ABSTRACT. Our study aims to enhance the assessment of lower limb misalignment through an innovative approach that combines coordinate regression and region of interest (ROI) segmentation for landmark detection on full lower limb X-rays. Traditionally, surgeons manually measure limb deformities by assessing angles between axes drawn on X-rays connecting specific landmarks. While various deep learning solutions exist, our method introduces a segmentation-guided regression network. Comparing our approach with landmark regression and ROI segmentation, we evaluated detection errors for nine landmarks and measured misalignment metrics. While ROI segmentation achieved accurate landmark identification, it faced challenges in malalignment measurement due to undetected landmarks. On the other hand, regression had no failed detections but exhibited lower accuracy. Our segmentation-guided regression strikes a balance, ensuring no missed landmarks, improved landmark accuracy, and precise misalignment quantification.

By encouraging the regression network to focus on specific areas through segmentation guidance, our method positions landmarks more accurately and effectively measures malalignment. Despite ROI segmentation's superior performance in landmark identification, it is more sensitive to image artifacts. Our approach provides surgeons with a reliable tool for comprehensive lower limb malalignment assessment, combining the strengths of both coordinate regression and ROI segmentation.

ABSTRACT. Navigated orthopaedic surgery relies on bony landmarks and the accuracy of their acquisition can impact the surgery outcomes. We propose an automatic workflow to determine 11 femoral bony landmarks on virtual 3D meshes. The studied landmarks were first determined on the mean shape of a statistical shape model of the femur. Then the statistical shape model was fitted to the virtual 3D meshes, and the landmarks of the mean shape were projected onto the fitted mesh. The proposed method was validated by comparing the computed landmarks to ground truth landmarks acquired manually on 41 knees. We also investigated the impact of the landmarks’ accuracy on the variability of axes and resection planes derived from the considered landmarks. The 11 femoral bony landmarks were automatically determined in less than 2 minutes with an accuracy of 2.81 ± 1.86mm. Such error impacted the accuracy of the derived axes and planes with less than 0.5° angular deviation. Three landmarks had poorer accuracy and precision attesting how ambiguous their definition is and the difficulty to identify them. The proposed method allows the fast acquisition of femoral bony landmarks, with similar to better accuracy than manual approaches.

ABSTRACT. 3D planning of a corrective osteotomy of a radius malunion requires a healthy reference which is not always available. A shape completion model could create a reference bone based on the unaffected part of the malunited bone. The aim of this study is to develop and validate a shape completion model for clinical use. A statistical shape model (SSM) was developed based on CT scans of 80 healthy radii. This SSM was expanded into a shape completion model to predict the distal 12% of the radius, based on the proximal 88%. Nine other CT scans were used to validate this model and to set hyper parameters that dictate factors of the available data. Finally, eight more CT scans were used to test the performance of the shape completion model. The average accuracy of the shape completion model, measured through a root mean square difference, was 0.42mm (SD 0.10) and the average Hausdorff distance was 3.57mm (SD 1.09). The predicted radii were comparable to the actual radii, with the differences mainly found at the radial styloid, Lister tubercle and the sigmoid notch. The radiocarpal articular surface was often well predicted. In conclusion, the current shape completion model showed to predict clinically useable models that have error margins that are comparable with previous models described in literature. Future research should compare the use of the shape completion model to the use of the contralateral radius to find what technique works more optimal.

ABSTRACT. The purpose of this study was to utilize a large database of lateral functional radiographs to develop a deep learning (DL) pipeline to rapidly classify functional position and measure pelvic tilt (PT), sacral slope (SS) and lumbar lordotic angle (LLA) for patients undergoing total hip arthroplasty (THA). A DL pipeline was developed for identifying functional positions and accurately detecting key anatomic landmarks. The pipeline was trained on imaging from an international joint registry, comprising 52,770 images for classification, 9,875 for object detection, and 22,578 for landmark detection. A 70:20:10 data split was implemented across training, validation, and testing and enhanced through data augmentation. The pipeline's performance was evaluated through area-under-the-curve (AUC), F1-score and mean absolute error (MAE) analyses. The DL pipeline's angular accuracy was compared in a prospective series of three expert engineers’ manual annotations and validated by two senior engineers and two surgeons. The DL pipeline processed images in 1.96±0.04 seconds per image, achieving precision, recall, Receiver-operator-characteristic-AUC, and Precision-Recall-AUC metrics above 0.994. It accurately identified anatomical landmarks with errors of: PT: 1.6°±2.1°, SS: 3.3°±2.6°, LLA: 4.2°±3.2°. Prospective landmarking by expert engineers showed no difference in PT and LLA compared to the DL pipeline, and a 0.5° difference in SS (p=0.043). Clinical validation revealed no significant difference in landmark rejection rates between the DL pipeline and expert engineers (p>0.05). Using deep learning models, we developed and clinically validated a fully automated image processing pipeline that can rapidly and accurately measure patient specific spinopelvic mobility from lateral functional radiographs.

ABSTRACT. Deep learning techniques for diagnosing Developmental Dysplasia of the Hip (DDH) in newborns from ultrasound (US) images of the hip have demonstrated improved reliability over manual annotations of US scans. While volumetric 3D US has been shown to better represent hip bone morphology, most of the proposed automatic diagnostic approaches to measure 3D equivalents of the commonly used 2D Graf angle rely on strong morphological (geometric) priors. We have found that a significant fraction of cases (~20%) result in metrics which expert assessors regard as incorrect or implausible. We hypothesize that the lack of robustness of existing algorithms is due to their assumption that selected morphological priors are always valid, and this may not hold in a number of cases. In this study, we evaluate the differences between extracted DDH metrics based on expert labels and automatic segmentations. We show that a metric extraction process that uses morphological priors is sensitive to relatively small variations in the segmentation results.

ABSTRACT. Abstract Introduction: In orthopaedic surgery, diagnosing musculoskeletal disorders is challenging due to limitations in conventional methods such as costly medical imaging and complex optical markers. To overcome this, we propose an innovative algorithm that infers bone shape exclusively from skin data, focusing on the femur and thigh. Our primary goal is to predict the femur based on the thigh, with a subsequent focus on deriving clinically relevant anatomical features, particularly the femur's mechanical axis.

Materials and Methods: Using a dataset of 25 angioscanner thigh-femur pairs, we construct a robust Statistical Shape Model (SSM) capturing correlations between skin and underlying bone, including femur and soft tissues' depth. The variable "Depth" represents the distance between femur points and their corresponding projections on the thigh. Femur prediction from a known thigh employs the SSM and the Markov Chain Monte Carlo (MCMC) method, generating samples approximating the target thigh. The selected thigh aligning most closely with the target provides the final prediction.

Results: The SSM exhibited high specificity and generality (0.8mm/2mm and 0.9mm/2.3mm for femur/thigh). Femur inference yielded a Root Mean Square Error (RMSE) between 7-14 mm, outperforming existing methods and closely matching with results using real landmarks on the thigh (RMSE [7-11] mm).

Conclusion: In summary, our algorithm accurately infers femur features from thigh skin, showing promise for non-invasive musculoskeletal parameter estimation in orthopaedic surgical applications. Future improvements may involve addressing gender distinctions and incorporating additional parameters such as BMI.

ABSTRACT. This study assesses the accuracy of a deep learning (DL) algorithm for automating orthopaedic parameter calculations on anteroposterior (AP) radiographs. A U-Net DL segmentation algorithm was developed using 297 calibrated AP radiographs, manually segmented by an engineer. The masks were used to define landmark positions on the AP radiographs and subsequent computation of orthopaedic measurements. During model development, images are divided into training (70%), validation (20%), and testing (10%) sets, and resized to 512x512 pixels. Leg length discrepancy (LLD) was assessed using three reference lines relative to the lesser trochanter: 1) obturator foramen, 2) teardrop, and 3) ischia. The femoral axis was defined using proximal femoral diaphysis and distal femoral canal landmarks. Algorithm accuracy was evaluated using the intersection over union (IOU) of the segmented regions and mean absolute error (MAE) between predicted and ground truth values. Error in offset, LLD, and caput-collum-diaphyseal (CCD) angle was compared. The model produced an average IOU value of 0.85 for all masks. The MAE varied across the landmarks with the lowest error observed in the identification of head centers (0.9±0.6mm) and the highest error in the proximal femoral diaphysis (3.1±2.8mm). Accuracy was highest for acetabular offset (0.6±0.5mm) and femoral offset (1.5±1.2mm). When assessing LLD with reference lines to the lesser trochanter, the least error was noted when referencing the obturator foramen (1.6±1.7mm). CCD angle calculation error was 1.7±1.6°. The DL algorithm shows promise as an efficient and accurate tool for defining landmarks in AP radiographs and automating the calculation of clinically relevant parameters.

ABSTRACT. Automated virtual reconstruction of complex fractures could be beneficial for preoperative surgical planning. We created a statistical shape model (SSM) based on data from 200 intact 3D hemipelves. This model can quantify shape variations and can reconstruct a new, abnormal shape. We used our SSM to reconstruct both simple and complex acetabular fractures, then evaluated the model’s reconstruction accuracy by comparing the reconstructed shape with the intact opposite hemipelvis. In this retrospective diagnostic imaging study, we used our statistical shape model (SSM) to virtually reconstruct the fractured hemipelves of eighty-three patients with an acetabular fracture. We calculated the root mean square error (RMSE) between the reconstructed shape and the intact contralateral shape for the entire hemipelvis and for regions crucial for plate fitting. These key regions for plate fitting were defined as: (1) the length and radius of the iliopectineal line, (2) the length and radius of the ischial body line, (3) the diameter of the acetabulum, (4) the slope of the quadrilateral space, and (5) the weight-bearing acetabular dome. The median RMSE for the entire hemipelvis in cases of elementary fractures was 2.2 mm (interquartile range (IQR): 1.7 – 2.5 mm), compared to 3.2 mm (IQR: 2.2 – 3.9 mm) for associate fractures. The median RMSE for the plate-fitting regions in elementary fractures was 1.7 mm (IQR: 1.4 – 2.1 mm), whereas for associate fractures it was 2.7 mm (IQR: 2.0 – 4.1 mm). A statistical shape model enables precise virtual reconstructions of elementary acetabular fractures within a clinically acceptable range, particularly in areas critical for plate-fitting. SSM-based reconstructions can be a valuable asset for preoperative planning in clinical practice, especially when the contralateral hemipelvis template is not available.

ABSTRACT. Sacroiliac (SI) joint dysfunction leading to debilitating pain can be effectively treated with minimally invasive sacroiliac joint fusion (SIJF). This treatment is commonly performed using 2D fluoroscopic guidance. This makes placing the implants without damaging surrounding neural structures challenging. Virtual surgical planning (VSP) using simulated fluoroscopic images may improve intraoperative guidance. We describe a VSP workflow for SIJF using simulated fluoroscopic images and evaluate it for achieved implant placement accuracy. Ten SIJFs were performed on 10 patients by the same surgeon, giving a total of 30 planned and placed implants; the median age was 39 years, and all patients were female. The overall mean implant placement accuracy was 4.9 ± 1.3 mm and 4.0 ± 1.4°. Malpositioning complications were absent. VSP assisted the surgeon in mentally reconstructing the complex 3D anatomy leading to a clinically acceptable accuracy in positioning and length selection of the implants.

ABSTRACT. We aimed to investigate the morphological and osteotomy design factors to minimize the pubic osteotomy gap and to avoid lateralization and superiorization of the femoral head in rotational acetabular osteotomy (RAO). Fifty hips that underwent RAO and had pre- and post-operative CT scans were analyzed. Postoperative pubic osteotomy gap (mm) in the coronal plane was measured. The lateralization and superiorization of the femoral head center (mm) was determined. Iliac osteotomy angle on the coronal plane and Ischial osteotomy angle on the axial plane were defined according to previous report. Univariate and multivariate analysis were performed using the pubic osteotomy gap, lateralization and superiorization of the femoral head as outcome variables. Mean Lateral CE angle (LCEA) correction angle was 24º, iliac osteotomy angle was 28º, ischial osteotomy angle was 46º, pubic osteotomy gap was 12 mm, and lateralization and superiorization of the femoral head was 0 mm and 1 mm, respectively. The larger pubic osteotomy gap was associated with a smaller ischium osteotomy angle (β=-0.40) and a larger LCEA correction angle (β=0.38) (R2=0.52). The larger lateralization was associated with a larger iliac osteotomy angle (β=0.36) and a smaller LCEA correction angle (β=-0.60) (R2=0.59). The larger superiorization was associated with a larger iliac osteotomy angle (β=0.39) and a larger preoperative LCEA (β=0.37) (R2=0.38). In conclusion, to improve head coverage while minimizing the pubic osteotomy gap, it was suggested to be appropriate for RAO osteotomy designs to be the iliac osteotomy through the inner plate cranially and ischial osteotomy through quadrilateral surface dorsally.

ABSTRACT. Definition of target zones for prosthesis alignment is a common method in preoperative planning for THA. Several criteria for calculation of these target zones are discussed in literature and a combined target zone (CTZ) has been defined on the basis of morphological and morphofunctional parameters. Especially the definition of a patient-specific target ROM is important for prediction of postoperative prosthesis´ functionality. Prosthetic impingement can limit movements und thus restrict patients in their activities of daily living. A precise analysis of the postoperative target ROM can help to provide prosthetic impingement during movements, that are essential for the individual patient and also improve the size of the resulting CTZ for prosthesis alignment. We analyzed 4 different ROM definitions and their impact on the size of the ROM based target zone (ROMTZ) as part of the CTZ on a database of 200 patients. We found significant differences between the sizes of the resulting ROMTZs. Our study underlines the effectiveness of a differentiated evaluation of the individual target ROM for optimized preoperative planning in THA.

ABSTRACT. 3D imaging technologies allow for more accurate planning of multiplanar corrective osteotomies around the knee. However, 3D planning introduces complexity. For example, varus/valgus is expressed in the coronal plane of the leg, while posterior tibial slope is expressed in the sagittal plane of the tibia. We present a software platform that deals with this complexity.

A high tibial osteotomy was parametrized using 5 parameters: hinge axis rotation in axial plane; hinge axis tilt in axial plane; hinge axis position along longitudinal axis; hinge axis distance to cortex; osteotomy opening angle. A MATLAB-algorithm was developed in which the effect of each possible combination of osteotomy parameters was obtained on change in alignment parameters expressed in coordinate systems of the tibia and leg. By finding the set of osteotomy parameters that approaches the predefined change in alignment parameters the closest, a 3D osteotomy planning can be linked to a desired outcome.

A digital interactive osteotomy planning platform was developed that automatically computes the required hinge axis orientation and osteotomy opening/closing angle for a given predefined outcome of change in posterior tibial slope and %-weight-bearing-line in the coordinate systems of the tibia and leg, respectively, with an error of <0.1°. During planning, the user is able to manually tweak the osteotomy and is given real-time feedback on the resulting trade-offs.

The presented osteotomy planning platform allows for automatic and complex 3D planning of multiplanar leg malalignment correction, while integrating that alignment parameters are measured in different coordinate systems.

ABSTRACT. Complex bone deformation of the femur can occur in bone disease. This can cause significant functional impairment and is often an indication for surgery. Current technology enables preoperative planning and execution of these complex multi-planar osteotomies with high accuracy. A novel solution is presented that generates preoperative plans using Blender. The model automatically optimizes femoral shaft shape within the boundaries of clinical constraints. 20 cases of femoral deformity were processed by the model retrospectively. Eighteen of twenty preoperative plans were accepted by two independent clinical experts. This shows the ability of the model to generate preoperative plans of sufficient clinical quality. Four random cases were compared to manual preoperative planning. These showed comparable results. The collum anteversion angle was normalized in four cases in the automatic group and only in one for the manual group. The manual planning process was slower in every case, ranging from 31% to 575% slower. The proposed method, allows fast interactive surgical planning within a constraint based design and is a promising aide for complex femoral correction planning.

ABSTRACT. Morpho functional interbone parameters of the knee are often used in clinical practice to assess the functional anatomy of the individual patient. Respective parameters, such as the TT-TG distance, are regularly measured manually on CT or MRI images. To overcome this time-consuming process, an automated framework for knee morphological analysis was previously presented by our group. A relevant remaining limitation, however, is the imaging performed in supine and not in an active, weight bearing position, which was addressed in this study. Data from 7 patients (14 knees) scheduled for total hip arthroplasty were used for this study. After segmentation, the CT-derived bone surface models were matched to the EOS images based on landmarks, with manual control and optional correction. Subsequently, the automated framework for morphological analysis was applied. A statistically significant difference in mean was found for the joint rotation. The other mean deviations were not statistically significant and below the parameters’ standard deviation, whereas the absolute deviations were higher. This fact highlights the relevance of inter-individual differences in supine versus standing interbone parameter measurements. In general, large discrepancies regarding the changes in interbone parameter measurements from supine to standing position were found in the literature. The results of this study are plausible with regard to theoretical biomechanical relations. Overall, the study motivates an interbone parameter assessment in a standardized, active weight bearing position. With further automation in data pre-processing, the workflow could be applied to large databases and hence be used to define reference ranges of interbone parameters in various poses.

ABSTRACT. Musculoskeletal (MSK) modelling and predictive simulation software, such as OpenSim and OpenSim Moco, are powerful tools that can evaluate the causal relationships between changes to the musculoskeletal system and its function. Statistical Shape Models (SSM) are a computational anatomy method that can efficiently describe variation in geometric datasets such as a clinical population’s bony anatomy. Thus, it can be a powerful tool to systematically vary the geometry of musculoskeletal models to approximate population variance. Therefore, the goal of this work was to develop a computational framework that can be used to predict the effect of patient morphology and implant placement on post-op biomechanics using SSMs, MSk modeling, and predictive simulations. The computational framework is composed of four phases: i) Scapula geometry generation with SSM mode of variation changes; ii) Virtual surgery parametrically adjusting RTSA characteristics based on orthopaedic surgeon guidelines; iii) Definition of the musculoskeletal model for each scapular geometry, muscle attachment and joint locations; iv) Each model implemented in predictive simulations of an activity of daily living. Twenty-five morphologies were generated by varying six modes to four levels (±1 and ±3 SD) plus the average morphology. Each MSK model was used for a muscle driven predictive simulation with an objective function with goals of: a) minimise simulation time bounded to [0.75 2.5] seconds, b) minimise effort (sum of controls squared) and c) to minimise the 3D spatial error between markers on the model’s wrist and global target markers. By generating the scapular morphologies used in our study’s MSK model from a validated SSM we were able to link specific morphological changes in the MSK model (e.g. moment arms) with the resulting biomechanics during simulated activities of daily tasks. The use of clinical SSMs is an important advancement in predictive surgery workflows as computational studies can strategically and systematically sample from a clinical population to identify morphologies that could be at high risk of complications. Future studies should build on these results to see the combined effect of morphology, construct design, placement and muscle function on post-operative biomechanical results.

ABSTRACT. Patient-specific computational models of the shoulder have the potential to further our understanding of joint biomechanics and optimize treatments for common pathologies such as osteoarthritis and rotator cuff tears. Since active motion and stability of the shoulder are mainly governed by the surrounding soft tissues, such models must be able to reliable and accurately predict muscle paths. Our aim was to develop and validate a computationally efficient line-segment muscle mapping model capable of mapping patient-specific muscle paths of the multi-pennate muscles of the deltoid and rotator cuff.

A triangular surface mesh was generated from segmentations of an anonymous male subject (75 years old, BMI 23) and muscle origin/insertion points were identified based upon previously reported data. A ‘convex hull’ algorithm identified optimal muscle fibre paths during 0-90° coronal plane abduction, sagittal plane flexion and axial rotation at neutral elevation. The model was capable of computing muscle lengths, moment arms and line of action for each muscle fibre.

The model had acceptable correlation with in-vivo and cadaveric data from the literature. The model was also highly efficient, capable of mapping 42 muscle segments at 2.5° intervals of joint motion in less than 17 seconds.

The current model presents a patient specific method for modelling muscle multipennate muscles of the shoulder with high computational efficiency, requiring only the surface mesh inputs of the bony anatomy and muscle origin/insertion points without the need for commercial finite element contact detection software. The model may be used further for the study of shoulder musculoskeletal disorders.

ABSTRACT. Computer Assisted Surgical (CAS) systems have been used successfully in joint arthroplasty to improve the accuracy of resections. Total ankle arthroplasty (TAA) is a surgical treatment for end-stage ankle osteoarthritis, and the latest generation of TAA is associated with favorable clinical outcomes as a modern alternative to ankle arthrodesis. Alignment of the implants during TAA can be challenging because of limited surgical exposure and reliance on fluoroscopic visualization. Therefore, a TAA application for CAS system was developed using CT-based alignment alongside required fluoroscopy with the intent of facilitating the procedure and improving accuracy of bone resections. TAA was performed by a board-certified, fellowship-trained orthopedic surgeon on twelve artificial ankle joint specimens using a CAS system featuring a dedicated ankle application. Bone resections were individually virtually planned by the surgeon performing the operation using template software to choose appropriate implant position and size relative to the bony anatomy. The resected bones were scanned with structured light before and after the procedure and the resultant models overlaid for assessment of the error relative to the original plan. For all eight angular and positional cut parameters across both the tibia and the talus, the mean signed overall intraobserver error was less than 2mm and 2° relative to the plan, and the 95% confidence interval was less than 2mm and 2°. The results point to resections with errors of less than 2mm and 2° and therefore the system can offer both accurate and precise intraoperative surgical resection measurements during computer-assisted TAA.

ABSTRACT. Continuum dexterous manipulators (CDMs) have shown great potential when integrated with computer assisted orthopaedic surgery (CAOS) systems for minimally invasive surgery (MIS). We hypothesize that the enhanced dexterity of CDMs may allow for greater access to target tissue through a single port when compared to traditional, rigid MIS instruments. To assess such CDMs for intervertebral disc removal applications, a phantom study in the scope of MIS transforaminal lumbar interbody fusion (TLIF) was conducted to evaluate the achievable surgical workspace of the intervertebral disc (IVD) during disc space preparation. A CDM with 6 mm diameter and a conformable nitinol whisk tip was evaluated against three 135° lumbar curettes in a 2D L4-L5 IVD phantom by an experienced spine surgeon. Improvements of up to 41.7% in reachable IVD workspace are achieved with the CDM, demonstrating its viability in improving outcomes for MIS spinal fusion.

ABSTRACT. Introduction: The patient’s native kinematics is an important factor to consider when striving for restoration of function after Total Knee Arthroplasty (TKA). The goal of this study was to define phenotypes based on intra-operative kinematics. Methods: Data from 2967 navigated TKA knees (OrthoPilot®, Aesculap AG) were included. A neutral knee flexion was performed by the surgeon. The antero-posterior (AP) position of the knee center and of the medial and lateral condyles were extracted at 0° and 60° of flexion. Categories were defined according to the amount and direction of the AP shift at 60°. Results: 49% of the knees had a shift smaller than 4mm, 16% (respectively 17%) had a 4mm to 8mm shift anteriorly (respectively posteriorly). 19% of the knees had a shift beyond 8mm anteriorly or posteriorly. 56% showed a medial pivoting pattern, with 41% having a pure lateral posterior shift. Only 6% showed a lateral pivot pattern with an anterior shift of the medial condyle. 27% of knees showed no pivoting behavior, having either limited rollback on both sides (11% of knees) or a parallel shift of more than 4mm on both the medial and lateral condyles. Discussion: Clear kinematic phenotypes could be defined from a neutral passive knee flexion. A standardized tool will support intraoperative decision-making, such as implant selection, which is an important step towards an optimal personalized treatment for each TKA patient.

ABSTRACT. For computer-assisted percutaneous scaphoid fixation, a patient-specific bone model is required for surgical planning. This bone model is commonly derived from pre-operative computed tomography (CT) or magnetic resonance imaging (MRI) data. We propose an approach for bone model derivation based on intra-operative 3D ultrasound (US) imaging for the cases, where pre-operative diagnostic CT or MRI are not indicated or available. As scaphoid bone surfaces are only partially visible in sonographic images, we employ the Transformer-based AdaPoinTr architecture to incorporate statistical morphological knowledge for the completion of partial bone surfaces extracted from sonographic images. For the generation of datasets, we built a statistical shape model (SSM) based on 85 scaphoid bone models. From this SSM, we generated 12288 full scaphoid models for training. 20 additional scaphoid bone models were used for testing. Partial models for both training and testing were generated by subsampling the full models, mimicking 3D US imaging from a volar probe position. Evaluation of the final trained model on the test subset showed a mean symmetric distance of 0.3 mm between original and completed scaphoid models, with an inference time of 0.2 s per model. We furthermore planned screws based on the completed test models and evaluated their fit for the original models. We found no screw protrusion for any tested model, with a mean safety margin to bone surface of 0.7 mm. This study shows feasibility of our approach for US-based bone model generation; future work may aim at integrating dorsal surface information.

ABSTRACT. Artificial intelligence (AI) and machine learning (ML) take an ever-growing place in medical care. Anatomical segmentation and reconstruction is one of the fields where ML reveals to be very efficient. Yet, verification of ML results still requires human verification and correction especially on pathologic morphologies. In this study, we introduce and evaluate the accuracy of a deep learning-based model, designed to assess automated ML-reconstructions of the scapula. The model provides an index indicating how much correction work the predicted reconstruction needs to guarantee its accuracy. It also flags potential correction areas for human intervention. The index can be used to separate predictions requiring little to no revision from predictions where corrected voxels represent more than 1% of the scapula, with an accuracy of 80% and a sensitivity of 96%. Combined with a formerly developed automated scapula segmentation model, this model is anticipated to reduce the need for exhaustive manual oversight, moving closer to a fully-automated system.

ABSTRACT. Objectives: Large language model (LLM)-based chatbots increase access to health information for those searching their symptoms online. The open-source information from the internet utilized by LLM includes content that may be inaccurate. Understanding potential misinformation in LLM responses can alert users to the important limitations of this technology and help inform strategies for development and safer utilization of LLMs. We asked what misinformation is common when providing a LLM-based chatbot with symptom descriptions. Methods: We provided a LLM-based chatbot (ChatGPT 3.5) with symptoms related to a range of common upper musculoskeletal disorders using 4 different descriptions; twice specifically worded and twice non-specifically. Potential misinformation was coded using a checklist, based on misinterpretations in 3 main categories; reinforcement of unhelpful thinking, misleading statement, and inaccurate/misdiagnosis. We performed analyses of the percentage of sentences containing potential misinformation for queries structured with specific and non-specific symptoms. Results: There was no difference in the percentage of sentences containing statements with potential misinformation when presenting ChatGPT with specific symptoms (82%) compared to non-specific symptoms (80%). For both specific and non-specific symptoms, the most frequent category of potential misinformation was "reinforcement of unhelpful thinking", with the most common subtheme being “reinforcement of kinesiophobia.” Conclusion: LLM based chatbots present a notable amount of potential misinformation when responding to simulated patients inquiring about their symptoms, regardless of the style and wording of the inquiry. This finding signals future model developments of LLMs to build in safeguards that mitigates the overemphasis on unhelpful thinking.

ABSTRACT. Introduction: Distal radius fractures (DRFs) are very common in adults. It is challenging to accurately predict the risk of secondary displacement of DRFs after successful closed reduction. One could argue that fracture stability is currently “guesstimated”, and is limited by surgeons’ bias. A reliable and interpretable deep learning model to predict loss of threshold alignment in DRFs, with a higher accuracy than surgeons, is of great interest to reduce undesired treatment variation. Subsequently, we aim to deploy AI to answer the clinically relevant question: Can we develop a Convolutional Neural Network (CNN) that predicts fracture instability on radiographs of DRFs? Methods: 2365 radiographs of 733 patients with conservatively treated DRFs were collected in two Trauma Centers. All radiographs (trauma, post-reduction, follow-up) were assessed for acceptable alignment defined by the Dutch DRF Guidelines. This resulted in two categories: 424 stable and 309 unstable fractures. To augment CNN training, a method called landmark annotation was used. All available trauma- and post-reduction radiographs were annotated by two researchers. Results: Our model achieved 76% accuracy, 82% AUC, 84% sensitivity and 68% specificity in predicting future DRF loss of threshold alignment when combining PA- and LAT radiographs. Discussion: We developed a CNN to predict distal radius fracture instability with an accuracy of 76%. This CNN has great potential to empower both surgeons and patients with personalized risk stratification for fracture instability in the treatment of DRFs.

ABSTRACT. This study aimed to investigate the muscle volume and computed tomography (CT) values of knee osteoarthritis (KOA) patients measured by the artificial intelligence-based method and to evaluate its relationship with health-related quality of life (HRQoL). 43 knees undergoing total knee arthroplasty (TKA) were evaluated in this study. 9 men and 34 women were included in the study. Computed tomography (CT) was performed preoperatively. Ten muscle groups on the affected side were subsequently segmented, and each muscle volume and the mean CT values (Hounsfield units [HU]) as a muscle fatty degeneration evaluation were calculated using the artificial intelligence-based method. HRQoL was evaluated using the Knee society score (KSS), preoperatively. The correlations between the KSS and muscle volume or CT values were statistically assessed. Spearman’s rank correlation test and multiple regression analysis were performed, and statistical significance was set at P < 0.05. Multiple regression analysis revealed that Functional Activities were significantly associated with muscle fatty degeneration of the gluteus medius and minimus muscles, the anterior compartment of the leg, and the lateral compartment of the leg. (β=0.42, p=0.01; β=0.33, p=0.038; β=0.37, p=0.014, respectively) We found that fatty degeneration of the gluteus medius and minimus muscles, the anterior compartment, and the lateral compartment of the leg was significantly related to Functional Activities in the patients with end-stage KOA. Based on these findings, we suggest these muscles should be targeted during rehabilitation before TKA.

ABSTRACT. In this study we develop a neural network to predict post-operative total knee arthroplasty (TKA) pain. The neural network takes patient demographics, pre-operative PROMS, joint anatomy and soft tissue profile as inputs, and predicts a Knee Injury and Osteoarthritis Outcome Score-12 (KOOS12) pain value. Data was extracted from an international joint replacement registry of patients who received a tibia first robotically assisted TKA with a digital joint tensioning device who completed pre-operative and post-operative questionnaires. A total of 914 patient outcomes from 368 individual patients were extracted from the registry. A fully connected neutral network with 3 hidden layers was generated in AzureML using Keras. Optimization of neurons per hidden layer was performed using a sweep analysis in which the neurons per layer was varied between 10 – 2000. The mean squared error (MSE) and mean absolute error (MAE) were recorded. Sweep mode analysis returned a model with a normalized MAE of 0.61 with hidden layers of 2000, 500 and 300 neurons in the first, second and third hidden layer respectively. Using this model, a KOOS12 MAE of 8.1 points, and correlation coefficient of 0.756 was recorded using test data. Here we successfully developed an exploratory neural network to predict post-operative outcomes in TKA based on pre-operative patient information and intra-operative data. Using a test data set, we validated the model and found a mean absolute error well below the minimum clinically important difference for KOOS12 pain.