CSS-Net: Classification and Substitution for Segmentation of Rotator Cuff Tear

Abstract

Magnetic resonance imaging (MRI) has been popularly used to diagnose orthopedic injuries because it offers high spatial resolution in a non-invasive manner. Since the rotator cuff tear (RCT) is a tear of the supraspinatus tendon (ST), a precise comprehension of both is required to diagnose the tear. However, previous deep learning studies have been insufficient in comprehending the correlations between the ST and RCT effectively and accurately. Therefore, in this paper, we propose a new method, substitution learning, wherein an MRI image is used to improve RCT diagnosis based on the knowledge transfer. The substitution learning mainly aims at segmenting RCT from MRI images by using the transferred knowledge while learning the correlations between RCT and ST. In substitution learning, the knowledge of correlations between RCT and ST is acquired by substituting the segmentation target (RCT) with the other target (ST), which has similar properties. To this end, we designed a novel deep learning model based on multi-task learning, which incorporates the newly developed substitution learning, with three parallel pipelines: (1) segmentation of RCT and ST regions, (2) classification of the existence of RCT, and (3) substitution of the ruptured ST regions, which are RCTs, with the recovered ST regions. We validated our developed model through experiments using 889 multi-categorical MRI images. The results exhibit that the proposed deep learning model outperforms other segmentation models to diagnose RCT with 6 ∼ 8% improved IoU values. Remarkably, the ablation study explicates that substitution learning ensured more valid knowledge transfer.