DDH-Net: A Graf-Compliant Method for Computer-Aided Assessment Using Infant Hip Ultrasound.

Ultrasound assessment of developmental dysplasia of the hip (DDH) in infants is highly operator-dependent, particularly during standard plane acquisition and Graf angle measurement. This paper proposes DDH-Net, a fully automated computer-aided assessment method for infant hip ultrasound that follows the complete Graf workflow. First, YOLOv8n-pose is used to detect eight anatomical structures and two ilium orientation keypoints for standard plane determination. Regions of interest (ROIs) are then cropped according to the detection results, and U-Net is employed to perform fine segmentation of the ilium, labrum, and lower limb of the ilium. Finally, Graf reference lines are constructed from the segmentation results to enable automatic measurement of the α and β angles. The study included 1409 infants, comprising 2648 ultrasound images and 50 ultrasound videos. On an independent test set of 529 images, the accuracy, sensitivity, and specificity of standard plane detection were 96.8%, 94.5%, and 100.0%, respectively. The model automatically saved 187 candidate frames from the videos, of which 93.1% were rated by experts as having no or only minor clinically relevant discrepancies; 48 of the 50 videos (96.0%) contained at least one clinically acceptable candidate frame. Compared with full-image segmentation, ROI-based segmentation reduced processing time by 33.7%. In 307 standard plane images, the mean absolute errors (MAE) for the α and β angles were 1.61° and 2.13°, respectively, with corresponding intraclass correlation coefficients (ICC) of 0.913 and 0.766. The complete pipeline achieved a processing speed of 28.87 frames per second, indicating that DDH-Net has the potential to provide efficient and interpretable analysis of infant hip ultrasound images.