D generative adversarial network (GAN) for fast dose prediction in prostate radiotherapy with automated treatment plan optimization targeting the on-the-fly predicted 3D dose using our previously developed Adaptive Sequencer (ASEQ). We have trained a 2.5D GAN network by extracting anatomical/dose slice pairs in sagittal, coronal and transversal directions of 50 prostate patients. Each slice encoded the binary masks of the relevant clinical volumes of interest along with the background CT Hounsfield Units. Then, for each test patient the respective slices were extracted and fed into the trained network yielding one predicted 3D volume per slice direction. These volumes were averaged leading to the final predicted dose of the patient. For 13 independent test patients, the average difference between the clinical and predicted dose for the PTV and EBV was 0.6%±2.1% and 1.4%±1.6%. The predicted 3D dose was then used to drive the automated ASEQ treatment plan optimization process. During a two-step IMRT optimization, ASEQ uses this predicted dose as the ideal dose that should be calculated/delivered, and iteratively generates deliverable segments while converging to the target dose. For 2 patients we show that the ASEQ automated plan produces similar dose distributions to the original clinical plan yielding (-0.3%±02.7% PTV relative difference). All three dose distributions (clinical, predicted and ASEQ) share similar DVH parameters especially in the high dose region. The automated dose prediction with deep neural networks combined with fast plan optimization can lead to clinical grade plans and can be used in an offline setting as a standalone reference during manual treatment planning or as a fully automated system in online environments in the context of MRI-guided radiotherapy.

ABSTRACT. We present the application of a generative adversarial network (GAN) to creation of synthetic computed tomography (sCT) scans from volumetric T1-weighted magnetic resonance imaging (MRI). Our motivation is to produce sCT datasets that are sufficiently realistic for dose calculation in MRI- based adaptive-radiotherapy workflows. Paired CT and MRI datasets from 42 patients who underwent stereotactic radiosurgery were used for development and initial testing of the algorithm. A 3D GAN for image-content transfer was created in PyTorch, based on the pix2pix architecture for 2D image translation. The dataset was split as 38 patients for training and 4 patients for testing. The training time was approximately 26 hours on a Nvidia GTX 1050 GPU. In the test phase, sCT generation took only 25 s per patient. The performance of the GAN was assessed qualitatively, and by calculation of the mean absolute error (MAE) of intensities in Hounsfield Units (HU). Initial results showed faithful reproduction of bone-air interfaces. MAEs for the whole body, bone, air and soft tissue were respectively 75 ± 8 HU, 172 ± 28 HU, 281 ± 9 HU and 38 ± 5 HU. The MAE values compare favourably with previously published sCT algorithms based on convolutional neural networks, and our 3D approach outperforms a 2D GAN architecture applied to the same dataset. Further cross-validation will be performed using the existing dataset, and the network will be tested on an additional unseen dataset of 24 patients. Final validation of the algorithm will be performed by comparison of calculated dose based on true and synthetic CT data, using a computational pencil-beam model of photon and proton delivery at clinical energies.

ABSTRACT. To employ daily cone beam computed tomography (CBCT) images for online-adaptive radiotherapy workflows, fast CBCT image correction to correct Hounsfield units (HU) is required. While U-shaped convolutional neural networks (Unet) trained in 2D have been used for this task, the 3D nature of CBCT images suggests a 3D Unet may yield improvements in terms of slice-by-slice consistency.

For training, CT to CBCT deformable image registration (DIR) was performed for 42 prostate cancer patients as part of a previously validated projection-based correction method called CBCTcor. The Unet2D was trained on the raw and CBCTcor image slices. For Unet3D all 2D convolutions were replaced by 3D convolutions and the training was performed on 3D patches of 32 slices. On a Nvidia P6000 GPU with 24 Gb of memory a batch size of 2 could be used with Unet3D. Patients were distributed in training (27), validation (7) and testing (8) groups, and CBCTcor was used as reference throughout this study when evaluating CBCT image accuracy.

Unet3D took considerably longer to train than Unet2D, requiring 853 epochs (3 days) instead of 140 (9 hours) to reach a similar loss function value. The mean error (ME) and mean absolute error (MAE) over all patients for Unet2D/3D were 2/4 HU and 56/58 HU. For the testing patients not seen during training the ME and MAE ranged from -10 to 12 HU for Unet2D and from -8 to 17 HU for Unet3D. In coronal/sagittal slices Unet3D showed less slice-by-slice variability and appeared smoother.

Using a 24 Gb GPU, it was possible to train a 3D Unet on patches of 512×512×32. The value of the optimal loss function as well as the ME and MAE were similar between 2D and 3D training, while the 3D training showed smoother coronal/sagittal slices.

ABSTRACT. The MR-linac system combines a linear accelerator (Elekta AB, Stockholm, Sweden) with a 1.5 T MRI scanner (Philips Medical Systems, Best, the Netherlands). In the workflow of the MR-linac, verification of the delivered plan is of high importance. To that aim, the use of Electronic Portal Imaging Devices (EPIDs) as a quality assurance tool is being investigated and developed. EPIDs are detectors placed behind the patient containing information about the actual treatment delivery. Recent work has shown the feasibility of using EPIDs as dosimeters in the MR-linac, and the first pre-treatment verifications have been performed using an adapted back-projection algorithm that accounts for the extra scatter and attenuation of the beam due to the presence of the MRI scanner between the patient and the EPID.

However, given mechanical constraints of the MR-linac, the acquisition of un-attenuated beams with the EPID is limited to an irradiation field of a maximum of ±4.8 cm in each direction of the longitudinal axis at the isocenter. For larger fields, the exit beam’s dimensions exceed the coil-free region and therefore, the exit beam is strongly and inhomogeneously attenuated.

The purpose of this study is to correct the undesired effect of the in-homogenous attenuation in the EPID dose reconstructions by using a U-net based deep neural network. 94 IMRT beams were irradiated to a slab phantom and EPID images were acquired and then back-projected to the isocenter. The same beams were measured with a 2D array detector and used as the ground truth for training.

A subset of 5 EPID and array dose distributions was used for independent test, and the averaged gamma-evaluation results (global, 3%, 2mm, 10% isodose) are reported for the array images compared both to the original back-projected EPID images and to the EPID images after the deep learning (DEEPID).

ABSTRACT. Prostate cancer is the most common solid tumor among men in the United States. It is multifocal in up to 87% of cases and the ability to distinguish between benign and malignant tissue within the prostate is a step toward optimizing diagnosis and treatment. This paper aims to accurately classify benign and malignant lesions in multiparametric MRI (mpMRI) using a convolutional neural network in the absence of big datasets. Learned features at each layer are systematically evaluated with different combination of mpMRI inputs for understanding and visualizing the process of convolutional networks.

ABSTRACT. Introduction: Simultaneously motion compensated image reconstruction (SMEIR) has been developed to improve 4 dimensional (4D) cone beam computed tomography (CBCT) image quality using deformation vector fields (DVFs) to perform motion compensated reconstruction. Biomechanical modeling has been introduced to SMEIR to improve the inner lung region DVFs accuracy by deriving inner lung DVFs from lung boundary DVFs using its physics properties. We proposed a convolutional neural network (CNN) based method to derive inner lung DVFs with expectation to improve the accuracy as well as increase the efficiency and further benefit the reconstruction of 4D-CBCT. Methods: Two U-net based architectures are developed to estimate DVFs inside of lung. The first one (U-net-3C) contains 3 channels where each channel is one direction of DVFs from SMEIR method generated by 2D-3D deformable registration. The output also contains 3 channels where each denoting one updated high-quality DVFs. The second architecture (U-net-4C) has an additional channel which is CT image of reference phase as input, with expectation to find informative features from heterogeneous properties of inner lung tissue. 11patients with 85 DVFs of size 128×128×128×3 are used to train the model and we performed a 5-fold cross validation. Predicted DVFs accuracy is evaluated by comparing displacement extracted from DVFs with manually-tracked landmarks displacement. Results: Average magnitude of displacement 3D vector between U-net-3C predicted DVFs and landmarks is 3.34 mm and that between U-net-4C predicted DVFs and landmarks is 3.02 mm, while this evaluation parameter is 5.04 mm for SMEIR and 3.67 for SMEIR-Bio. The trained model takes 10 seconds to obtain an updated 3 channels- DVF. Conclusion: In summary, we developed a CNN based DVFs derivation approach which can improve the DVFs accuracy compared with 2D-3D-registration and biomechanical modelling as well as decrease the computational time substantially.

ABSTRACT. Real-time accurate localization of the target based on single x-ray projection is very challenging and of great use for the tumor-tracking radiotherapy. In this work, a deep learning based tumor localization method with single x-ray projection at arbitrary angle was proposed to address this problem. With the aid of principal component analysis (PCA)-based motion model, a deep learning network was established before treatment to recover the complex nonlinear mapping from single x-ray projection at any angle to the tumor motion. During treatment, volumetric image and tumor position could be obtained by applying the deep learning model on the acquired x-ray projection and the gantry angle. The method was validated with simulated data (XCAT). It was found that the method could achieve real-time tumor localization with high accuracy (< 0.2 mm in 3D).

ABSTRACT. Introduction: Iterative CT reconstruction problem typically contains regularization parameters to control resulting image quality. Tuning these parameters is of central importance. Manual parameter tuning is not only tedious, but becomes impractical when there exists a number of parameters. We develop a novel quality-guided deep reinforcement learning (QDRL) framework to intelligently evaluate the image quality and automatically adjust parameters in a human-like fashion.

Methods: The proposed QDRL framework is inspired by the natural learning process of human. It aims at training a parameter-tuning policy network (PTPN) to automatically adjust a large number of regularization parameters to improve reconstructed image quality. A quality assessment network (QAN) is trained together with PTPN to differentiate high-quality image patches in our training data pool of high-quality CT images and low-quality image patches generated during the reconstruction process. QAN outputs a probability for a patch to be of high quality, which serves as a reward function to guide the reinforcement learning process of PTPN. We consider an example iterative reconstruction problem with pixel-wise total-variation regularization, i.e. each pixel in CT image has a regularization parameter to be tuned.

Results: We successfully performed QDRL to train PTPN and QAN. Compared to manual parameter tuning, the trained PTPN is able to reduce reconstruction error by 7.7% on training data, and 6.1% on testing data. QAN is capable of evaluating the image quality, assigning higher scores to images of better quality.

Conclusion: We have developed a novel QDRL framework to automatically tune regularization parameters in the iterative CT reconstruction problem. Experimental results demonstrate the effectiveness of this approach. The framework is able to effectively handle a large-scale parameter-tuning problem with the number of parameters involved clearly beyond human capability.

ABSTRACT. Radiation therapy is one of the major cancer therapy modalities, accounting for two-thirds of cancer patients in the US, either standalone or in conjunction with surgery, chemotherapy, immunotherapy, etc. Radiotherapy treatment planning currently requires many feedback loops between the planner and physician. The physician’s preferences for a particular patient cannot be fully quantified and precisely conveyed to the planner. We present a real-time Pareto surface dose generation deep learning neural network that can be used immediately after segmentation by the physician, adding a tangible endpoint goal for the planner. From 70 prostate patients, we first generated 84,000 7-beam IMRT plans sampling the Pareto surface, using an in-house GPU-based proximal-class first-order primal-dual algorithm, Chambolle-Pock. We divided the data to 10 test patients and 60 model development (training and validation) patients. We then trained and fine-tuned a hierarchically densely connected convolutional U-net (HD U-net), to take the PTV and an avoidance map representing OARs masks assigned their respective w_s, and predict the optimized plan. We trained the model for 100,000 iterations using the Adam optimizer, with a learning rate of 1×10^(-4), using an NVIDIA V100 GPU. The HD U-net is capable of accurately predicting the Pareto optimal 3D dose distributions, with mean dose errors of 3.4%(PTV), 1.6%(bladder), 3.7%(rectum), 3.2%(left femur), 2.9%(right femur), and 0.04%(body) of the prescription dose. The PTV dose coverage had errors of 1.3% (D98) and 2.0% (D99). The neural network is capable of predicting the dose distribution in 1.7 seconds. Clinically, the optimization and dose calculation for IMRT takes approximately 5-10 minutes to complete. The implementation of such a framework would drastically reduce the number of feedback loops between the planner and physician. The valuable time that is saved would allow for the physician and planner to focus on more challenging cases and produce the best achievable plan.

ABSTRACT. In intensity modulated radiation therapy, the optimal choice of beam orientations has a significant impact on the treatment plan quality, influencing the final treatment outcome. In current treatment planning workflow, the beam orientation selection is from a standard protocol, and/or manually done by the planner, typically yielding suboptimal solutions. Beam Orientation Optimization (BOO) methods are used to find the optimal beam directions, and has been a research field of interest for many in radiation oncology. Modern BOO methods typically solve the problem in the radiation dose domain, which requires the precomputation the dose influence matrices (Dij) for all candidate beam orientations, and use them to perform fluence map optimization (FMO) during the beam selection process. The number of candidate beams can be very large. For example, there are 180 candidate beams for a coplanar geometry with 2 degree separation, and 1162 candidate beams for a non-coplanar geometry with 6 degree separation. Both the computation of Dij matrices and FMO are very complex and time intensive operations, taking on the order of hours for Dij calculation and minutes for the FMO process, ultimately hampering the BOO implementation in clinical routine. In contrast, artificial intelligence (AI) is an attractive tool for solving BOO problems, given its superior speed and promising results for medical applications. This work aims to develop a fast BOO method based on deep neural networks that provides a solution in about a second, and therefore, can be implemented in clinical routine. The images of 70 prostate cancer patients, which converted to more than 30000 input scenarios with augmented techniques, were used in exclusive sets to train, validate and test the AI model.

ABSTRACT. We present an improvement to our formalism in [1] wherein we devised the beam orientation problem in intensity-modulated radiation therapy as an approximate dynamic programming problem. Here, we improve the Monte-Carlo planning portion of the algorithm using a variant of the approximate policy iteration [2]: a batched alternating iteration of approximate policy evaluation and approximate policy improvement. Suppose that η is the cardinality of all beam angles in a treatment setup, which together with a patient’s geometry, Υ, constitute a deep neural network (DNN) policy input, ℧=[Υ,η]. A column generation DNN policy, Π, chooses a feasible angle based on a ℧, and generates a profile map, ζ ∈R^n, whose argmax informs the next candidate beamlet that must selected (n=180 in our implementation) until we fulfil a beam plan requirement. This is our a policy evaluation stage. We repeat the selection process for multiple patients and re-evaluate Π based on a mean square error cost function. Π essentially learns a mapping f:℧ ⟶ζ. With Monte-Carlo evaluations and tree search, we generate new training data in a policy improvement stage. We then project the new training data to the cost function space of Π. An alternating iteration of training Π and improving Π via our computed tree search policy, Γ, produces an improved policy, Π^* which we eventually use to evaluate treatment plans. A detailed description of our methods and results are presented in the supplementary document to this submission.

[1] Ogunmolu et al. Deep BOO! Automating Beam Orientation Optimization in Intensity-Modulated Radiation Therapy. Algorithm Foundations of Robotics XIII, Workshop, Merida, MX. Available at https://goo.gl/GDDEiQ. [2] Bertsekas, D. P., Approximate policy iteration: A survey and some new methods. Springer-Verlag, 2011.

ABSTRACT. Patients with locally advanced cervical cancer receive a high dose rate (HDR) brachytherapy (BT) boost after external beam radiotherapy (EBRT) as part of their treatment. Typical for this type of treatment is the prescription of a “pear-shaped” dose distribution. In this study we have developed and validated a system for fully automated multi-criterial HDR-BT planning for cervical cancer.

Our in-house developed system for automated multi-criteria EBRT planning (Erasmus-iCycle) was extended for optimisation of cervical cancer HDR-BT plans. Automated planning in Erasmus-iCycle is driven by a carefully tuned wish-list containing prioritised planning objectives and hard constraints. The developed wish-list aims at favourable dose delivery to the high-risk CTV (CTVhr) and sparing organs-at-risk (OARs) as much as possible, while also generating the desired pear-shaped dose distribution.

Automated plans were generated for 22 patients. The optimised dwell-times were then imported in our clinical treatment planning software for final dose calculation and evaluation. Automatically generated plans were compared to manually generated plans that have been delivered, so called clinical plans, both by analysing plan parameters and in blind pairwise plan comparison by an expert clinician (JWM).

All automatically generated plans were clinically acceptable. On average, the dose to the CTVhr was 8% (range -3% – 17%) higher for the automated plans (p < 0.001), without exceeding the constraints on the OARs (p > 0.05). For 20/22 plans, the clinician preferred the automated plan, for 1 patient the clinical plan was preferred, and for 1 patient there was no preference. Computation times were 20 seconds on average (range 5 s – 55 s).

This study demonstrates that fast (20 seconds) automated multi-criterial treatment planning for cervical cancer HDR-BT is feasible. Both according to attained plan parameter values, and in blind comparisons by an expert clinician, automated plans were of higher quality than manually generated plans.

ABSTRACT. Background: In a recent article, our group proposed a new direct aperture optimization (DAO) algorithm for intensity-modulated radiation therapy (IMRT) called fast inverse direct aperture optimization (FIDAO). When tested on fixed-gantry IMRT plans, we observed up to a 200-fold increase in the optimization speed. The purpose of this work is to extend and evaluate FIDAO for inverse planning of volume-modulated arc therapy (VMAT). Methods: A prototype FIDAO algorithm for VMAT treatment planning was developed in MATLAB using the open-source treatment planning toolkit matRad (v2.2 dev_VMAT build). Single 360° arc VMAT treatment plans were generated on the AAPM TG-119 phantom, as well as a sample liver and prostate cases. VMAT treatment plans were created by first performing fluence map optimization on 24° equispaced beams, followed by aperture sequencing and arc sequencing with a gantry angular sampling rate of 4°. After arc sequencing, a copy of the plan underwent DAO using our prototype FIDAO algorithm while another copy of the plan underwent DAO using matRad’s DAO method, which served as the standard algorithm. Results: Both algorithms achieved similar plan quality while optimization was considerably faster with FIDAO. The optimization times (number of iterations) for FIDAO and matRad respectively were: 65s (245) vs. 602s (275) in the TG-119 phantom case; 25s (85) vs. 803s (159) in the liver case; and 99s (174) vs. 754s (149) in the prostate case. Conclusion: This work demonstrated that FIDAO can offer faster DAO than conventional DAO methods for VMAT plan optimization.

ABSTRACT. Volumetric-modulated arc therapy (VMAT) treatment planning is an efficient treatment technique with a high degree of flexibility in terms of dose rate, gantry speed, and aperture shapes during rotation around the patient. However, the dynamic nature of VMAT results in a large-scale nonconvex optimization problem. Determining the priority of the tissues and voxels to obtain clinically acceptable treatment plans poses additional challenges for VMAT optimization. The main purpose of this paper is to develop an automatic planning approach integrating dose-volume histogram (DVH) quality criteria in direct aperture optimization for VMAT, by adjusting the model parameters during the algorithm. The proposed algorithm is based on column generation, an optimization technique that sequentially generates the apertures and optimizes the corresponding intensities. We take the advantage of iterative procedure in this method to modify the weight vector of the penalty function based on the DVH criteria and decrease the use of trial-and-error in the search for clinically acceptable plans. We evaluate the efficiency of the algorithm and treatment quality using a clinical prostate case and a challenging head-and-neck case, both from the CORT dataset. In both cases, our methodology obtained clinically acceptable plan for random initial structure weights with only a 10% increase in the computational time, while no plan was acceptable in simple VMAT. The results demonstrate the ability of the proposed optimization algorithm to obtain clinically acceptable plans without human intervention and also its robustness to weight parameters. Moreover, our proposed weight adjustment procedure proves to reduce the symmetry in the solution space and the time required for the post-optimization phase.

ABSTRACT. Non-coplanar volumetric modulated arc therapy (VMAT) has been shown to offer better prescription dose delivery to the tumor and better organ-at-risk sparing for cancer treatments. Previous methods limit the search space for the optimal beam orientations due to the large size and the combinatorial nature of the problem leading to local solutions. Our hypothesis is than in order to obtain a high quality non-coplanar VMAT treatment plan, globally optimal beam orientations have to be employed. The non-coplanar VMAT problem is formulated as a beam selection problem and solved using a simulated-annealing inspired algorithm capable of producing global solution to a combinatorial optimization problem. Treatment plans for three patient cases: a TG-119 case, a liver case and a prostate case are prepared using our proposed non-coplanar simulated annealing VMAT algorithm and compared the generated plans to our implementation of a state of the art approach on the basis of prescription dose accuracy, organ-at-risk sparing and delivery time. The results show an accurate delivery of the prescription dose to the target tumor volume and an improved organ at risk sparing when using our proposed method but at a cost of having longer delivery time. An increase in the number of control points has been observed when using our method compared to the state of the art method but conversely there is a reduction in the average delivery time spent at each control point. The reduced average delivery time spent each control point using our proposed method could be helpful in order to achieve better organ at risk sparing e.g. on the skin.

ABSTRACT. Although automatic dwell time optimization is common practice in HDR prostate brachytherapy (BT) treatment planning, this is less the case for automatic catheter position optimization for pre-planning. Recently, a bi-objective optimization model has been introduced, to automatically optimize dwell times and catheter positions simultaneously, creating a set of plans with different trade-offs between target coverage and organ sparing. The model can be optimized with the Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA), but this requires too large running times on an ordinary Central Processing Unit (CPU). In this work, we parallelize this optimization on a modern Graphics Processing Unit (GPU), and improve on the optimization model to ensure clinically feasible catheter positions.

Catheter positions are constrained to be inside or close to prostate and/or seminal vesicles, and outside of rectum, urethra, and bladder. The directions of all catheters are parallel. Plans are optimized using GOMEA on 20,000 dose-calculation points, and re-evaluated on 100,000 points. Parallelization is performed on an NVIDIA Titan Xp GPU by calculating the dose in the dose-calculation points in parallel. Experiments were performed on data of 3 prostate cancer patients treated with HDR BT. The catheter optimization model was run for 16 catheters, both sequentially on a CPU for 4.5 hours and in parallel on the GPU for 5 minutes. Next, the catheter optimization model was extended with a 2mm margin around rectum and urethra, taking into account the part of the urethra that extends below the prostate, and again run on the GPU.

Parallelizing catheter position optimization for HDR prostate BT on a GPU reduced the running time from 4.5 hours to 5 minutes, which is a necessary step for future applications. Moreover, the use of a 2mm margin around rectum and urethra improved the clinical feasibility of the treatment plans.

ABSTRACT. Portal dose images (PDIs) acquired using the electronic portal imaging device (EPID), are used to detect dose deviations during radiotherapy treatment. Objective automatic models to flag relevant errors in EPID-based doses are still lacking, providing potential for artificial intelligence. Gamma maps are often reduced to a few numbers, losing potentially important information. This work presents a deep learning method for flagging relevant dose errors using gamma maps based on EPID dosimetry simulations. Simulated PDIs were used as input for a convolutional neural network (CNN). Anatomical changes (tumor shift/tumor regression/pleural effusion) were introduced into CT images of 6 lung cancer patients. PDIs were simulated using the original treatment plan. PDIs of modified CTs were compared with the PDI of the original CT, by gamma analysis. Of the 372 resulting gamma maps, 248 maps (4 patients) were used for training, 60 maps (1 patient) for validation during training and 64 maps (1 patient) for testing. The gold standard classification was DVH-based. The CNN classification was compared with the clinical classification, where gamma maps are flagged if the gamma fail rate (GFR) exceeds 10%. The CNN consisted of 2 convolutional/pooling layers followed by fully connected and Softmax layers. The model was trained for 100 epochs using the Adam optimizer, with learning rate 0.0001 and dropout probability 0.2 to prevent overfitting. The CNN achieved test accuracy, area under the curve (AUC), sensitivity and specificity of 93.8%, 0.88, 78.6% and 98.0%, respectively. The clinical GFR classification was not able to flag any errors in the test set. Deep learning provides a promising, novel way of analyzing EPID dosimetry results for detection of dosimetric errors during radiotherapy treatment. This is a step forward towards automatic objective decision models for determining which patients may need treatment plan adaptation. This work is being extended to real patient data.

ABSTRACT. Introduction: Inverse treatment planning of radiotherapy requires human efforts to operate an optimization engine and adjust optimization parameters, e.g. dose volume constraints’ locations and weights, to achieve the best plan for each patient. This process is time-consuming, and the plan quality depends on planer’s experience and available time. Motivated by advances in deep reinforcement learning (DRL) to model human decision-making behaviors, we develop a virtual treatment planner that can automatically adjust optimization parameters in a human-like manner.

Methods: We develop a virtual planner network (VPN) under the Q-learning framework to operate an optimization engine. Similar to a human’s behavior, VPN repeatedly observes dose-volume histogram of a plan produced by the engine under a set of optimization parameters and outputs actions to adjust the parameters. This is continued until a high-quality plan is achieved. VPNs are trained via an end-to-end DRL process with experience replay technique and the epsilon-greedy algorithm. We demonstrate the feasibility and effectiveness of VPN in intensity modulated radiation therapy (IMRT) for prostate cancer and high-dose-rate brachytherapy (HDRBT) for cervical cancer.

Results: We successfully trained VPNs in IMRT and HDRBT scenarios. VPN spontaneously learns how to use an optimization engine to perform treatment planning by adjusting optimization parameters. We apply the trained VPN to five HDRBT and 63 IMRT testing patient cases. For HDRBT, quality score of plans generated by VPN is 10.7% higher than that of human-generated plans. For IMRT, all the VPN-generated plans reach the highest PlanIQ quality score.

Conclusion: This is the first time that intelligent treatment planning behaviors are autonomously generated through a training process and the trained VPN is capable of behaving in a human-like way to produce high-quality plans in testing cases.

ABSTRACT. Introduction: Imaging dose reduction in CT is important to reduce cancer risks caused by x-ray. The CT image quality is usually degraded due to amplified sinogram noise. Most existing methods consider suppressing noise in CT images. Sinogram denoising remains a challenging task, since reconstruction algorithms amplify high-frequency signals in the sinogram and hence inaccurate sinogram processing can causes reconstruction error deteriorating image quality. To achieve effective sinogram domain denoising, we propose a deep-learning based reconstruction-aware sinogram denoising approach that is trained by incorporating the end goal of reconstruction high-quality CT images.

Methods: We construct a denoisng network with a U-net structure. The clinical standard filtered back projection (FBP) reconstruction algorithm is implemented as fixed network layers with frozen parameters, and attached to the output layer of the U-net to perform reconstruction. To train the U-net, we consider the loss function, which is the weighted sum of the CT image domain loss function and the sinogram domain loss function. A higher weight is given to the image domain loss to emphasize the importance of the image reconstruction task. The two functions in the two domains are evaluated as the l_1 norm of the difference between the low-quality and the ground truth images in the corresponding domain, respectively. For comparison purpose, we also train another U-net with an identical structure by minimizing only the loss in the sinogram domain.

Results: Compared to the U-net trained with only sinogram domain loss, the proposed framework reduces the average reconstruction error from ~32 HU to ~28 HU. PSNR has been improved by 1.73dB on training images and 2.08dB on testing images.

Conclusion: We have developed a deep learning-based reconstruction-aware framework to address the challenging task of low-dose CT denoising in sinogram domain. It achieves high-quality CT reconstruction and outperforms the standard deep learning-based denoising method.

ABSTRACT. The aim of study was to investigate the influence of different cGAN generator structures on the image conversion and their impact on the mean absolute error (MAE). Further we transferred the learned features to different MRI scanners to identify if the network-learned features can be globally applied.

The four different generators were based on SE-ResNet (SN), DenseNet (DN), U-net (UN) and Embedded net (EN) using pelvic T2-weighted MR (0.35T) images of 40 patients. The test data set contained 12 patients. A gold-atlas dataset (GA) from the Swedish gentle radiotherapy project was used to evaluate the impact of the trained networks on other MR scanners (1.5 and 3T). Finally, an Ensemble model (EM) was designed by combining the converted images with a voxel-wise median calculation. The performance of the networks was evaluated by MAE in the whole-body contour.

MAE at the 100th epoch count was about 27-30HU for DN and SN and 39-42HU for EN and UN for the training data. During the testing phase, MAE of all 4 networks ranged between 43-46HU. For the GA dataset, EN performed worse with an MAE of 59-66HU. For both 3T systems the SN performed best in comparison to the other networks. EM produced the best results for all datasets.

Differences between networks should be considered if applied to new data. Detailed information on the networks’ performance should be reported in studies that utilize such methods. Our EM results suggest that a combination of multiple networks can increase the overall performance. The results indicate that a comprehensive comparison between generator types is necessary when using deep learning methods on medical image post-processing.

ABSTRACT. The application of artificial intelligence (AI) and deep learning in medicine often based on a large set of clinical data, which is suffering from nomenclature inconsistency and annotation inconsistency in most cases. The researchers must complete a time-consuming data-cleaning process before the analysis. Previous works tried to utilize deep convolutional neural network (CNN) to build a classifier and recognize organs at risk (OARs) automatically. Due to the extremely imbalanced dataset, the classifier tends to bias towards the majority categories and cannot generalize well on other datasets. To cope with these challenges, we propose a multiscale down-sampling data augmentation method and a composite mask feature to provide extra contextual information of OARs. Meanwhile, the model captures the global dependency from high-level semantic by adding the non-local block. Testing on the MICCAI-2015 head-and-neck dataset and the UTSW head-and-neck dataset, the final model gains 99.46% and 97.06% average recall rate respectively.

ABSTRACT. Purpose: The aim of this work is to adapt and evaluate a 3D U-Net convolutional neural network for the segmentation of pelvic bones in dual-energy CT (DECT). The results can be used to increase the accuracy of radiation therapy treatment planning by making more realistic assumptions on the material composition of patient tissues.

Materials and Methods: A 3D U-Net network was implemented using the Keras framework. The network was trained and validated for 25 patients and 3 input configurations denoted as H, M and LH: (i) high tube voltage images, (ii) mixed images, and (iii) low and high tube voltage images. A 5-fold cross validation scheme was used to utilize the data, which had to be downsampled to 128 × 128 × 128 to fit to the GPU's memory. Xavier and Adam algorithms were used for the initialization of weights and optimization, respectively. Dice coefficient was used as the loss function. Manual annotation of the mixed images was done with ITK-SNAP. Statistical significance was tested with a paired-samples t-test.

Results: The training took 16 hours for one model, in total 15 models were trained. Mean Dice coefficients obtained during the validation stage for the H, M, and LH configurations were 0.971, 0.972 and 0.973, respectively.

Discussion and Conclusions: The proposed 3D U-Net could segment bones in the DECT data sets; the Dice coefficient was 0.973. The method performed better for DECT images than for mixed images, which simulate images taken at 120 kV. The corresponding statistically significant (p=0.017) increase in the Dice coefficient from 0.972 to 0.973 was relatively small since the enhancements were mainly at the edges of the bones. The method can be easily extended to multi-energy CT by using additional input channels of the 3D U-Net network.

ABSTRACT. [Purposes] Eyeing on the future implementation of the marker-less real-time tumour detection using stereo x-ray colour fluoroscopic images, we tackle the problem of detecting lung tumour. The purpose of this study is to construct the model to detect the position and the shape of a lung tumour using the convolutional neural network. [Materials and Methods] A patient who underwent respiratory-gated radiotherapy using SyncTraX was enrolled in this study. Seven fields were created for treatment plan. For each field, colour fluoroscopic images were acquired at 30 fps for 10 s from two directions and total 4200 images were acquired. Colour fluoroscopic images were randomly selected 90 images for each direction. For RGB components, a lung tumour-recognizable one was selected. For one component images, supervised images were created and data augmentation was performed. The model that detects the lung tumour was created using U-net. To verify the model, ten colour fluoroscopic images during treatment simulation were selected randomly and the centroid of detected lung tumour using proposed model was compared with those of ground truth to evaluate the tracking accuracy. The dice coefficient was also calculated to evaluate segmentation accuracy. Furthermore, the proposed model was applied to colour fluoroscopic images acquired during treatment and the centroid of detected lung tumour were compared with a fiducial marker used as the internal surrogate. [Results] For all fields, the mean ± standard deviation of dice coefficient in test images was 0.95 ± 0.02 (0.85–0.98). However, the centroid positional errors were ranged from 0.1 mm to 4.6 mm. and it was very difficult to detect the invisible lung tumour in colour fluoroscopic images. [Conclusion] We have constructed the model to detect the position and the shape of a lung tumor. Further, it is needed to improve the model and investigate for many clinical cases.

ABSTRACT. Purpose: To quantify the modulation degrees in accelerated partial breast irradiation (APBI) intensity-modulated radiation therapy (IMRT) plans. Methods: A total of 52 APBI patients’ plan was selected. The treatment plan was generated with tri-Co-60 MR-image guided radiation therapy system equipped with the double-focused multi-leaf collimator (MLC). The modulation degrees were quantified by the following modulation complexity metrics (MCM): the plan averaged beam area (PA), plan averaged beam irregularity (PI), and plan averaged beam modulation (PM). Absolute differences in the total beam-on time, and the leaf position were obtained by using the DICOM RT plan objects and log files. Global gamma analysis with 3%/3 mm, 3%/2 mm, and 2%/2 mm criteria, was performed to validate the plan deliverability. The Pearson’s correlation coefficients (r) between gamma passing rates (GPR) and the sum of absolute parameter differences were obtained to demonstrate the impact of mechanical uncertainties on plan deliverability. Also, those between MCM and GPR were obtained. Results: The average GPR with 3%/3 mm, 3%/2 mm, and 2%/2 mm showed 98.72±1.42%, 96.33±4.12%, and 97.05±2.68%, and the average PA, PI, and PM were 40.13±14.40, 1.97±0.28, and 0.70±0.13, respectively. The sum of absolute differences in time and leaf position were 0.05±0.03 seconds and 4.21±1.77 mm, respectively. Beam-on time differences, PI, and PM had significant correlations with GPR of 3%/2 mm showing r of -0.306, -0.375, and -0.405 (p < 0.05). Conclusions: Statistical analysis with beam modulations, gamma analysis, and parameter differences present that the small aperture area and the excessive degrees of modulation could degrade in beam deliverability, and mechanical uncertainties are closely related to the plan complexity. The study outcomes make it possible to predict the treatment accuracy in advance of QA delivery in APBI IMRT plans.

ABSTRACT. Optimal beam orientation selection remains to be a challenging problem for intensity modulated radiation therapy today. Column generation (CG) has been shown to produce plan vastly superior to that of human selected orientations, especially for highly non-coplanar plans such as 4π Radiotherapy. While efficient, CG is a greedy algorithm and typically still finds suboptimal solutions. Recently, artificial intelligence (AI) plays an essential role in automation of complicated processes. In this study, we utilized a reinforcement learning structure involving a supervised learning network to guide Monte Carlo tree search to explore the beam orientation selection decision space. We have previously trained a deep neural network (DNN) that takes in the patient anatomy, organ weights, and current beams, and then approximates beam fitness values, indicating the next best beam to add. This DNN is used to probabilistically guide the traversal of the branches on the Monte Carlo decision tree to add a new beam to the plan. To test the feasibility of the algorithm, we solved for 5 beam plans, using 13 test patients that are different from the training and validation patients that originally trained the DNN. On average, the CG algorithm needed 700 seconds to find its solution, while the proposed method found solutions with higher quality, with respect to the objective value, in 100 seconds. Using our guided tree search (GTS) method we were able to maintain a similar planning targe volume (PTV) coverage within 2% error, and reduce the organ at risk (OAR) mean dose by 0.10 ± 0.08%(body), 2.44±2.01%(rectum), 4.94±4.65%(L fem head), 6.40 ± 3.94%(R fem head), of the prescription dose, but a slight increase of 1.31±1.96% in bladder mean dose. In this study we demonstrate that our GTS method produces a superior plan to CG, in less time than it takes to solve the CG algorithm.

ABSTRACT. Enabling dynamic couch and collimator rotations leading to dynamic trajectory radiotherapy (DTRT) might improve the delivery of coplanar standard volumetric modulated arc therapy (VMAT). This work investigates potential dosimetric benefits of DTRT when compared with VMAT for different treatment sites. A dedicated optimization framework for DTRT was developed. First, fractional volume-overlaps of the organs at risk (OARs) with the target is determined for potential beam directions penalizing directions where OARs are proximal to the target. The resulting two-dimensional gantry-table map is used as input for an A* pathfinding algorithm returning an optimized gantry-table path considering CT and collision restrictions. The A* algorithm is used again to determine the gantry-collimator path by optimizing the area between MLC leaves and the target contour. The resulting dynamic paths serve as input for the intensity modulation optimization using a research VMAT optimizer. This procedure was applied for different clinically motivated cases. Resulting dose distributions for DTRT and standard VMAT plans were compared based on dose volume histogram (DVH) parameters. Additionally, the delivery of all DTRT treatment plans was validated by means of measurements on a TrueBeam linear accelerator with either using the Delta4 system or gafchromic EBT3 films. Comparison of the DVHs for the target volumes showed at least the same coverage and dose homogeneity for DTRT using equal or a lower number of fields compared with VMAT for all cases studied. Depending on the case, improvements in mean and maximum dose for OARs of up to 50% for the DTRT plans were achieved compared with VMAT plans. Measured and calculated dose distributions agree generally within 3% or 3 mm. The results of this investigation demonstrate that DTRT has enormous potential to reduce dose to OARs, while target coverage is preserved compared with VMAT treatment plans. This work was supported by Varian Medical Systems.

ABSTRACT. Treatment planning for Intensity Modulated Proton Therapy (IMPT) can be significantly improved by reducing the time for plan calculation, which facilitates efficient sampling of the huge solution space characteristic of IMPT treatments. Additionally, fast plan generation is a key for online adaptive treatments, where the adapted plan needs to be ideally available in a few seconds. However, plan generation is a computationally demanding task and, although dose restoration methods for adaptive therapy have been proposed, computation times remain problematic. Here, we describe a fast implementation of our in-house, clinically validated, dose and optimization algorithm for IMPT. Calculation times have been reduced by the development of dedicated graphical processing unit (GPU) kernels, which perform all dose calculation and optimization steps required to generate a complete treatment plan. These include generation of water equivalent depth-maps, calculation of the distances to delineated structures, generation of the dose deposition matrix and optimization of all pencil beam fluences. Using an off-the-shelf GPU, the fast implementation is able to generate a complete new treatment plan in 5-10 seconds for typical IMPT cases, and in under 25 seconds for plans to very large volumes as for cranio-spinal axis irradiations. Although these times do not include the manual input of optimization parameters, they include all required computational steps, including reading of CT and beam data. In addition, no compromise is made on plan quality, with the resulting plans being close to identical, or better, to those generated using the clinical implementation. In conclusion, fast plan generation with a clinically validated dose calculation and optimizer is a promising approach for daily adaptive proton therapy, as well as for automated or highly interactive planning.

ABSTRACT. Introduction Prostate SBRT with a large number of non-coplanar beams may significantly enhance plan quality at a cost of long treatment times. Using our in-house developed optimizer for fully automated and integrated multi-criterial beam angle selection and IMRT profile optimization we have investigated reduction of the number of non-coplanar beams while maintaining high plan quality.

Materials & Methods Per investigated treatment approach of a patient (e.g. 30-beam non-coplanar IMRT), the applied Erasmus-iCycle optimizer automatically generates a single Pareto-optimal plan. Due to the multi-criterial optimization generated plans are also clinically favourable.

Initially, Erasmus-iCycle was used to generate for 20 study patients 1) a plan for axial dual-arc VMAT combined with 5 individually selected non-coplanar IMRT beams, and 2) an IMRT plan with 30 individualized non-coplanar beam directions (30NCP). It was observed that these plans were highly comparable in quality. Strong preferences occurred in the selected non-coplanar beam angles. A novel treatment approach was then defined, designated VMAT+CS, consisting of dual-arc VMAT supplemented with a class solution (CS) of 2 highly prefererred non-coplanar beam directions, i.e. for all patients beam profiles of the same 2 non-coplanar beam directions were simultaneously optimized with VMAT.

VMAT+CS was compared to 30NCP.

Results VMAT+CS and 30NCP had highly similar plan quality; differences were considered clinically irrelevant. However, treatment time reduced from 28.8 min to 8.2 min. Dosimetry demonstrated that all plans were clinically deliverable.

Conclusions Using a clinically applied, in-house developed optimizer for fully automated and integrated multi-criterial beam angle selection and IMRT beam profile optimization, a novel treatment approach was developed for prostate SBRT, consisting of dual-arc VMAT supplemented with 2 fixed non-coplanar IMRT beams. Compared to 30-beam non-coplanar treatment the treatment time reduced drastically while plan quality was highly similar. Fast, high quality treatment using only a few, well-selected non-coplanar beams is feasible.

ABSTRACT. In-room collisions between the linear accelerator and either the patient or the couch can cause significant injury and damage to the equipment if not prevented. Currently the workflow at our center for the Elekta Versa HD requires the use of Mobius 3D to estimate minimum clearance of the gantry with the patient and couch. For patients with clearance below tolerance, a clearance check is performed on the unit prior to treatment with the patient, adding several minutes to each patient setup. Additionally, collisions cannot be detected a-priori for breast patients laying on a breast board with their arms raised above their heads as portions of the patient anatomy are not included in the planning CT volume. We have developed a collision detection plugin for the Eclipse treatment planning system for non-coplanar geometries using the Eclipse Scripting Application Programming Interface (ESAPI). A detailed mesh model of the Elekta Versa HD was constructed using a series of 3D scans of the linac acquired with an Xbox Kinect. A video game physics engine (Bullet Physics) is used to detect collisions between the model of the linac and the couch as well as the patient. The couch is modelled using a series of simple rectangular prisms while the patients are represented using a cylinder. The current mesh model has been compared to in-room measurements and can differ by up to 1 cm due to the process of stiching together multiple scans to construct the model. Computer aided design (CAD) models of the linac and Hexapod couch will be implimented in the futre in order to refine the accuracy of the plugin. Verification using retrospective data of previously detected collisions along with further comparison to in-room measurements is ongoing.

ABSTRACT. Purpose: To demonstrate the impact of RIRAS, a Web-based onymous reporting system on safety in radiation oncology across the Veterans Health Administration nationwide. While most incident learning systems are designed to aggregate user reported incident data, RIRAS features secondary causal analysis and feedback provided by subject matter experts from the National Office. This generates greater contextual depth to accurately analyze each reported incident.

Methodology: Over 500 reported incidents were analyzed to obtain contextual information on the categories, causes, and safety barriers related to specific incidents. Broad categories were further divided into subcategories for the most commonly encountered failure modes. Causal analysis was performed to determine appropriate interventions depending on the process step where the incident occurred. Special attention was placed on actual events that were reported despite existing barriers.

Results: The most frequently reported incidents belonged to dosimetry (30%), patient setup (13%), and contouring (9%) categories while the highest number of subcategory events resulted from geometric misses due to isocenter placement and shifts (11), treatments delivered with incorrect plans (5), and setups requiring bolus (5). The majority of incidents were caused by human factors, with inadequate attention and training combining for 59% of events and 38% of total incidents. Also, 63% of events occurred despite existing barriers already in place, with 41% attributed to deficiencies during time-out. It was discovered during analysis that several events were misidentified as good catches at the user-end.

Conclusion: Secondary analysis of RIRAS data presents opportunities to classify radiotherapy incidents with greater contextual detail. This allows corrections to misattributed incident types, accurate determination of cause in ambiguous cases, and highlights to areas with systematic failures. The result is an improvement in the understanding of failure points, which are specific to each reported incident thus facilitating real learning and process improvements at each reporting service.

ABSTRACT. Online adaptive radiotherapy treatments on MR-Linac systems present a challenge to traditional plan checking methods. Plan checking remains crucial to prevent treatment errors but must take place in a time critical setting. A fast automated comprehensive checking system was developed for fully adaptive treatments of the Elekta Unity MR-Linac (Elekta, AB, Stockholm, Sweden). The checking system was designed to perform all the necessary treatment plan checks needed to confidently deliver hypofractionated radical treatments and includes an independent dose recalculation. The checking system consists of two independent parts developed separately and with independent codebases. These are an in-house developed program written in C++ and a set of python scripts running inside a commercial treatment planning system [RayStation (RaySearch, Stockholm, Sweden)]. Checks are duplicated in both systems to prevent missed errors due to false negative checks. The results of both checking processes are presented to the user in a single combined user interface. The checking process has been incorporated into the clinical workflow in such a way as to proceed in parallel with other essential planning tasks. The net effect of the speed of the checking system and the parallel workflow mean that plan checking results in zero time extension relative to a treatment fraction without plan checking.

ABSTRACT. Introduction For adaptation and automation to be safely implemented on a large scale in radiotherapy, plan QA needs to become trivial and fast, while filling the gap left by less human interaction. We propose a data-driven framework to replace conventional plan QA, built on three principles: 1. Predicting dosimetrical deviations of complex plans 2. Independent MU check 3. Determining likelihood of plans To this end we use plan complexity metrics (CM) derived from RT-plans, fractionation and diagnosis. This work presents the proof of principle of the framework for VMAT plans using Pinnacle3 v9.10.

Materials & Methods 1. Dosimetrical deviations: we use linear regression of CM to model a) the approximation of VMAT by discrete control points and b) relative mean dose differences due to ambiguities of the beam fit (BF). We compare this model to dosimetric outcome from in-vivo EPID dosimetry. 2. To capture gross MU errors we use a random forest regressor to use CM to predict the relative area weighted sum of MUs, normalized to 200 cGy fractions (MU200). 3. We perform outlier detection using an isolation forest, using the previous CMs. We artificially contaminated 3% of the plans and evaluated the AUC.

Results 1. Using a single CM to predict deviations due to arc discretization gives an AUC of 0.98. Linear regression using 6 CMs could predict BF ambiguity with an RMS of 0.04. The model correlates with the 90th percentile isocgamma from in vivo EPID dosimetry with R2=0.52 (p<0.01). 2. With 3 CMs 97% of all data is predicted within 50 MU200. 3. The AUC of outlier detection using 9 CM was 0.97.

Discussion & Conclusions We propose a framework for data driven plan QA predicting clinically relevant deviations. This approach provides an immediate QA assessment and fits in an automated workflow.

ABSTRACT. In the current era of machine learning, many new predictive models are generated yearly. However, their use is not widespread in clinical practice due to many barriers like time constraints during consultations and the difficulty of accessing relevant models. In order to encourage this, we have developed an easily accessible web-based interface to host clinically relevant models.

An internally validated linear regression model was created for the survival probability of 151 small-cell lung cancer patients 26 weeks after the start of Prophylactic Cranial Irradiation (PCI). Subsequently, a web-based interface was created to make it accessible to physicians. A Java service was built underneath the interface to allow direct communication between the model and the EHR system (HiX, Chipsoft, The Netherlands). Communication was performed by SOAP (Simple Object Access Protocol) queries through a business intelligence software suite (BusinessObjects). This allowed for automated filling of the input parameters and subsequent calculations.

The internally validated AUC was 0.71 and provides the probability of surviving 26 weeks after the start of PCI radiation. Input parameters included the tumour N- and M-staging, gender, and concentrations of lactate dehydrogenase and haemoglobin in the blood upon diagnosis. A web-based interface with a generic architecture was created that could directly communicate with the HiX EHR system in real-time. Additionally, adjustment of input parameters with subsequent recalculations. Finally, the interface was added inside the HiX EHR system, allowing physicians to access the model directly during the consultation. The interface is currently being tested in a prospective clinical trial.

This interface allows for automated generation of model outcomes at a moment’s notice with up-to-date information and within the EHR system. By increasing the usability of a model using this interface, clinicians are encouraged even more to utilize the added benefit that predictive models can bring in routine clinical practice.

ABSTRACT. Introduction

An inhouse-developed plan quality control tool (planQC) was built to check the quality of our automated prostate plans. This tool predicts a personalized dose-volume histogram (DVH) for each organ-at-risk (OAR) based on the anatomy of each individual patient. In this study the planQC tool was clinical implemented.

Materials & Methods

Historical data of 129 automatically planned clinical prostate patient plans which all fulfil the clinical criteria were used to train and test a prediction model. In our clinical workflow the planQC tool is initiated when the dicom data of a final clinical plan is exported from the treatment planning system to the planQC DICOM node, where the planQC tool application is executed. The application calculates the predicted and planned DVH curves and generates a personalized scorecard in a PDF report. Threshold levels of 3 Gy for dose and 3% for dose and volume metrics at the OARs are set to detect if a plan is eligible for replanning. Plans are replanned by changing the treatment technique template for each of the OAR dose parameters that can be improved and starting the Auto-Planning again.

Results

The planQC tool was incorporated into the clinical workflow of automated prostate planning on March 1st 2018. Currently, 50 clinical prostate plans have been checked by planQC. For 11 of the 50 patients the plans did not pass the check and further optimization of the plan was done. In general, the mean improvement for the 11 plans was 3.6 Gy/% for to be improved parameters of the OARs.

Conclusions

A planQC tool was introduced in the clinical workflow and can be used to guarantee the quality of automated prostate plans. It can be used to detect plans that deviate from the trained cohort and therefore possible can be improved.

ABSTRACT. As the complexity of radiotherapy treatments has increased there has been increasing focus on in vivo dosimetry to detect potential errors. In some countries in vivo has become mandatory or highly recommended by the national bodies, however, it can be a resource burden. EPID based dosimetry is ideal due to its fast acquisition, high resolution, digital format and ready availability at the treatment unit. The SunCheck platform PerFraction uses forward projection and an automated workflow to perform absolute dose in vivo dosimetry. A set of calibration images were acquired in a water equivalent phantom at a defined range of depths and SSDs to create a dose-per-signal calibration factor matrix for the EPID. The transit dosimetry is a 2D planar gamma analayis of individual beams against the expected dose map. The expected dose is calculated by creating patient specific factors using the plan file and calibration matrix and projecting the planned beams through the planning CT dataset . Anthromorphic phantoms were used to create a series of static, IMRT and VMAT plans. We have assessed the sensitivity to set-up and machine errors by introducing known errors to the delivered plans. The plans were created in Eclipse TPS and delivered on Varian TrueBeam linear accelerators. Patient data was collected for breast and palliative treatments to assess the efficacy of performing daily in vivo dosimetry. Gamma analysis of the delivered beams indicated that PerFraction is sensitive to geometric shift of 2mm, incorrect bolus placement and incorrect plan delivery. Machine errors due to output changes, MLC errors and incorrect energy were detected. The average gamma score for 19 palliative beams was 96.6% at 3%/3mm, 91.4% for 45 breast beams. Forward planned transit dosimetry can detect set-up and machine errors. The automated workflow allows this to be utilized for all patients in a busy department.

ABSTRACT. Introduction HighTech radiotherapy capable to provide complex dose delivery modalities became widespread, making essential to evaluate clinical machine performances. Patient-specific verifications consist of dose delivery in a phantom before patient treatment. This operation of Delivery Quality Assurance (DQA) is repetitive and involving both workforce and Linac bunker occupational time. To work around this problem we developed a neural network capable of predicting passing rates a priori for Helical Tomotherapy (HT) DQA. In this paper we evaluated net performances, focusing on learning quality in function of specific machine parameters.

Materials & Methods For net training we used 734 DQA plans calculated with HT treatment planning station and measured in Octavius phantom with a PTW 2D-729 chamber array. We chose as input planned sinogram and plan parameter information extracted from the machine database files. As output of the training, we used the measured map of 729 dose values.

Results We focused on conv-net as: LENET, VGG-Like and ResNet-like network. Then we performed a tuning of the hyperparameters with a k-fould x-validation as a result of limiting data sample. As Loss Function the Mean Squared Error (MSE) was chosen, to test the accuracy we have done average calculation on all map dose value, calculating the difference between every single pixel of the two maps: simulate and measured. MSE mean value on all dataset plan is 12 cGy^2.

Discussion & Conclusions Failure in reproducing dose distribution for some DQA could be ascribed to low statistic in the training set or to an unusual behavior of the machine respect to the mapping function represented by our neural network. This issues will be investigated. In conclusion, results of this study demostrated that deep learning with convolutional neural networks can become a valid support for an efficient routinely patient- specific quality assurance in HT.

ABSTRACT. Spot scanning proton therapy is a promising form of external beam radiotherapy but comprehensive Proton Therapy software packages are not readily available commercially. The need for added quality assurance and improved treatment planning led to the development of a powerful, comprehensive Proton Therapy application with an intuitive web-based interface that seamlessly integrates into the Mayo Clinic Radiation Oncology workflow. The challenge was to architect the system that would incorporate multiple highly complex algorithms into the institutionally standardized framework. The current set of modules includes dose verification, patient-specific quality assurance, and robust plan optimization. The core technology consists of highly optimized, compute-intensive executables designed for a Linux environment. Each module is designed for efficiency, plan improvement, and plan quality/safety. For Dose Verification and Robust Optimization (RO), users choose between a CPU-based analytical and a GPU-based Monte Carlo model. Due to computational requirements, RO is carried out on a Linux cluster with 480 CPU nodes and 32 GPU nodes. Dose Verification runs on a single Linux server. The front end and Patient Specific QA (PSQA) module of the Proton therapy application is Windows based. All three modules are currently being used at Mayo Clinic Proton Therapy centers. Robust Optimization completion times vary significantly depending on the size/complexity of the treatment plan, the RO method chosen, and the cluster computing resources specified/available, but the range is from several minutes to several hours. Compared to TPS optimization, preliminary tests of the Mayo RO module show clear improvements in plan quality and robustness. The PSQA application has resulted in an average time savings of 55% per patient plan compared to previous manual methods. We conclude that incorporating the Mayo developed Proton Therapy application into the overall clinical workflow provides additional safety and more effective treatment for proton patients, without significantly impacting practitioner workload.

ABSTRACT. Introduction Our institute’s clinical workflow critically depended on a database infrastructure in dBase III file format, which is vulnerable as it is file based. We replaced this database with a SQL-server based infrastructure. We faced the challenge of replacing the infrastructure and twelve applications simultaneously, without interfering with clinical routine. Besides a higher degree of maintainability, a modern infrastructure enables our institute to follow an innovative automation strategy.

Scope & design We built the SQL infrastructure next to the clinical database, running both for the duration of the project. We first adapted the applications to write to both databases, and read only from dBase III files. Once we had a full dual-write infrastructure, we started a second round of adaptation to read from the new database. We adopted a layered database design. The database is called through stored procedures, which are accessible through an in-house developed DLL. The applications access the database through a module that handles the communication with the DLL.

Testing & QA All fully new code was written with relevant unit tests in place. Each piece of code adaptation was reviewed by a second developer. We adopted the Jenkins suite for continuous integration, and we used a risk analysis of each application to define manual tests that cover the risks to an acceptable level.

Agile development In order to be flexible enough to adapt to changes in requirements based on new information, we adopted the Scrum methodology. This enabled us to incorporate changes in the design and functionality when needed.

Conclusion Dealing with legacy code in a clinical infrastructure is challenging, due to the mission critical nature of the code. Using a layered architecture, unit and integration testing and Agile organization we replaced our legacy with modern software, while maintaining clinical safety and continuity.

ABSTRACT. Domiciliary care for oncologic patients may have clear advantage in terms of improved quality of life, psychological well-being and cost effectiveness. Coordination and collaboration between oncologists and patients are essential in order to guarantee the healthcare provision and treatment adherence in cancer home therapy. FILO BLU is a software application that aims to improve the communication in the doctor-patient relationship composed of two messaging APPs for smartphone, one for the patient/caregiver and one for the medical team, it is equipped with a module for the interoperability with portable medical monitoring systems and it is integrated with the patients’ electronic medical records. This allows the doctor to respond to requests having always available all the clinical information. To improve decision making and workload management we develop an expert system for the analysis of medical-patient communications that aims to score the patient's clinical status and that analyzes the flow of communications in order to signal to doctors, through an "attention" score, potential critical situations keeping into account both the written texts and any physiological values monitored. We generate synthetic data, composed of a simulation of the patient physiology and semi-automatically generated sentences, to test operative workflow. To easily combine numerical and textual data sources for the classification we choose a deep learning approach. We tested multiple neural networks architectures used in sentiment analysis to classify patients’ messages according to the severity of the clinical status both alone and in conjunction with physiological parameters recordings. While we do not expect that our synthetic data can replace data gathered through the usage in clinical settings, it creates a controlled test environment for classification and context sensitive spelling correction algorithm.

ABSTRACT. Uncertainty simulation for radiotherapy provides insight into the robustness of plans. For training, a web based implementation is beneficial. Because the simulations are computationally expensive, it is important to choose the appropriate web platform. The aim of this study is to develop such a simulator and compare its performance on two web platforms. We assume shift invariance, which is reasonable for photon therapy. Simulation of a treatment course is achieved by moving the planned dose distribution using randomly generated parameters followed by accumulation over fractions. JavaScript was tested first. Transforming a single dose slice in on my computer took 40ms. E.g. simulating 36 fractions in 2D takes ~1.5s. This is too slow for our requirements. The second option is WebGL. This platform uses the GPU and the software is provided in the form of a shader program, passed with the HTML page, which is run for each screen pixel and automatically parallelized on the GPU. The shader code contains a pseudo-random Gaussian random generator with passed seed (each pixel must move the same) to generate the systematic and random errors. The total dose per pixel and fraction is looked up using the sum of random and systematic errors translating the screen coordinate. Finally, the dose is summed and combined with the CT and segmented CTV and OAR structures for display. The performance is amazing, a 40 fraction treatment is easily rendered in 2D at 60 fps (Fig. 2), even on low-end hardware. A full 3D calculation (50 slices), takes less than 1s. Changing parameters immediately updates the result. In conclusion, we compared two web platforms to perform radiotherapy uncertainty simulation. Only WebGL meets speed requirements, i.e., it is feasible to perform uncertainty simulation in a web browser with WebGL. A limitation is that older browsers may not support WebGL.

ABSTRACT. Introduction QATrack+ is a free, open source and web based radiation therapy machine quality control (QC) application. Among its many features, it has the capability to execute python code in the background. One recent project at our institution was to integrate within the QATrack+ application as many QC test analyses as possible. The objectives of the project were to move all QC analysis to an open platform, improve QC time efficiency and provide a streamlined solution for QC analysis code development through distributed version control. Materials & Methods A python 3.6 package was created containing all scripts, functions and classes needed for test analysis of every imaging and treatment units. It is installed on the QATrack+ department server. The package source code is version controlled with Git and a “central” Git repository runs on our departmental server. Results The package is divided into several functional modules that correspond to the various tests used at our institution: 3D imaging, DailyQA3 data import, jaw and MLC field sizes, MLC picket fence, beam profiles, Winston-Lutz test, Tomotherapy specific tests and Brachytherapy HDR treatment parameters verification. An estimated 75 hours of monthly QC time per year are saved by having these tests integrated within our QC application. About 150 hours of daily QC time are saved per year by using the DailyQA3 data import module. Discussion & Conclusions The integration of all QC test analysis under a single python package with QATrack+ has freed our QC analysis from expensive licensed software. It has lead to a sizable reduction in QC time spent at treatment and imaging units. By using Git to manage the package source code, we have simplified version control and distributed edition of the source code for our department staff.

ABSTRACT. The Swiss cheese error detection (SCED) method performs real-time treatment delivery QA by intercomparing predicted and measured EPID images. SCED divides error detection into a series of tests to identify specific errors. The purpose of this study was to develop a reliable method to detect aperture shape errors. EPID image frames at ~10 Hz were acquired for 27 beams from 15 VMAT/SBRT clinical treatment plans. The deliveries were assumed to be error free. Delivery errors were simulated by systematically shifting each active MLC leaf for each control point for each beam of each plan 1 to 5 mm outward for the aperture prediction. A logistic regression model with a 2-dimensional input space (missed in-aperture radiation and excess out-of-aperture radiation) and pass/fail output was tuned using machine learning for optimal per-frame error detection. The model accuracy per systematic MLC shift was determined by grouping each beam in pairs, i.e., baseline results (no error) and results with the MLC shift (error). With a per-frame mask composed from union the apertures anticipated during the frame acquisition plus a ±2 mm buffer to accommodate in-tolerance MLC positioning errors to distinguish areas of missed and excess radiation, the SCED aperture check was unable to distinguish 1 mm systematic MLC shifts (as expected). Accuracy and AUC with the logistic classifier improve with for larger systematic shifts, with per-frame accuracy >95% and >99% for ≥3 mm MLC shifts. With ~1000 frames/beam, the single-frame false-positives rate would prohibit clinical usage. However, action thresholding based on adjacent-frame error-detection or requiring a given fraction of the total delivery to be in error can reduce the per-beam false positive rate to ~0. The aperture error logistic classifier trained on systematic MLC shifts has high accuracy for other aperture error scenarios. Missed and excess radiation detection yield reliable aperture error detection.

ABSTRACT. To propose a new software application (JAWS3D) to help dentists on assessing the radiation dose delivered to the teeth, maxilla and mandible using an interactive 3D visualization environment. The proposed software is implemented in Python and runs Mayavi engine for an interactive 3D visualization. The graphical user interface is built using Qt librairies. This software is cross-platform and can access to CT images, structure set and dose distribution maps stored in our radiation oncology database using the DICOM-RT format. As a first proof of concept, two clinical process are presented to assess dose distribution for dental application using the same patient dataset. First approach make use of a clinical software specifically designed for treatment planning in radiation oncology to visualize in 2D a set of CT images with isodose curves at 40, 50 and 60 Gy. The second approach employs JAWS3D for the dentist to examine rapidly the dose distribution data using a simple and fast 3D visualization environment. JAWS3D allows the dentist to assess dose distribution without requiring the staff in radiation oncology to prepare a specific dosimetry report. This assisting tool provides an interactive environment to inspect a 3D surface of a region of interest and adjust the dose range to be displayed. JAWS3D is designed specifically for dental applications, which avoids complicated and expansive proprietary software often lacking in 3D features. The development of JAWS3D will include clinical trial, automatic DICOM file transfer and web-based implementation. This software may also provide a useful pre-radiation platform for radio-oncologists and dentists to make concerted efforts to deliver better treatment. The 3D visualization environment is customizable and meant to be extended to other applications in radiation oncology.

ABSTRACT. Purpose: The primary aim of this study was to investigate whether CT image-derived radiomics are able to predict overall survival (OS) of patients diagnosed with primary Head and Neck Squamous Cell Carcinomas (HNSCC).

Material and methods: 5 independent cohorts, 411 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. CT scans with visible artifacts within the GTV were excluded. The dataset was split into training set (cohort 1, 2 and 3, n=308) and validation set (cohort 4 and 5, n=103) with the ratio of 3:1. A total of 1105 features were extracted from the GTV via PyRadiomics. Radiomic features were pre-selected to reduce the probability of over-fitting, and then we analysed for their prognostic power using the median value in the training set as the threshold value in the univariable analysis. The logrank test was used to assessed whether the individual feature can stratify patients into high and low risk groups. Multivariable Cox proportional hazards regression model was used to examine the association between survival and radiomic features. The concordance index (c-index) was determined to assess the models discrimination power.

Results: Eight radiomic features were selected: (1) log-sigma-1-0-mm-3D_glszm_LowGrayLevelZoneEmphasis, (2) log-sigma-2-0-mm-3D_gldm_DependenceVariance, (3) log-sigma-2-0-mm-3D_glszm_SmallAreaLowGrayLevelEmphasis, (4) log-sigma-3-0-mm-3D_gldm_DependenceEntropy, (5) wavelet-LHH_firstorder_Median, (6) wavelet-LLH_gldm_DependenceVariance, (7) original_firstorder_Maximum, and (8) original_glszm_SmallAreaLowGrayLevelEmphasis. The Cox regression model resulted in a c-index of 0.70 (95%CI: 0.63–0.74) on the training set and a c-index of 0.65 (95%CI: 0.56–0.68) on the validation set. The prognostic score demonstrated significant differences in OS between risk groups in both training (X2 17.0, p<0.001) and validation sets (X2 5.4, p=0.02). Conclusion: The radiomics signature showed a promising performance to predict overall survival of HNSCC patient The further study will investigate the prognostic performance models combining radiomics and clinical factors.

ABSTRACT. Introduction

Abundant literature suggests that radiomic features are significantly associated with overall survival in many cancers; in one such example, gray-level co-occurrence matrix (GLCM) features were found to be significant in pancreatic ductal adenocarcinoma (PDAC). In order to probe potential links between texture features and underlying biology, we utilize sub-micron resolution pathology images to investigate the effect of progressive image blur on such features in PDAC.

Methods

A custom-built algorithm was leveraged to classify individual pathology images into different cell types. Binary masks were created from sample regions containing epithelial, adipose and stromal tissue. From these binary masks, synthetic parameterized pathology images were generated by matching cell concentration, image entropy and average cell diameter.

The synthetic images were down-sampled using a simple box filter across three orders of magnitude – from conventional pathological resolution (µm) to clinical imaging resolution (mm). For each synthetic pixel size, five standardized GLCM textural features were extracted: energy, entropy, correlation, dissimilarity and inverse difference normalized.

Results

Qualitatively, similar patterns in the behaviour of these textural features can be seen across all tissue types. Inverse difference normalized varies little across pixel sizes of three orders of magnitude. Values for entropy and dissimilarity show a distinct maximum at the average cell diameter for the tissue type; conversely, energy shows a minimum at this pixel size. Feature values for correlation diminish as a function of image blur. For all tissue types, down-sampling from the highest resolution to the lowest one caused the distribution of feature values to become binary.

Discussion

This preliminary study indicates that image blur from micro- to macro-scale can significantly alter GLCM texture feature values; additional effects of image noise will be included in future work. With this type of investigation, a biological basis for CT radiomics classifiers could be postulated.

ABSTRACT. Purpose In radiation oncology, Machine Learning (ML) classification publications are typically related to two outcome classes, e.g. the presence or absence of distant metastasis. However, multi-class classification problems also have great clinical relevance, e.g., predicting the grade of a treatment complication following lung irradiation. In multi-class cases, AUC is not defined, whereas correlation coefficients are. Thus, it may seem like solely quoting the correlation coefficient value (in lieu of the AUC value) is a suitable choice. We illustrate via the use of Monte Carlo models why this choice is misleading.

Materials & Methods A predictive feature was modeled as the sum of Gaussians (unit standard deviation, each of size 10^4), the number of Gaussians being equal to the number of outcome classes (i.e., each Gaussian corresponded to a specific outcome class). There was always a Gaussian centered at 0 (associated with outcome = 0). The other Gaussians were centered such that the distance between two adjacent Gaussians was the same. The inter-Gaussian separations were chosen by trial-and-error. There were two experiments conducted: (1) the Pearson correlation of the predictive features was conserved as the number of outcome classes increased from 2 to 5, and (2) the prediction accuracy of the features was conserved as the number of outcome classes increased from 2 to 5.

Results In Experiment 1, the prediction accuracy of features decreases as the number of outcome classes increases. When going from 2 to 5 classes (R fixed at 0.45), the accuracy plummets from 0.69 to 0.31. In Experiment 2, the Pearson correlation increases as the number of outcome classes increases. When going from 2 to 5 classes (accuracy fixed at 0.69), R spikes from 0.45 to 0.93.

Conclusions No multi-class results should be allowed to pass scrutiny unless performance metrics in addition to correlation coefficients are presented

ABSTRACT. Prediction modelling with radiomics is a rapidly developing research topic that requires access to vast amounts of imaging data. Methods that work for decentralized data are urgently needed, because concerns about patient privacy and data ownership often prohibit data sharing. Previously published computed tomography medical image sets with gross tumour volume (GTV) outlines for non-small cell lung cancer have been updated with extended follow-up. In a previous study, these were referred to as Lung1 (n = 422) and Lung2 (n = 221). The Lung1 dataset is made publicly accessible via The Cancer Imaging Archive but Lung2 is remains a private collection due to patient privacy concerns. We performed a decentralized multi-centre study for a lung cancer radiomic signature validation that did not involve data exchange between institutions. Validation results were the same as those performed with fully centralized analysis. We foresee that the public dataset can be further re-used for testing radiomic models and investigating feature reproducibility.

ABSTRACT. Purpose. Radiomics are quantitative features extracted from routinely acquired medical images. These features can be used to predict, for instance, treatment toxicity or survival. However, many radiomic features depend not only on tumor properties but also on non-tumor related factors: scanner signal-to-noise ratio (SNR), reconstruction kernel and other image acquisition settings. This causes undesirable ‘noise’ in the radiomic features that will reduce the performance of prediction models or even obstruct model development. Phantom radiomics studies refer to reproducibility –removal of unstable features. However, this instability is a measurement artifact, and there is no a-priori reason why these ‘unstable’ features could not contain relevant information about the tumor. In this paper, we investigate whether we can use phantom measurements to characterize and even correct for the major source of instability – scanner SNR. Method. We used a phantom with 17 regions of interest (ROI) to investigate the scanner SNR influence. CT scans were acquired with 9 different exposure settings. We developed an additive correction model to reduce scanner SNR influence. Result. Sixty-two of 92 radiomic features showed high variance due to the scanner SNR. Of these 62 features, 47 showed a significant standard deviation reduction at least by a factor of two. We assessed the clinical relevance of radiomics instability by using a 221 NSCLC patient cohort measured with the same scanner. Conclusions. 1) 47 out of 62 non-reproducible radiomic features are correctable with additive linear regression models, 2) different ROIs react to scanner SNR differently affecting their radiomic feature versus scanner SNR trend, 3) radiomics correction is relevant in clinical studies. Corrected radiomics are less obstructed by scanner SNR and thus are more reliable. Our approach can be relevant to the deep learning radiation oncology applications: without knowledge of scanner SNR, a model is blind to this irrelevancy in data.

ABSTRACT. Predicting radiotherapy treatment outcomes from radiomic features is subject to many limitations. One of the major limitations currently is the scale of data sets that can be compiled to produce generalizable model predictions and to sufficiently validate models to warrant wider use. In this study we use distributed machine learning techniques, to develop and validate models without the need to pool data.

Data sets for patients with unresectable Stage I-IIIB, non-small cell lung cancer (NSCLC) treated with curative radiotherapy (>45Gy) were collected at three cancer therapy clinics. The target was to predict 2-year overall survival and a total of 538 patient data sets were included in the study. A set of 76 radiomic features including first-order statistics, shape and texture features, were computed for the Gross Tumour Volume (GTV) region of interest in each patient planning CT as segmented in routine clinical practice. A distributed learning system developed in-house was used to train and validate a logistic regression model with Tikhonov regularization. We randomly sampled and stratified 60% of the data at each centre for training and used the remainder for testing.

The AUC on the combined training data set was 0.66 whereas on the test data set it was 0.62. The weights with the largest magnitude were Gray-Level Run Length Matrix Gray-Level Variance (GLRL_GLV), Gray-Level Co-occurrence Matrix Sum of Squares Variance (GLCM_SOSV) and GLCM Sum Average Variance (GLCM_SAV). For the two significant co-occurrence features there was a high covariance and that overall 15 principal components of the combined data set account for 95% of the variance.

Distributed learning platforms enable the combination of data sets across clinics while preserving the privacy of the individual patients. This work demonstrated the feasibility of conducting statistical modelling on radiomics data with our software platform.

ABSTRACT. Metallic dental implants produce noisy artifacts in Head and Neck (H&N) CT images. Such artifacts distort image characteristics relevant for applications such as radiomics-based modelling and image segmentation. We propose a Random Forest (RF) classifier to automatically identify CT slices affected by dental artifacts. The classifier was trained using axial slices from CT scans of 44 (H&N) cancer patients from our institution. A total of 26 gray-level co-occurrence matrix (GLCM)-based radiomic texture features were used in building the RF model. The classifier was evaluated on two datasets: (1) internal set of 10 H&N CT scans from our institution, resulting in 84.06% sensitivity and 98.45% specificity and (2) external set of 24 H&N CT scans from an open-source archive, resulting in 77.78% sensitivity and 99.2% specificity. The reduced sensitivity when applied to the external dataset may be explained by heterogeneity in acquisition parameters across multiple institutions. The classifier also misidentified noisy slices in both internal and external datasets when the streaks were less prominent or localized to a limited section of the image. This could potentially be improved by exploring other relevant image characteristics.

ABSTRACT. Introduction: Radiomics is an emerging field of research that has potential to directly impact the clinical practice of radiotherapy. Radiomic features can be implemented in a variety of ways, differing for example by shifts or scales, and implementation differences can also impact repeatability. The International Biomarkers Standardization Initiative (IBSI) was created to establish a unified description of radiomic features and provide benchmarks for validation. A radiomics module was implemented using the open source DICOMautomaton software platform. DICOMautomaton provides a compelling base for a radiomics extraction tool since it was designed to quantify images and radiotherapy dose metrics. In particular, radiomic evaluation can be conveniently integrated into an automated clinical workflow via DICOMautomaton. In this work we validate the DICOMautomaton radiomic feature extraction module against the IBSI benchmarks for a non-small-cell lung carcinoma patient CT data set, IBSI feature definitions, and the reported consensus-derived benchmarks.

Methods and Materials: A subset of 27 radiomic features defined by the IBSI that are consistently found to be strong factors in radiotherapy applications were implemented. DICOMautomaton uses contours directly rather than converting to bitmap masks. While this has advantages for contour demarcation and resolution-independent adaptive resampling, the definition of some IBSI features were modified or simplified to suit. The IBSI corpus comprises image sets from four patients. Features were compared using a Wilcoxon's sign test with alpha=0.05.

Results: The value of individual features differed from the IBSI consensus benchmarks by no more than 0.75%. The two-tailed Wilcoxon sign test for all paired features resulted in a z-statistic of -1.603 and p>0.11, implying there is no significant difference between the IBSI and DICOMautomaton feature definitions and implementations.

Conclusions: DICOMautomaton is an open source platform ideal for robust radiomic analysis. The DICOMautomaton feature extraction module conforms to IBSI benchmarks.

ABSTRACT. Throughout the radiotherapy process, geometrical changes in the patient often occur resulting in organ deformations. These changes introduce deviations between the planned and delivered dose to organs at risk (OARs). To account for this uncertainty, the effect of deformations must be modelled in order to be incorporated into advanced treatment planning. We propose a method to statistically model time varying organ deformations for later use in probabilistic planning.

We used the planning CT (pCT) and five weekly cone-beam CT (CBCT) scans from 10 head and neck cancer patients. We performed a group-wise registration between all patient’s pCTs to register them to the same frame of reference. We then registered each week’s CBCT to the pCT for each patient and mapped the resulting vector field to the group-wise reference. We created 5 organ deformation models using principal component analysis (PCA), one for each week.

30 fractions (6 per week) were simulated by creating potential deformations from each PCA model to warp the dose distribution. The warped dose distributions were summed over all fractions to evaluate the dose delivered to each OAR for each of 100 simulated treatments per patient.

Averaged across all patients, the mean dose to the left parotid increased by 0.1 Gy and the right decreased by 0.3 Gy. The average maximum dose to the brainstem decreased by 1.0 Gy. The largest change in a single treatment lead to an increase of 7.0 Gy, 5.7 Gy and 6.0 Gy for the maximum brainstem dose and mean left and right parotid doses respectively.

We have implemented a methodology that allows us to simulate weekly organ deformations for individual patients. This method can be used in probabilistic planning to optimise treatment plans directly accounting for organ deformation. For the treatments simulated, we found large changes in OAR doses.

ABSTRACT. Cone beam computed tomography (CBCT) images can be used for dose calculation in adaptive radiation therapy (ART). The main challenges are the large artefacts and inaccurate Hounsfield unit (HU) values. Currently, deformed planning CT images are often used for this purpose, although anatomical accuracy might be a concern. Ideally, we would like to convert CBCT images to CT images with artifacts removed or greatly reduced and HU values corrected while keeping the anatomical accuracy. Recently, deep learning has achieved great success in image-to-image translation tasks. It is very difficult to acquire paired CT and CBCT images with exactly matching anatomy for supervised training. To overcome this limitation, we developed and tested a cycle generative adversarial network (CycleGAN) which is an unsupervised learning method and does not require paired training datasets to synthesize CT images from CBCT images. The synthesized CT (sCT) images have been compared with the deformed planning CT (dpCT) showing visual and quantitative similarity with artifacts being removed and HU value errors being reduced. Mean absolute error (MAE) and root mean squared error (RMSE) of sCT using CycleGAN are decreased to 27.98 HU and 86.71 HU from 71.78 HU and 167.22 HU of CBCT. Structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) of sCT are increased to 0.85 and 30.67 from 0.77 and 25.22 of CBCT against dpCT. HU line profiles of the sCT image show less noise than those of the CBCT image and approach to those on the dpCT image.

ABSTRACT. The use of deep neural networks to learn from previous clinical cases and automate different parts of the radiation treatment workflow (i.e. diagnosis, segmentation, planning, and delivery) is becoming very popular. In particular, convolutional neural networks (CNN) have been used to predict the optimal dose distribution for a new patient, which can be later used as guidance for the optimizer to automatically generate a treatment plan. However, the existing methods only use patient anatomy as input to the CNN, assuming consistent beam configuration in the training database. The purpose of this work is to develop a more general model that, in addition to patient anatomy, also considers variable beam configuration, which is crucial in cases where the beam configuration varies from patient to patient. The proposed anatomy and beam (AB) model combines two architectures, UNet and DenseNet, and uses mean squared error as loss function. Ten input channels are used to include anatomical and beam setup information (expressed in the dose domain). The AB model is compared with our previous work, the anatomy only (AO) model, where only nine channels are used for anatomical information. Both models were trained/validated with a set of 100 lung IMRT patients and tested with 29 patients. The AB model outperformed the AO model, improving the average error of the predicted doses with respect to the clinical ones in about 1% to 2% of the prescription dose, for relevant DVH metrics. In addition, Dice similarity coefficients were used to address the spatial correspondence of the predicted and clinical dose distributions, showing an improvement of 10% in the low to medium dose region for the AB model with respect to the AO model. Therefore, the AB model is robust to variable beam configuration, achieving a more comprehensive automatic planning with a potentially easier clinical implementation.

ABSTRACT. Radiation therapy treatment planning for head and neck cancers is a complex and time consuming task. Our lab has developed a deep learning based model that quickly generates 3D dose distributions that are very close to the realistic clinical plans delivered to patients. These predicted plan represents an average dose distribution based on the clinical data used to train the model. The quality of a predicted plan is therefore dependent on the quality of clinical plans in the training dataset. We believe if we can gradually identify and remove the lower quality plans from the training dataset, the overall quality of the remaining dataset will become higher, and the trained model will be able to predict dose distributions with increasing quality. Here, we propose a method for the training dataset to self-evolve to drive the predicted plans towards higher quality. 126 H&N patients treated in our clinic were selected. The patient cohort includes over 30 different H&N cancer sites, that include 1 to 5 PTV levels with prescription doses that ranged from 42.5 to 72 Gy. A plan quality scoring system considered the following dose volume criteria: D95 to PTV-high, D105 to PTV-high, D0.03cc to PTV-high, Conformation Number for PTV-low, D0.03cc to Spinal Cord, D0.03cc to Brain stem, D0.03 cc to right cochlea and D0.03 to left cochlea. Our adaptive retraining procedure increased the mean plan score from 19.5 to 22.3 for the testing patients and reduced the mean percentage difference in scores from 42.3% to 30.8%. Our method allows a training dataset to self-evolve by iteratively identifying and removing lower quality plans. A model trained on the improved dataset can predict dose distributions of higher quality. This self-evolving dataset allows the dose prediction model to get better over time as higher quality dose plans are created in the clinic.

ABSTRACT. Purpose: Radiation therapy treatment planning is a trial-and-error process and often time-consuming. Dose distributions are often optimized based on not only patient-specific anatomy but also physician preferred trade-offs. In this work, we use deep learning to develop a 3D dose distribution prediction algorithm with dose volume histogram (DVH) curves and patient’s anatomy as inputs. Methods: In this work, we build a modified U-Net network to predict a 3D dose distribution. Firstly, all critical structure contour masks, each being represented with one input channel for the network, are encoded into feature maps. Then, each DVH curve is discretized and represented by a vector which is treated as a separate channel for the input. The DVH vectors are concatenated with feature maps from contours along channel axis. Finally, the feature fusion maps containing both contour information and DVH information are decoded with deconvolution to acquire dose distribution. Results: The same patient associated with multiple DVHs which represent physician desired trade-offs was firstly optimized by threshold-driven optimization for reference-based auto-planning (TORA) algorithm interfacing with Varian Eclipse, then this dataset was employed to train and evaluate our model. The difference map between predicted dose and clinical deliverable dose shows an accurate prediction, implying that model can predict the dose distributions with desired DVH curves. We calculate the dose difference for all critical structures as a quantitative evaluation. For the PTV and all OARs, the largest average error in mean dose is about 1.6% of the prescription dose, and the largest average error in maximum dose is about 1.7% of the prescription dose. Conclusions: In this preliminary study, we have developed a 3D U-Net model with patient’s anatomy and DVH curves as inputs to predict the dose distribution with physician desired trade-offs.

ABSTRACT. Introduction: Many patients receiving radiotherapy may benefit from treatment in the upright position due to inability to lay prone/supine and through improved tumour-organ at risk geometry. A major challenge of upright radiotherapy is acquisition of volumetric imaging for treatment planning and image guidance. Our aim was to develop a workflow for upright imaging using a cone-beam CT (CBCT) system on a commercial linear accelerator, and test its suitability for treatment planning.

Materials & Methods: A CIRS adult female phantom was imaged using a TrueBeam® linac (v2.5, Varian Medical Systems) in developer mode. An XML control file rotated each phantom twice through 180° using the treatment couch while acquiring continuous fluoroscopy (125kVp, 30mA, 20ms, 15fps). Extended field of view was achieved by offsetting the kV detector in each of the two half-rotation acquisitions (+13cm,-13cm). The combined projections were reconstructed with the Reconstruction Toolkit (RTK) using an alternating direction method of multipliers with total variation regularisation algorithm. Hounsfield unit (HU) calibration was obtained from upright CBCT images of the electron density phantom. A VMAT plan for a simulated lung tumour on the supine planning CT image was created and copied to the upright (60Gy/30x fractions) to compare the calculated dose.

Results & Discussion: The planning CT was found to have similar structure and intensity to the HU calibrated upright CBCT. Mild ring and projection joining artifacts were visible. The dose distribution was very similar, with some hot spots in bone in the upright plan. Dose metrics show similarity near the tumour regions and vary in lung away from the treatment area.

Conclusions: We have demonstrated that dose calculation on upright CBCT may be feasible, with future work to focus on removing image artifacts and a site-specific HU to relative electron density calibration method to improve dose calculation accuracy.

ABSTRACT. The purpose of this investigation is to calculate the delivered dose in a single fraction of a dose-escalated plan in adaptive radiation therapy for a locally advanced pancreas patient. The assumed breath-hold treatment is aiming for a fraction dose of 9 Gy (10X-FFF) delivered within 15 different 300 degree arcs using volumetric arc therapy optimized on a reference patient model. The length of a single breath-hold is assumed to be 15s at maximum. During the first 2.5s the breath-hold level is verified using kV imaging. If the breath-hold level is within the threshold, the beam is turned on for max. 12.5s to apply the 300 degree arc. For this investigation 15 different breath-hold levels and therefore multiple patient models are used, which were created from various phases of a free-breathing 4D-CT scan. The maximum and the average GTV shift were 12.3 and 5.3 mm respectively. The planning optimization parameters for this single fraction were PTV D95 ≥ 8.55 Gy, Dmax Duodenum ≤ 6.3 Gy, Dmax small bowel, stomach ≤ 5.9 Gy. Individual arcs were recalculated on their corresponding patient model to calculate the delivered dose per breath-hold. Then the dose was accumulated breath-hold by breath-hold on the reference patient model using deformable image registration and dose accumulation to calculate the delivered dose volume histograms. This modeled breath-hold level variation led to a significant decrease in GTV/PTV dose of 2 Gy and to an overdose of the duodenum of 1.3 Gy and stomach of 1.9 Gy. Further motion scenarios, patients and anatomical sites will be studied. The results of this work may help to define patient individual breath-hold levels and may allow to base treatment adaptations on delivered opposed to planned dose. This may help to safely deliver escalated dose protocols on standard linear accelerators.

ABSTRACT. The widespread adoption of electronic medical records (EMRs) is creating rich databases documenting the cancer patient’s care continuum. However, the majority of these data, especially narrative oncologic histories, staging and diagnosis assessments are locked within free text (unstructured) portions of notes. Most of the information entered in the EHR is for routine care delivery. There is a wealth of information in these clinical notes for big data application but the challenge is to capture these data in a discrete format. Keeping this goal in mind, we performed a study in which tumor staging, initial prostate-specific antigen (PSA) and primary and secondary Gleason scores were extracted from radiotherapy initial consultation notes for prostate cancer patients. For this study, radiotherapy initial consultation notes of 800 patients were extracted. Using rule-based pattern matching, these data elements were automatically extracted from the notes. As part of training the model, the dataset was randomly split into 80/20 as training and testing respectively. F-measure was used to measure the accuracy of the extraction model; with 0.74 for primary and secondary Gleason score, 0.72 for total Gleason score, 0.36 for PSA, 0.67 for NCCN risk, 0.80 for tumor stage, 0.74 for the nodal stage, and 0.78 for the metastatic stage. The extraction of initial PSA has the lowest F-measure because PSA scores over multiple years were recorded in these notes. Such temporal attributes are difficult to identify with rule-based extraction methods. Based on these analyses, we infer that (i) rule-based information extraction (IE) works best with data elements that follow a standardized format, such as NCCN risk classification and Gleason scores; (ii) data elements with high variance in a representation such as PSA, or verbose description require ontology-based entity extraction. Extraction methods combining multiple NLP strategies are needed to yield overall effective results.

ABSTRACT. In recent years, machine learning (ML) approach in radiation oncology is a growing interest. These ML-based applications aim to provide automated assistances to radiation oncology workers to complete the complex workflow of radiation therapy. In order to build an automated ML application for clinical use, automated method of extracting data for model construction and updating is required. However due to the non-standardized schema in different oncology information systems (OIS), implementation of these ML-based applications to different clinics is limited as it is difficult to extract relevant clinical data for ML purpose consistently.

In this talk we present our work on automatically extracting clinical data from two main OIS, Aria (Vairan Medical System, Palo Alto, USA) and Mosaiq (Elekta AB, Stockholm, Sweden), for creating an error detection Bayesian network (BN) model using SQL queries generated via a schema mapping between the OIS databases and a dependency layered ontology for radiation oncology (DLORO). The aim of this work is to demonstrate the advantage of formalizing terminology into ontology on automated data extraction from different OIS databases for ML model constructions and updates, which is the foundation of spreading ML-based applications in the community.