The Evaluation of Developmental Dysplasia of Hip from Pediatric Ultrasound Images by Key Point-Based Deep Learning

Background: Developmental dysplasia of the hip (DDH) is a common hip disorder in infants. Ultrasound is the primary imaging modality for infants younger than 6 months, but measurements using the Graf and Harcke methods depend on examiner experience. This study aimed to develop and internally validate a key point-based deep learning model for the automated measurement of the Graf α and β angles and femoral head coverage. Methods: A total of 300 standard-plane hip ultrasound images from 165 infants aged 0–6 months were retrospectively collected at a single tertiary center, and reference key points were determined by consensus of pediatric orthopedic surgeons. We developed a dual-branch network, named “DDH-KPNet,” which combines a UNet++ convolutional branch and a transformer branch fused with attention modules. Performance was evaluated with fivefold cross-validation at three levels: accuracy of key point localization; intraclass correlation coefficient (ICC); and DDH classification. Results: The experimental results demonstrated that Graf-based classification achieved accuracy = 0.903, sensitivity = 0.81, and specificity = 0.9497 (under 95% conference interval); Harcke classification achieved accuracy = 0.942, sensitivity = 0.888, and specificity = 0.968 (under 95% conference interval); and ICCs for α = 0.832, β = 0.728, and coverage = 0.923. Conclusions: The proposed model showed promising agreement with expert measurements under internal validation. Multicenter external validation and independent diagnostic analysis by different physicians are required before clinical implementation.