Abstract
BACKGROUND: The current method of classifying fatty infiltration is highly subjective and has low reliability, which may impact the decision-making for the management of rotator cuff tears. The purpose of this study was to present and evaluate a new deep-learning (DL) approach to automatically and objectively classify fatty infiltration of rotator cuff muscles on magnetic resonance imaging (MRI). METHODS: A validated dataset of 1,149 images of segmented rotator cuff muscles, derived from 383 patients, were classified using a simplified grading system (normal, mild, severe) proposed based on the original Goutallier classification. These images and their classifications were used to train the artificial intelligence models. A novel DL pipeline comprising key components of in-domain transfer learning, feature fusion, and machine learning classifiers was proposed for automatic fatty infiltration classification. Pretrained DL models Xception, InceptionV3, and MobileNetV2 were trained separately. Then, K-Nearest Neighbor, Support Vector Machines, and Naive Bayes classifiers were trained using fused features extracted by 3 DL models from the delineated rotator cuff muscle areas. Performance metrics, including accuracy, precision, recall, F1-score, and Gradient-Weighted Class Activation Mapping visualizations, were used to evaluate the model's performance. RESULTS: Among the individual models, MobileNetV2 demonstrated the highest overall performance, with accuracy of 89.5%, specificity of 94.7%, recall of 89.5%, precision of 90.5%, and F1-score of 90.0%. After feature fusion, K-Nearest Neighbour achieved the highest performance, with accuracy of 91.1%, specificity of 95.5%, recall of 91.1%, precision of 93.1%, and F1-score of 92.1%. Overall, the performance metrics of the feature fusion were higher compared to the individual models and approached the consistency of clinical experts (intraclass correlation coefficient 0.91). CONCLUSION: This study provides evidence for the effective utilization of artificial intelligence advancements in the automated classification of fatty infiltration of rotator cuff muscles on MRI using in-domain transfer learning, feature fusion, and machine learning classifiers. By combining the power of these 3 components, the proposed approach has excellent potential to achieve accurate, robust, and enhanced classification, with a level of consistency in line with expert agreement. As such, this approach offers a promising solution for automating the classification of fatty infiltration on MRI which may have potential benefit for daily clinical practice.