Deep Learning for the Automatic Quantification of Rotator Cuff Muscle Degeneration from Shoulder CT Data Sets

Purpose: Assessing and quantifying the condition of rotator cuff muscles is useful in various shoulder disorders including glenohumeral osteoarthritis. Although magnetic resonance imaging (MRI) is the examination of reference due to its high soft tissue contrast capabilities, computed tomography (CT) is more accessible and increasingly being performed in the preoperative planning of shoulder arthroplasty, notably for measuring glenoid version and glenohumeral subluxation. However, manual segmentation and quantitative analysis of rotator cuff muscles in CT images would benefit from full automation. We aimed to propose and evaluate a deep-learning method for the automatic quantification of rotator cuff muscle degeneration on shoulder CT images.

Methods and Materials: The presumed contours of healthy/premorbid rotator cuff muscles from 127 patients scanned with CT during their preoperative planning for shoulder arthroplasty were manually drawn by three independent musculoskeletal radiologists on a standardized sagittal-oblique CT section. These premorbid muscle segmentations were also predicted by a deep convolutional neural network (CNN), following a fivefold cross-validation to train and test iteratively the CNN based on the three manual segmentations. Automatic segmentations from the CNN were evaluated against reference segmentations for each muscle created by aggregating the three manual segmentations using an expectation-maximization algorithm. Automatically predicted premorbid segmentations were then used to quantify the ratio of muscle atrophy, fatty infiltration, and overall muscle degeneration. Automatic results for each of these three parameters were compared with values obtained semiautomatically by radiologists.

Results: Average Dice coefficients for automatic segmentations with the CNN (89% ± 9%) were comparable with manual segmentations (89% ± 6%). No significant differences were found for the subscapularis, supraspinatus, and teres minor (p ≥ 0.120); Dice coefficients were significantly higher for automatic segmentations of the infraspinatus (p = 0.012). The automatic method was able to provide comparable and reliable estimates of muscle atrophy (R² = 0.87), fatty infiltration (R² = 0.91), and overall muscle degeneration (R² = 0.91).

Conclusion: Deep learning allows for the rapid automatic quantification of rotator cuff muscle atrophy, fatty infiltration, and overall degeneration from shoulder CT data sets, with a diagnostic accuracy comparable with human observers. This method could be integrated as an additional parameter in the preoperative CT planning of shoulder arthroplasty and other shoulder disorders. It could further be used to correlate the condition of rotator cuff muscles with surgical and clinical outcomes.