Automated landmark detection and view positioning assessment of shoulder grashey view radiographs using cascade deep learning: A dual-center validation study.

BACKGROUND: Shoulder Grashey view radiographs are fundamental diagnostic tools in orthopedic practice, with the critical shoulder angle serving as a key predictor of rotator cuff pathology. However, critical shoulder angle measurements are highly sensitive to projection angles, making consistent radiographic acquisition technically challenging in clinical settings. PURPOSE: This study aimed to develop and validate a deep learning model for automatically localizing anatomical landmarks on shoulder Grashey view radiographs and accurately assessing radiographic view positioning. METHODS: We developed a cascade deep learning framework combining RetinaNet and U-Net architectures to automatically detect 14 anatomical landmarks and measure radiographic parameters including critical shoulder angle, acromial index, and ratio of transverse to longitudinal glenoid diameter. The model was trained on 500 radiographs from Seoul Medical Center, with internal testing on 100 radiographs and external testing on 50 radiographs from Asan Medical Center. Performance was compared against two orthopedic residents. RESULTS: On internal validation, our U-Net model achieved superior landmark detection accuracy with average landmark error of 2.04 ± 3.29 mm and critical shoulder angle error of 1.34 ± 1.33°, significantly outperforming both RetinaNet and resident observers (p < 0.001). External testing confirmed robust generalizability, with the U-Net model achieving landmark detection accuracy (3.18 ± 3.88 mm) superior to RetinaNet (4.95 ± 3.64 mm) and first-year residents (5.25 ± 7.09 mm), and comparable to second-year residents (3.79 ± 3.92 mm), (p < 0.001) while maintaining consistent critical shoulder angle measurements (1.52 ± 1.49°). Importantly, model's performance remains consistent regardless of radiographic positioning quality, suggesting potential to standardize measurements, improve diagnostic accuracy, and enhance clinical workflows. CONCLUSION: Our cascaded deep learning model achieves high accuracy in automated anatomical landmark detection and enables reliable measurement of radiographic parameters on shoulder radiographs. The model's performance remains consistent across varying image qualities, suggesting potential to standardize measurements, improve the diagnostic accuracy, and enhance clinical workflows.