Abstract
This paper proposes an AI-based diagnostic method using MRI images for rotator cuff injuries to assist in treatment by segmenting tear areas and assessing tear severity. A multi-model deep learning network based on Unet + FPN architecture was developed to automatically segment rotator cuff injury images and determine tear grades. A dataset of 376 patients with 5640 images was used for training, with an additional 94 patients and 1410 images reserved for testing. To optimize segmentation, a tailored matching strategy was applied, achieving an Intersection over Union (IoU) of 0.79 ± 0.01 and a Dice coefficient of 0.75 ± 0.01, indicating high accuracy in segmenting tear areas. For tear severity indicators, the accuracy of estimating retraction (ER) reached 0.92 ± 0.02, and the accuracy of estimating stop tear width (ESTW) reached 0.79 ± 0.01. As the first AI algorithm specifically developed for diagnosing rotator cuff injuries, this platform demonstrates promising accuracy in both tear segmentation and severity assessment, aiming to support doctors in providing efficient, accurate diagnoses of rotator cuff tears.