Defect-adaptive landmark detection in pelvis CT images via personalized structure-aware learning

Area of Science:

• Medical Imaging
• Computer Vision
• Orthopedic Surgery

Background:

• Accurate localization of pelvic anatomical landmarks is vital for orthopedic preoperative planning.
• Existing automatic methods struggle with defective bone structures common in clinical cases.

Purpose of the Study:

• To propose DADNet, a defect-adaptive detection network for accurate and robust landmark detection in defective pelvis CT images.
• To incorporate personalized structural priors to enhance landmark detection performance.

Main Methods:

• DADNet constructs a structure-aware soft prior map encoding landmark spatial distribution.
• A patch-based context-aware network performs landmark regression guided by the prior map.
• A bone-aware detection loss enhances robustness in defective regions, with dynamic weight adjustment.

Main Results:

• DADNet achieved an average detection error of 1.252 ± 0.075 mm on severely defective pelvic CT cases.
• The method significantly outperformed existing techniques on public and private datasets.
• Demonstrated strong adaptability to anatomical variability and structural incompleteness.

Conclusions:

• DADNet offers accurate and robust landmark detection in challenging clinical scenarios with pelvic bone defects.
• The proposed framework shows promise for improving preoperative planning in orthopedic surgery.
• Personalized structural priors and defect-adaptive strategies are key to enhanced performance.