MRI-based automated multitask deep learning system to evaluate supraspinatus tendon injuries

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Musculoskeletal Disorders
• Radiology and Diagnostic Imaging

Background:

• Supraspinatus tendon (SST) injuries are common and require accurate diagnosis.
• Current diagnostic methods for SST injuries can be subjective and time-consuming.
• Advanced imaging analysis using artificial intelligence offers potential for improved diagnostic accuracy.

Purpose of the Study:

• To develop an automated, multitask, MRI-based deep learning system for detailed evaluation of supraspinatus tendon (SST) injuries.
• To classify various types of SST injuries, including normal, degenerative, and different tear configurations.
• To compare the diagnostic performance of the deep learning system against experienced radiologists.

Main Methods:

• A dataset of 3087 patients with arthroscopy-confirmed SST findings was used for training and internal validation.
• The Visual Geometry Group network 16 (VGG16) was employed for initial image screening, followed by a rotator cuff multitask learning (RC-MTL) model for classification.
• External validation was performed on 573 patients, and model performance was assessed using ROC curve analysis, comparing results with radiologists via McNemar's test.

Main Results:

• The automated multitask deep learning system achieved high diagnostic performance, with an average AUC of 0.98 for the in-group dataset and 0.97 for the out-group dataset.
• The system demonstrated superior performance compared to radiologists in classifying SST injuries across all defined groups (p < 0.001).
• The intraclass correlation coefficients (ICCs) for radiologists ranged from 0.97 to 0.99, indicating high reliability.

Conclusions:

• An automated, MRI-based deep learning system effectively diagnoses and classifies supraspinatus tendon injuries with high accuracy.
• The developed system demonstrates performance comparable to that of experienced radiologists.
• This AI tool has the potential to improve diagnostic efficiency, reduce inter-observer variability, and enhance the assessment of SST injuries.