MR-Transformer: A Vision Transformer-based Deep Learning Model for Total Knee Replacement Prediction Using MRI

Comprehensive Summary

This study investigates MR-Transformer, a transformer-based deep learning model designed to predict the progression of knee osteoarthritis (OA) to total knee replacement (TKR). The model was trained on two separate cohorts, OAI and MOST, using four MRI sequences. The model leveraged ImageNet pretraining and captured 3D spatial correlations within MRI scans. Cross-validation revealed that MR-Transformer achieved an average AUC of 0.865 across all MRI types. When evaluated against pre-existing models, MR-Transformer outperformed two of the three comparison models. Overall, the model effectively integrated large-scale training data, 3D features, and long-range MRI dependencies. Pretraining data was found to be critical to the model’s accuracy.

Outcomes and Implications

Knee OA is a debilitating disease that significantly lowers the quality of life in patients. Understanding the progression of knee OA to TKR can lead to reduced complications and more personalized treatment plans. This study lays the groundwork for future work in multimodal imaging and improved interpretability in deep learning-assisted OA prognosis. However, the model is limited by its high computational cost and lack of participant diversity, which currently hinders its clinical applicability.