Multi-View Cross Attention for Guiding Surgical Planning in Rotator Cuff Repairs With Limited MRI Data

초록

Accurately determining surgical strategies, such as complete or partial repairs, is essential for effective treatment planning in rotator cuff injuries. This study proposes a novel two-stage framework leveraging multi-view cross attention to analyze magnetic resonance imaging (MRI) data from coronal and sagittal views, addressing the challenges posed by limited datasets. In Stage 1, segmentation of six clinically relevant regions from each view was performed for feature extraction. In Stage 2, segmentation results were combined with original MRI images using cross-view transformers, enabling the extraction of interdependencies between segmented and original data. To further enhance prediction accuracy, multi-head cross attention was employed to merge features from both views, capturing their spatial and anatomical relationships. The proposed framework achieved an area under the curve (AUC) of 0.95 in predicting surgical repair types, effectively distinguishing between complete and partial repairs. While segmentation performance in Stage 1 showed variability (Intersection over Union, IoU: 0.45-0.83) due to the limited dataset size (211 samples), the incorporation of multi-view cross attention significantly enhanced prediction accuracy by leveraging complementary information from both sagittal and coronal views. This demonstrates the framework's capability to mitigate the challenges of limited data in segmentation tasks by improving overall classification performance.

키워드