Comprehensive Summary
In this study, Pulik et al. developed a machine learning model that identifies anatomical features in hip ultrasound (US) images to support the diagnosis of Developmental Dysplasia of the Hip (DDH). A dataset of 685 US images were obtained and annotated by qualified orthopedic surgeons using the widely-accepted Graf method. Two segmentation models were evaluated: Model-5 and Model-8. Model-5 used data from five anatomical structures, while Model-8 incorporated two additional anatomical points and a baseline. Five architectures (SegFormer, OCRNet, U-Net, U-HRNet, and SegNeX) were tested for each model, with performance assessed with Intersection over Union (IoU) and Dice Similarity Coefficient (DSC). The best performance was achieved using the SegNeXt architecture with an MSCAN_L backbone, yielding IoU and DSC scores of 0.632/0.774 for the chondro-osseous border, 0.916/0.956 for the femoral head, 0.625/0.769 for the labrum, 0.672/0.804 for the cartilaginous roof, and 0.725/0.841 for the bony roof. While there is room for further research and improvement, findings suggest that this novel model can be used to aid in reliable clinical diagnoses of DDH.
Outcomes and Implications
The diagnosis of DDH is imperative to ensure that osteoarthritis or disability does not develop in patients later in life. However, ultrasound-based interpretation currently relies on operator expertise and is time consuming, leading to variability in interpretation and outcomes. The successful implementation of an automated segmentation model has the potential to reduce operator dependence, increase diagnostic consistency, and enhance large-scale, efficient screening for DDH in the future. Future developments should incorporate real-time angle measurement and automated classification to enhance clinical decision-making and guide treatment planning.