Multihipposenet: multi-task deep learning for automated DDH assessment in infant hip ultrasound

Developmental dysplasia of the hip (DDH) is a common skeletal condition in infants and young children. For infants under six months, diagnosis relies on Graf ultrasonography to measure the α and β angles, which assess hip-joint development. However, this method requires experienced sonographers with expertise in complex anatomy.

In this study, we propose an end-to-end multi-task learning network that improves DDH assessment by jointly exploiting complementary anatomical information. First, we use ResNet-50 with a feature pyramid network (FPN) and a contextual semantic fusion module to extract multi-scale features from hip anatomy. Second, we employ a multi-expert gated channel transformation (ME-GCT) unit that explicitly models the relationships between tasks and channels to balance collaboration and conflict across tasks. Finally, a multi-task graph adaptive learning mechanism (MT-GALM) mines the hidden dynamic dependencies between adjacent channels across tasks. Two datasets comprising 1,355 annotated hip ultrasound images from 566 infants were constructed for this work.

Across the patient-level cross-validation test folds, 97.97% and 96.81% of the estimated α and β angles fall within 5 degrees of the expert reference.

Because validation was restricted to Graf type I and type II hips, these results should be interpreted as evidence of technical feasibility rather than of established clinical utility; evaluation on type III and type IV hips is required before clinical deployment.