AI-Based Automated CT Segmentation of Shoulder Muscles: Functional Muscle Volume and Postoperative Range of Motion After Arthroscopic Rotator Cuff Repair – A Retrospective Study

Background Muscle quality and volume are important determinants of functional outcomes after arthroscopic rotator cuff repair (ARCR). Although fatty infiltration of rotator cuff muscles has been widely evaluated using semi-quantitative methods, comprehensive three-dimensional quantitative assessment of both intrinsic and extrinsic shoulder muscles remains limited. This study aimed to evaluate AI-driven automated shoulder muscle segmentation on preoperative computed tomography (CT) and to investigate the associations between quantitative muscle characteristics and postoperative shoulder function after ARCR. Methods This retrospective study included 30 patients (31 shoulders) who underwent ARCR, with a mean follow-up of 20.2 months. Preoperative CT scans were obtained using a standardized protocol with 1-mm slice thickness. Shoulder muscles were automatically segmented using an AI-based deep learning model. Six muscles—the deltoid, trapezius, subscapularis (SSc), supraspinatus (SSP), infraspinatus and teres minor complex (ISPTm), and teres major (TM)—were analyzed. Muscle composition was quantified based on Hounsfield unit thresholds and classified into functional muscle (FM), low-attenuation muscle (LAM), and adipose tissue (AT). Muscle volumes were normalized to scapular bone volume. Associations between quantitative muscle parameters and postoperative active range of motion (ROM) at the final follow-up were evaluated using Pearson correlation analysis. Results FM and AT proportions differed significantly among the six muscles, whereas LAM showed no significant variation. The TM demonstrated the highest FM proportion, while the SSP showed the highest AT proportion. The total volumes of the SSc and TM were associated with postoperative anterior elevation (r = 0.410 and 0.415, respectively), as were the FM volumes of the SSc, ISPTm, and TM (r = 0.458, 0.430, and 0.427, respectively). In contrast, neither AT nor LAM measures were significantly associated with postoperative ROM. No muscle parameter was significantly associated with postoperative external rotation. Conclusions AI-driven automated segmentation enabled the quantitative evaluation of shoulder muscle volume and composition on preoperative CT. Both intrinsic and extrinsic shoulder muscles, particularly the SSc, ISPTm, and TM, were associated with postoperative anterior elevation after ARCR. Greater functional muscle volume, rather than fatty muscle composition, was associated with better postoperative anterior elevation after ARCR. AI-based quantitative muscle assessment may provide clinically relevant information to support preoperative evaluation.