Predicting Rotator Cuff Tear Severity Using Radiographic Images and Machine Learning Techniques

Rotator cuff (RC) tears can cause acromion sclerosis associated with quantitative correlation between the severity of tear and severity of acromial sclerosis. The object of this study was to determine the effectiveness of X-ray image processing by machine learning (ML) techniques in the assessment of the severity of rotator cuff (RC) tear. The accuracy of ML diagnosis was compared with the accuracy of physician diagnosis. 145 patients including 72 patients with full-thickness rotator cuff tears, 50 patients with partial rotator cuff tears, and 23 patients with Bankart lesions, who underwent arthroscopic repair were recruited in this retrospective study. Before surgery, X-ray radiography was performed to diagnose the RC tear type. Image processing software Tensorflow and Keras (Image-Data-Generator) (Sequential) with convolutional neural network (CNN) structure were used to differentially diagnose partial tear and full-thickness rotator cuff tears. 80% of images were used for model training and 20% of images for model validation. The results demonstrated that the accuracy of physician diagnosis-based X-ray was 72.6% for full tears and 60.3% for partial tears, respectively. The accuracy of CNN diagnosis-based X-ray was 79.6% for full tear and 87.5% for partial tear, respectively. CNN discriminated partial tear from no-tear with a higher accuracy than human vision (Chi Square, Pearson test, p<0.001). This study presents a novel approach for the diagnosis of RCT using X-ray images and ML techniques that can assist the orthopaedic surgeon using plain X-rays to determine future treatment plans.