Related Experiment Video
Updated: Jun 30, 2026
Measuring 3D In-vivo Shoulder Kinematics using Biplanar Videoradiography
Published on: March 12, 2021
Hybrid deep learning ensemble model for detecting small to medium rotator cuff tears from shoulder radiographs
Shun-Wun Jhan1,2,3, Tian-Hsiang Huang4, Jai-Hong Cheng3,5
1Doctoral Degree Program in Biomedical Engineering, College of Medicine, Kaohsiung Medical University, Kaohsiung, Taiwan.
Background:
Rotator cuff tears (RCTs) represent a common orthopedic condition, the diagnosis of which often requires advanced imaging techniques. Notably, plain radiography may not accurately differentiate between small to medium RCTs and other shoulder pathologies. Deep learning (DL) models may facilitate preliminary assessment of RCTs, helping avoid unnecessary advanced imaging and expediting patient care.
Methods:
For this study, we used a dataset comprising 587 shoulder radiographs (339 from patients with small to medium RCTs and 248 from without RCTs). The study dataset was divided into a training set (406 images [69% of the dataset]) and a validation set (101 images [17% of the dataset]). In addition, an independent test set consisting of 80 radiographs (39 with small to medium RCTs and 41 without RCTs [14% of the dataset]) was used for external validation of the model performance. The test set was strictly held out and not used during model training or model selection. Three pretrained DL models (ResNet-50, ResNet-101, and InceptionV3) were fine-tuned using transfer learning. These three models were integrated using a majority voting strategy to develop a hybrid DL ensemble model.
Results:
On the validation set, the hybrid DL ensemble model outperformed the individual models in classifying rotator cuff images, achieving an accuracy of 78% (95% CI 69-86), a precision of 79% (68-88), a recall (sensitivity) of 87% (77-95), a specificity of 66% (51-80), an F1-score of 0.83 (0.74-0.89) and an area under the receiver operating characteristic curve (AUROC) of 0.85 (0.76-0.92). On the independent test set, the hybrid DL ensemble model achieved an accuracy of 81% (73-90), a precision of 72% (60-84), a recall of 100% (100-100), a specificity of 63% (48-78), an F1-score of 0.84 (0.75-0.91), and an AUROC of 0.93 (0.87-0.97).
Conclusion:
Our hybrid DL ensemble model can detect small to medium RCTs from shoulder radiographs. Thus, it holds promise as a cost-effective and efficient tool for preliminary assessment of RCTs. This artificial intelligence-based approach has the potential to improve diagnostic reliability and assist clinical decision-making as a screening and triage support tool for RCTs.
Level Of Evidence Iii:
Retrospective study.
