Abstract
<title>Abstract</title> <p>Developmental Dysplasia of the Hip (DDH) is a prevalent congenital pediatric musculoskeletal deformity, and accurate early International Hip Dysplasia Institute (IHDI) grading is essential for timely clinical intervention and favorable prognosis. A novel framework for reliable IHDI classification of DDH from pelvic radiographs using a prior-constrained lightweight HRNet (PCL-HRNet) is presented to support clinical decision-making in primary care. This method builds multi-dimensional geometric prior loss to constrain keypoint estimation, and then implements lightweight optimization through knowledge distillation, structured channel pruning and INT8 static quantization sequentially. The compressed network is globally optimized by the Quantum-behaved Particle Swarm Optimization with Adaptive Weight Adjustment (QPSO-AWA) algorithm to search for optimal hyperparameters for stable multi-class IHDI grading. Keypoint localization error and anatomical logic error rate are adopted to evaluate the prior constraint performance, while classification performance is measured using Macro-F1, Kappa coefficient, mean distance error (MDE) of keypoints, and mean absolute error (MAE) of acetabular angle. The proposed method reduces the anatomical error rate to 0.24%, and achieves a Macro-F1 of 0.825 with merely 0.33% accuracy loss after lightweight compression. Finally, the proposed PCL-HRNet outperforms mainstream single-baseline deep networks in both diagnostic accuracy and CPU inference efficiency, providing a deployable intelligent decision support tool for DDH screening.</p>