2025 ISAKOS Congress in Munich, Germany

2025 ISAKOS Biennial Congress Paper


Application of Artificial Intelligence in Quantifying the Degree of Fatty Infiltration of Rotator Cuff Muscle: A Feasibility Study

Yee Lam Jasmine Louie, MBChB, Shatin HONG KONG
Ryan Hui, MBChb, Hong Kong HONG KONG
Chun Yat Jimmy Wu, BEng, Hong Kong HONG KONG
Jonathan Ng, Hong Kong, Hong Kong HONG KONG
Rex Mak, MBChB, FRCS(Orth), FHKCOS, FHKAM, FRCS, Hong Kong HONG KONG
Cham Kit Dennis Wong , Nt HONG KONG
Michael Tim Yun Ong, MBChB(UK), BSc(UK), MRCS(Edin), MSc(CUHK), FRCSEd, Shatin HONG KONG
Patrick S. H. Yung, MBChB, FRCS(Orth), FHKCOS, FHKAM, FRCS, Shatin, New Territories HONG KONG

The Chinese University of Hong Kong, Shatin, New Territories, HONG KONG

FDA Status Not Applicable

Summary

The study evaluates the feasibility of AI model using image segmentation to quantify fatty infiltration in rotator cuff muscles from MRI images, highlighting its respectable accuracy and reliability while noting challenges in class differentiation, and suggesting that expanding the dataset and incorporating prognostic factors may enhance personalized treatment for patients with rotator cuff tears.

Abstract

Background

Choosing the appropriate surgical treatment for rotator cuff tear (RCT) has been challenging owing to its multifactorial nature. Despite advancements in differentiating between healthy muscle, fibrous tissue and fat, interobserver variability in estimating fatty infiltration remains unresolved.

The aim of the study is to assess the feasibility of applying an image-based deep-learning algorithm in determining the degree of fatty infiltration of rotator cuff muscle.

Methodology

We employed image segmentation techniques to develop the AI model using a U-Net neural network, which features a contracting path for context capture and an expanding path for precise localization to segment MRI images. The model differentiates pixel classes while capturing high-level context and fine details essential for accurate segmentation.

To predict the degree of fatty infiltration of each rotator cuff muscle, we calculate the number of pixels for each class to produce an quantifiable percentage of fatty infiltration.

For training the model, we used a dataset comprising annotated shoulder MRI sections, with 539 images for training set and 135 images for validation set. We resized both masks and images to 256x256 pixels. To enhance model generalization, we applied data augmentation techniques during training, and used cross-entropy as the loss function. The annotation masks included a total of ten different classes, consisting of each of the four rotator cuff muscles, their fatty streaks, and the background.

Results

To evaluate our model, we assessed mean pixel accuracy (mPA), precision, and recall using the validation set. The model achieved a mPA of 62.94%, along with a precision of 64.33% and a recall of 62.94%.

We observed that increasing the number of classes resulted in lower validation accuracy, indicating a tendency to overfit and a struggle to generalize effectively under the increased complexity of class distinctions. Conversely, the model achieved higher validation accuracy with fewer classes, specifically reduced to the supraspinatus and its fatty streak only, resulting in an mPA of 80.55%, with precision and recall at 82.29% and 80.55%, respectively.

These results suggest that while our model excels in simpler classification tasks, it requires further optimization and advanced techniques to handle more classes effectively.

Furthermore, differentiating between the infraspinatus and teres minor was challenging, as both occupy the posterior part of the scapula and often appear as a single muscle on MRI. In contrast, the supraspinatus and subscapularis, each occupy distinct regions of the scapula. Accurately distinguishing them requires tracing muscle fiber striations, which exceeds our AI model’s current learning capabilities.

Discussion And Conclusion

Our AI model demonstrated respectable accuracy and reliability in assessing muscular atrophy of rotator cuff muscles, providing quantifiable fatty infiltration measurements. Higher accuracy can be achieved by expanding the patient database and upgrading the computational capacity, while training the model to trace muscle striation can enhance differentiation between infraspinatus and teres minor.

Conducting randomized controlled trials to compare treatment options in RCT is inconceivable considering the disease heterogeneity. Expanding our AI model to include various prognostic factors in correlation with postoperative outcomes may enhance the development of personalized treatment plans.