Developmental dysplasia of the hip (DDH), a developmental deformity with an incidence of 0.1-3.4%, lacks an objective and reliable definition and assessment metric by which to conduct timely diagnosis. This work aims to address this challenge by developing a system of analysis to accurately detect 22 key anatomical landmarks in anteroposterior pelvic radiographs of the juvenile hip, from which a range of novel salient morphological measures can be derived. A coarse-to-fine approach was implemented, with six model variations of the U-Net deep neural network architecture compared for the coarse model and four variations for the fine model; model variations included differences in data augmentation applied, image input size, network attention gates, and loss function design. The best performing combination achieved a root-mean-square error in the positional accuracy of landmark detection of 3.79 mm with a bias and precision in the x-direction of 0.03 ± 17.6 mm and y-direction of 1.76 ± 22.5 mm in the image frame of reference. Average errors for each morphological metric are in line with the performance of clinical experts. Future work will use this system to perform a population analysis to accurately characterize hip joint morphology and develop an objective and reliable assessment metric for DDH.
Gartland et al. (2025) studied this question.