Two-stage deep learning for circular landmark detection in hip radiographs

This study develops a high-resolution anatomical landmark detection method for hip anteroposterior radiographs, designed to enhance the accuracy of femoral head and acetabulum localization in both native and prosthetic joints. A two-stage deep learning framework is introduced that integrates convolutional and Transformer-based architectures. On the global stage, a U-Net estimates coarse landmark positions from downscaled images. These predictions guide the cropping of high-resolution patches, which are subsequently processed in the local stage using a Detection Transformer. Unlike conventional point-based methods, the proposed method models each landmark as a circle and directly regresses its center, radius, and status (natural or prosthetic). On a dataset of 637 annotated hip radiographs, the framework achieved an average center localization error of 1.226 mm (± 1.410) and a radius error of 0.968 mm (± 1.312). Compared with U-Net alone, the two-stage model improved hip center detection accuracy by 22% (p < 0.001), while maintaining robust performance under challenging anatomical conditions, such as osteoarthritis and surgical alterations. These results demonstrate that the proposed two-stage framework provides consistent performance improvements over single-stage approaches. This method enables reliable and precise localization of key hip landmarks, supporting clinical tasks such as surgical planning, implant evaluation, wear assessment, and longitudinal monitoring in orthopedic imaging.