Automated 3D segmentation of rotator cuff muscle and fat from longitudinal CT for shoulder arthroplasty evaluation

Comprehensive Summary

This study aims to develop and validate a deep learning model for automating the 3D segmentation of rotator cuff muscles on longitudinal CT scans. Using DeepLabV3 + and ResNet50 as the workflow backbone, the models were constructed and trained on 53 TSA (total shoulder arthroplasty) subjects. The trained models were then evaluated on a cohort of 172 patients. As a result, compared to the ground truth, models achieved a mean Dice score of 0.928 and 0.916, mean ASSD (average symmetric surface distance) of 0.844 mm and 1.028 mm, mean HD95 (95th percentile Hausdorff distance) of 3.071 mm and 4.173 mm, and mean RAVD (relative absolute volume difference) of 0.025 and 0.068 on the CT scans, respectively. Although these results are promising, it is important to mention that this study lacks external validation from other institutions; therefore, the data cannot yet be completely generalized.

Outcomes and Implications

With over 100,000 TSA surgeries being performed annually, it is extremely important to maximise post-operative outcomes. Currently, there is significant evidence that post-opperative success is highly dependent on the muscle volume and fat fraction of patients. These measurements, however, are highly dependent on the costly and low-visibility automated segmentations of MRI (magnetic resonance imaging) images of the shoulder. By automating CT scans instead, doctors can reduce treatment costs and increase pre-operative visibility within the images for more efficient pre-operative planning. Although the author does not explicitly comment on a timeline for clinical implementation, the author does point out clear clinical utility and segmental success; however, they will most likely need to gain external validation by utilizing subjects from other institutions.