Arthroscopic Images Predict Tendon Integrity After Arthroscopic Rotator Cuff Repair Using a Deep Learning Model.

PURPOSE: To assess the feasibility of a deep learning model for predicting early structural integrity after arthroscopic rotator cuff repair (ARCR) using intraoperative arthroscopic images obtained after the double-row bridging repair technique. METHODS: This retrospective study included patients who underwent ARCR between January 2016 and June 2024, provided that postoperative tendon integrity was assessed using ultrasonography at a minimum follow-up of 6 months. Intraoperative arthroscopic images of the postrepaired tendon were collected. The dataset was randomly divided into training, validation, and test sets in a 7:1.5:1.5 ratio at the patient level. Transfer learning was performed using the pretrained ConvNeXt-Tiny architecture integrated with a Convolutional Block Attention Module. Model performance was evaluated using the area under the receiver operating characteristic curve, precision-recall curve, sensitivity, specificity, and F1 score. Clinical utility was evaluated using decision curve analysis and a reliability curve. RESULTS: A total of 1047 intraoperative arthroscopic images were collected from 642 patients in the healed group, and 151 images were obtained from 85 patients in the retear group with a mean follow-up of 6 months. The performance of the model on the test dataset (81 patients [71 healed and 10 retear], corresponding to 170 images [146 healed and 24 retear]) achieved a sensitivity of 100% (95% confidence interval [CI], 1.0-1.0) and specificity of 83.3% (95% CI, 0.66-0.96), with an overall classification accuracy of 97.6%. Model performance yielded an area under the receiver operating characteristic curve of 0.95 (95% CI, 0.84-0.99), precision-Recall Area under the curve of 0.98 (95% CI, 0.96-0.99) and F1 score of 0.98 (95% CI, 0.97-0.99). Eigen-Class Activation Map visualizations revealed that the model predominantly focused on the repaired tendon edges and footprint zones during prediction. CONCLUSIONS: The ConvNeXt-Convolutional Block Attention Module deep learning model can predict early tendon structural integrity with high internal accuracy after ARCR using only intraoperative arthroscopic images of the repaired tendon from a single-center dataset. CLINICAL RELEVANCE: The image-based deep learning model may assist surgeons in screening patients at high risk of retear after ARCR, thereby supporting delayed postoperative rehabilitation, using a decision threshold of 0.40 healing probability.