Digital image analysis of ultrasound images using machine learning to diagnose pediatric nonalcoholic fatty liver disease

Purpose: Prevalence of nonalcoholic fatty liver disease (NAFLD) in children is rising with the epidemic of childhood obesity. Our objective was to perform digital image analysis (DIA) of ultrasound (US) images of the liver to develop a machine learning (ML) based classification model capable of differentiating NAFLD from healthy liver tissue and compare its performance with pixel intensity-based indices. Methods: De-identified hepatic US images obtained as part of a cross-sectional study examining pediatric NAFLD prevalence were used to build an image database. Texture features were extracted from a representative region of interest (ROI) selected from US images of subjects with normal liver and subjects with confirmed NAFLD using ImageJ and MAZDA image analysis software. Multiple ML classification algorithms were evaluated. Results: Four-hundred eighty-four ROIs from images in 93 normal subjects and 260 ROIs from images in 39 subjects with NAFLD with 28 texture features extracted from each ROI were used to develop, train, and internally validate the model. An ensembled ML model comprising Support Vector Machine, Neural Net, and Extreme Gradient Boost algorithms was accurate in differentiating NAFLD from normal when tested in an external validation cohort of 211 ROIs from images in 42 children. The texture-based ML model was also superior in predictive accuracy to ML models developed using the intensity-based indices (hepatic-renal index and the hepatic echo-intensity attenuation index). Conclusion: ML-based predictive models can accurately classify NAFLD US images from normal liver images with high accuracy using texture analysis features.

Automated summary

This study aimed to develop a machine learning classification model capable of accurately differentiating NAFLD from normal liver tissue in children through digital image analysis. Results showed that the texture-based ML model was superior in predictive accuracy compared to models developed using intensity-based indices.