A two-stage deep learning model for predicting rotator cuff tear reparability from preoperative MRI

This study investigates the effectiveness of the Patte classification and Goutallier grade in predicting the reparability of rotator cuff tears using advanced artificial intelligence (AI) models. An analysis was conducted on preoperative shoulder magnetic resonance image (MRI) scans (coronal and sagittal images presenting Patte grade and Goutallier classification) of 243 patients who previously underwent rotator cuff repair. Of these, 190 patients underwent complete repairs, while 53 had partial repairs. The proposed model operates in two stages, with MRI scans as its primary input. Initially, U-net based segmentation models are employed to precisely delineate critical areas within the shoulder anatomy relevant to the surgery. Subsequently, the model advances to the second stage, using DenseNet. This stage involves a classification process where the model evaluates the segmented images to predict the feasibility of either a complete or partial rotator cuff repair. The first stage, involving U-net based segmentation, achieved 62–73% accuracy in identifying relevant shoulder anatomy. The second stage, employing DenseNet for predicting reparability (complete vs. partial repair), attained a 56–75% accuracy level. This study presents a two-stage AI approach that uses the anatomical features underlying the Patte classification and Goutallier grade to predict rotator cuff reparability from routine preoperative MRI in a single-center cohort. Although current performance remains modest, the findings point to the need for incorporating additional clinical factors, expanding to multicenter datasets, and refining the model architecture to achieve clinically meaningful predictive accuracy.