Segmentation method using U-net for assessment of rotator cuff muscles degeneration

Abstract

Organic deterioration in the rotator cuff muscles often causes shoulder pain, limitation of motion and shoulder disabilities in patients. Fat content, as calculated by the Dixon method, is a quantitative assessment of rotator cuff deterioration and is relevant to the outcome of reconstructive surgery. However, Dixon method requires four different muscle regions of interest on shoulder MR images manually, which is a barrier to its clinical dissemination. In this study, rotator cuff muscles segmentation method is proposed to aid clinical quantitative assessment: using 28 MR images, Segmentation of the four rotator cuff muscles and supraspinatus fossa and from these the fat content and muscle atrophy rates of the Dixon method are estimated. As a result Dice coefficients of 0.93, 0.95, 0.88, 0.82, and 0.96 were shown for the subscapularis, supraspinatus, infraspinatus, teres minor, and supraspinatus fossa, and the atrophy rate was also highly accurate with RMSE = 4.29. In addition, it was confirmed that the accuracy of segmentation affects the prediction accuracy of fat content.