Text

Objectives and questions: The differentiation between traumatic and degenerative rotator cuff tears (RCTs) remains a diagnostic challenge with significant implications for treatment planning. While magnetic resonance imaging (MRI) is standard practice, traditional radiological interpretation has shown limited reliability in distinguishing these etiologies. This study aims to evaluate the potential of artificial intelligence (AI) models, to differentiate between traumatic and degenerative RCT.

Material and methods: A retrospective study was conducted using an annotated dataset of 683 patients (1470 samples) from Balgrist University Hospital. The proposed framework integrates magnetic resonance imaging (MRI) with structured clinical data. To address the issue of data scarcity, a Class-Conditional Latent Diffusion Model (LDM) was implemented to synthesise high-fidelity MRI samples, while a Tabular Variational Autoencoder (TVAE) was utilized to augment structured clinical records. The multimodal architecture employs a Vision Transformer (ViT) and TabNet architecture, utilising advanced fusion strategies to synthesise visual and tabular inputs. In order to ensure clinical interpretability, explainable AI techniques, including Grad-CAM++ and Saliency maps, were integrated into the workflow.

Results: The multimodal Vision Transformer with TabNet achieved an Accuracy of 87%, an AUC of 65%, a F1-Score of 38% with a precision of 38% and recall of 38% on the test set.

Discussion and conclusions: The study demonstrates that a multimodal learning approach combining MRI data and structured clinical variables can improve the differentiation between degenerative and traumatic rotator cuff tears. As a comparison with the results of the pilot study, the performance of the model was significantly enhanced, particularly through the incorporation of synthetically augmented MRI data and advanced fusion methods. The findings indicate that both data augmentation and multimodal feature integration contribute significantly to the model’s predictive capability. However, the experiments demonstrate that expanding to a more extensive and heterogeneous dataset poses significant challenges. The elevated complexity of the dataset, coupled with the pronounced class imbalance, appears to compromise the model’s capacity for generalization, particularly concerning the minority traumatic class. This finding suggests that the constraints on model performance are currently more attributable to issues of data quality and distribution than to architectural capacity. The utilization of explainable AI techniques facilitates the acquisition of valuable insights into the model’s decision-making process and supports interpretability, which is imperative for clinical and medicolegal applications. Consequently, future research should concentrate on curating cleaner and more balanced datasets, as well as refining labels, to fully exploit the potential of multimodal deep learning for robust and clinically reliable decision support.