PATIENT-LEVEL CLASSIFICATION OF ROTATOR CUFF TEARS FROM MAGNETIC RESONANCE IMAGING USING A GRAPH NEURAL NETWORK

Aşık S., Yazıcı A., Aşcı M.

14. Türk Omuz ve Dirsek Cerrahisi Kongresi , Adana, Türkiye, 16 - 18 Nisan 2026, ss.284-287, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: Adana
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.284-287
  • Bilecik Şeyh Edebali Üniversitesi Adresli: Evet

Özet

Aim: Rotator cuff tears are among the most common causes of shoulder pain and functional impairment. Accurate assessment of tear severity plays a key role in treatment planning. Although magnetic resonance imaging (MRI) is the gold standard for evaluating these pathologies, interreader variability and differences in experience can affect diagnostic accuracy. In this study, we aimed to develop an artificial intelligence model that accounts for anatomical continuity across MRI slices and automatically classifies rotator cuff tears at the patient level. Method: Shoulder MRI data from 2,447 patients were retrospectively analyzed. For each patient, 16 axial slices were evaluated. Patients were categorized into three groups: normal (n = 1,628), partial-thickness tear (n = 157), and full-thickness tear (n = 662). In the proposed AI model, each patient was treated as a whole; adjacency relationships among MRI slices were represented using a mathematical structure called a graph. Thus, the model evaluates the entire slice sequence for each patient rather than a single slice. A pretrained deep learning network was used for image feature extraction, while inter-slice relationships were modeled using an attention-based graph neural network. Class imbalance across groups was addressed using a specialized loss function. The dataset was split at the patient level into 80% training and 20% testing, and slice-level data leakage was prevented. Findings: In the test set (n = 490), overall diagnostic accuracy was 89.2%. Sensitivity for identifying a normal shoulder was 93.9%, and sensitivity for detecting full-thickness tears was 82.7%. In the clinically important binary classification task, sensitivity and specificity for distinguishing any tear from a normal shoulder were 94.5% and 93.9%, respectively. Sensitivity for differentiating fullthickness tears from the other groups was 90.2%. In the three-class evaluation, the area under the ROC curve (AUC) was 0.93. The reliability of the model’s probability estimates was supported by calibration analyses (Brier score: 0.056). Conclusion: The proposed AI model can classify rotator cuff tears with high accuracy by analyzing anatomical continuity across MRI slices in a holistic manner rather than evaluating slices independently. The high sensitivity for detecting tears suggests that this approach has potential as an assistive decision-support tool in clinical practice. Keywords: Computer-Aided Diagnosis, Deep Learning, Graph Neural Network, Magnetic Resonance Imaging, Rotator Cuff Tear