Comparative Evaluation of Hand-Engineered and Deep-Learned Features for Neonatal Hip Bone Segmentation in Ultrasound

Abstract

Developmental dysplasia of the hip (DDH) is one of the most common congenital disorders seen in newborns. Undetected cases may lead to serious consequences including limping, leg length discrepancy, pain, osteoarthritis, disability, and total hip replacement. Diagnosis typically relies on ultrasound (US) screening of the infant hip between 0–4 months of age. An inexpensive and safe non-ionizing modality, US imaging enables measurement of DDH metrics based on hip bone features such as the $\alpha$ angle. Though DDH assessment remains mostly a manual process in clinical practice, notorious for its significant operator variability, a number of automated measurement methods were recently proposed. These computational methods rely on highly engineered, hand-crafted features most notable of which are phase-based bone image features. Though promising, especially as they were shown to significantly reduce user variability, challenges remain with regards to robustness, as well as generalizability to new data. To improve bone localization, from which the metrics are calculated, we first build upon recent phase-based feature extraction by applying spatial anatomical priors to eliminate false positives and accurately segment the ilium and acetabulum contour. Second, we propose the use of deep-learned features, using the popular U-Net with single and multi-channel inputs. We observe superior performance of deep-learned features compared to the enhanced engineered features including shadow peak and confidence-weighted phase symmetry. We present quantitative evaluation on extensive data from two clinical datasets collected with two different ultrasound probes as part of a study we performed on a cohort of 103 pediatric patients.