Comprehensive Summary
This study aims to replace traditional methods of Adhesive Capsulitis of the Shoulder (ACS) diagnosis by developing and comparing machine learning (ML) and deep learning (DL) algorithms. To conduct this study, a total of 444 patients were separated into two groups: the primary cohort (387 subjects), and an external test cohort (57 subjects) from a separate medical center. The models were then developed, tested, and assessed using the area under the receiver operating characteristic curve (AUC). The PD_T2_LightGBM model (using radiomic features and trained with ML) outperformed all DL feature models with a training AUC of 0.975, validation AUC of 0.915, and test AUC of 0.886. In addition, a combined ML model using both radiomic and clinical features also showed beneficial results. This obtained a training AUC of 0.981, a validation AUC of 0.935, and a test AUC of 0.882. It is important to note, however, that the external test set was small and contained relatively few non-ACS patients, which may limit the generalizability of the results.
Outcomes and Implications
ACS severely affects a patient's quality of life and daily function. Current methods of ACS detection often fail to complete early and accurate diagnosis, leading to delayed treatment and prolonged disability. By utilizing ML/DL algorithms, physicians will be able to alleviate symptoms quicker, allowing patients an accelerated return to their normal lives. In addition, automation would also reduce radiologist workload and lower costs by streamlining MRI interpretation. The author mentions clinical implementation is in the future, but will require a multi-center observational study, regulatory approval, and compliance with institutional IT and security requirements.