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

Area of Science:

• Orthopedic imaging
• Medical image analysis
• Deep learning in radiology

Background:

• Accurate localization of hip anatomical landmarks is crucial for orthopedic imaging analysis.
• Existing methods often struggle with the complexity of native and prosthetic joints, and challenging anatomical conditions.

Purpose of the Study:

• To develop a high-resolution anatomical landmark detection method for hip anteroposterior radiographs.
• To enhance the accuracy of femoral head and acetabulum localization in both native and prosthetic joints.

Main Methods:

• A two-stage deep learning framework integrating convolutional neural networks (CNNs) and Transformer architectures.
• Global stage: U-Net for coarse landmark estimation on downscaled images.
• Local stage: Detection Transformer on high-resolution patches, modeling landmarks as circles and regressing center, radius, and status.

Main Results:

• Achieved an average center localization error of 1.226 mm and radius error of 0.968 mm on 637 annotated hip radiographs.
• The two-stage model improved hip center detection accuracy by 22% compared to U-Net alone (p < 0.001).
• Demonstrated robust performance in cases of osteoarthritis and surgical alterations.

Conclusions:

• The proposed two-stage framework offers consistent performance improvements over single-stage methods for hip landmark detection.
• Enables reliable and precise localization of key hip landmarks in orthopedic imaging.
• Supports clinical tasks including surgical planning, implant evaluation, and longitudinal monitoring.