Introduction
Fractures in the forearm are common and sometimes result in limitations of pronation/supination. Besides malunion as a possible cause, soft tissue involvement may play a more significant role as well. More insight in both causes of impaired forearm rotation could help to treat patients in the least invasive way as possible, potentially avoiding invasive corrective osteotomies in some patients.

Objectives
This study aimed to provide a deep-learning based framework for automated segmentation of anatomical structures involved in pronation/supination of the forearm on magnetic resonance (MR) images. This approach allows for visualization and quantitative analysis of the patient-specific anatomy, enabling efficient identification of soft tissue structures that may contribute to impaired forearm rotation.

Methods and materials
Manual ground truth annotations of six anatomical structures (radius, ulna, interosseous membrane, m. pronator quadratus, m. pronator teres, m. supinator) were performed on 24 fast-recovery fast spin-echo T2-weighted (FRFSE T2) in-phase Dixon images of the forearm. The dataset contained an equal distribution between affected and unaffected, and left and right forearms. Two nnU-Net configurations (2D and 3D) were trained on 20 manually segmented forearms using 5-fold cross-validation for segmentation of the six structures. An ensemble was created by combining predictions from both fully-trained models. A hold-out test set of 4 forearms was used to evaluate segmentation performance using the Dice similarity coefficient (DSC) and the average symmetric surface distance (ASSD) metrics. Additionally, relative volume difference (Δrel) between ground truth and predicted segmentations were computed to assess under- or oversegmentation.

Results
The 3D model achieved the best segmentation performance, with a median DSC score of 0.894 (IQR=0.094) and a median ASSD of 0.324 (IQR=0.386) mm. It slightly undersegmented the anatomy, with a median relative volume difference of -2.7% (IQR=7.1%). Qualitative results revealed that the 3D model produced segmentation masks that contained fewer and less severe segmentation errors compared to the 2D model and ensemble. Minor segmentation errors were observed in the interosseous membrane, the proximal part of the m. pronator quadratus and the insertion of the m. pronator teres in some cases.

Conclusion
The 3D nnU-Net model has proven its suitability for clinical use, enabling fast, reproducible and precise segmentation of structures involved in pronation/supination of the forearm. This approach facilitates bilateral comparisons of soft tissue structures through visual assessment and quantitative analysis, supporting patient-specific and minimally invasive decision-making.

Muscle-tendon paths vary between individuals, and musculoskeletal models show sensitivity to these variations when estimating muscle and joint forces. Accurate estimations of these internal forces allow for a more comprehensive approach to researching debilitating musculoskeletal pathologies such as osteoarthritis. However, defining subject-specific muscle-tendon paths is labour-intensive and requires expert knowledge, which is not as repeatable. Therefore, in this study, the accuracy of determining subject-specific muscle points and volumes based on lower limb Magnetic Resonance (MR) scans with the nnU-net is evaluated. Two models are trained using the open-source pipeline, referred to as the point and volume model. The volume model aims to segment muscle-tendon volumes and is trained on the open-source augmented dataset of Henson et al. (2023). The point model localises attachment and via points describing muscle action lines for a subset of relevant muscles of the volume model. Since U-net is not designed for predicting points, a workaround is introduced by creating cubes around the points as labels. The 3D volume model scored a median Dice Similarity Coecient of 92.7 % and shows some generalisation capability. The 3D point model scored a median Euclidean error of 5.1 mm. Compared to intra-operator variability for attachment points, this approach yields lower and more repeatable accuracy without required manual intervention. The nnU-net pipeline is capable of producing accurate models that can define subject-specific muscle-tendon paths based on MR scans.

Osteoarthritis (OA) is a chronic musculoskeletal joint disease that leads to disability. Osteophytes are a hallmark of OA in the knee, characterized by the formation of bone spurs that contribute to joint pain and reduced mobility. This study explores the application of deep learning (DL) techniques for the automatic detection and grading of osteophytes on magnetic resonance (MR) images of the knee. Leveraging the DenseNet-121 and ResNet-50 DL architectures from the Medical Open Network for Artificial Intelligence (MONAI) framework and a dataset, containing 1782 double echo steady-state (DESS) MR images from the Osteoarthritis Initiative (OAI), the study aims to enhance diagnostic accuracy and efficiency in medical imaging analysis. The dataset was split 8:2 for training and validation purposes, respectively. Through a series of numerical experiments, the research evaluates binary classification, region of interest (ROI)-based detection, and multi-class classification models, demonstrating that DenseNet-121 generally outperforms ResNet-50. The five-fold cross-validated binary DenseNet-121 model achieved an area under the receiver operating characteristic curve (ROC AUC) score of 0.90 and a balanced accuracy of 0.82, with a 95% confidence interval (CI) of 0.81-0.83 trained on resampled whole knee images. Moreover, the cross-validated ROI detection models for the patella inferior, superior, and tibia lateral subregions achieved balanced accuracy scores of 0.89 (0.88-0.90 CI), 0.86 (0.85-0.87 CI), and 0.85 (0.84-0.86 CI), respectively. However, the multi-class DenseNet-121 model achieved lower performance, with a balanced accuracy of 0.73 (0.71-0.75 CI), indicating the complexity of multi-class classification in this context. Furthermore, this research did not include hyperparameter optimization, as many settings were kept at their default values, suggesting the possibility for improved results. The cross-validated models were evaluated on an external test set, obtained from the Erasmus Medical Centre, comprising FSPGR-FS images from a significantly younger patient cohort, with a notable class imbalance. The models’ performance on this dataset was significantly lower than their validation results, underscoring the limitations in generalizing to different age demographics and class distributions. External testing underscores the need for more robust models to maintain high performance across diverse datasets and clinical settings. Key contributions of this study include the use of weighted categorical cross-entropy (WCCE) loss functions and analysis of the knee’s subregions to improve detection accuracy. The findings establish a solid foundation for further research, suggesting future work should focus on advanced optimization techniques, mixed imaging sequences in the training dataset, and comparative studies with other established models within the computer vision sector.