P0016 - Digital Image Analysis (DIA) of Ultrasound Images Using Machine Learning to Diagnose Pediatric Nonalcoholic Fatty Liver Disease (NAFLD)

Background: Prevalence of non-alcoholic fatty liver disease (NAFLD) is rising with the epidemic of childhood obesity. NAFLD is often a silent condition, so disease progression occurs without symptoms. Screening for NAFLD with point-of-care ultrasonography (POCUS) is convenient. Although POCUS can be done at the bedside by non-radiologists, awaiting a radiologist’s interpretation can delay diagnosis. Texture in a digital image refers to the distribution of gray-level values among pixels of a given “Region of Interest” (ROI). Texture analysis can be used to detect abnormal tissue. Pixel-intensity based indices such as hepatic-renal index (HRI) and hepatic echo-intensity attenuation index (HEAI) have been used to quantify NAFLD on ultrasound images (Gao 2014) but has low accuracy. An accurate and automated machine learning-based digital tool based on texture features can be useful in bedside screening for NAFLD using POCUS.

Objective: To perform digital image analysis (DIA) of hepatic ultrasound images and develop a machine learning-based classification model capable of differentiating NAFLD from normal liver in pediatric subjects and compare its performance with pixel intensity-based indices.

Methods: We used de-identified hepatic ultrasound images obtained as part of a prior cross-sectional study examining NAFLD prevalence in children conducted at the same institution (May 2018 - July 2019), to build an image database. Representative ROIs were digitally selected from ultrasound images of subjects with normal liver, and with NAFLD, as interpreted by a pediatric radiologist. Images were analyzed with ImageJ and MAZDA software. Texture analyses were performed using histogram, run-length matrix, co-occurrence matrix, gradient analysis, auto-regressive model, and Haar wavelet energy co-efficient indices. An ensemble of machine learning classification models was tested using the SPSS flow of IBM Watson studio software.

Results: 484 ROIs from images in 93 subjects without NAFLD and 260 ROIs in 39 subjects with NAFLD with 30 texture features extracted from each ROI were used to develop, train, and internally validate the model. Support Vector Machine, Neural Net, and Extreme Gradient Boost were the best classifiers based on the area under the curve (AUC) of the receiver operating characteristic curves (Table 1) with the best predictors shown in Figure 1. In an external validation cohort of 211 images from 42 children, the model was accurate in differentiating NAFLD from normal liver with an AUC of 0.92 and was superior to the intensity-based indices (hepatic-renal index, AUC 0.74 (95% CI, 0.60-0.88), p < 0.001 and the hepatic echo-intensity attenuation index, AUC 0.66 (95% CI, 0.51-0.82), p < 0.001.)

Conclusion: Machine learning-based predictive models can accurately classify NAFLD POCUS images from normal liver images with high accuracy using texture analysis features. Cloud deployment of this model may serve as an automated, practical, and cost-effective screening tool for NAFLD detection by clinicians at the bedside.