Uncertainty Estimation for Assessment of 3D US Scan Adequacy and DDH Metric Reliability

Abstract

Developmental Dysplasia of the Hip (DDH) is the most common paediatric hip disorder and a major cause of early hip replacement and osteoarthritis (OA) in young adults. Clinical practice for diagnosis remains reliant on manual measurement of pediatric hip joint features from 2D Ultrasound (US) scans, a process plagued with high inter/intra operator and scan variability. Recently, 3D US was shown to be markedly more reliable with deeply-learned image features effectively used to localize and measure anatomical bone landmarks. However, opaqueness of neural-net based analysis provides no means for assessing the reliability of computed results, a limitation that hampers deployment in clinical settings. We propose using interpretable uncertainty measures that can simultaneously measure bone segmentation reliability and quantify scan adequacy in clinical DDH assessment from 3D US. Our approach measures the variability of estimates generated from an encoder-decoder type CNN optimized for hip joint localization using random dropout. We quantitatively evaluate our proposed uncertainty estimates on a clinical dataset comprising 118 neonates. Results demonstrate smaller variability in dysplasia metrics to be markedly correlated with higher Dice scores for repeated segmentation estimates. Further, we observe that US scans with lower dysplasia metric variability are strongly associated with those labelled as clinically adequate by a human expert. Findings suggest that our uncertainty estimation may improve clinical workflow acting as a quality control check on deep learning based analysis. This in turn may improve overall reliability of the diagnostic process and the prospects of adoption in clinical settings.