Classical & neural network analysis of ultrasound images: estimation of fat & fibrosis content in diffuse liver disease

A system designed to detect diffuse liver disease and quantify the associated fat and fibrosis content using ultrasound images of the liver is described. Established and original image features were analysed using multivariate discriminant (MD) and artificial neural network (ANN) classifiers. An "Autosector" scanner with single element 5MHz transducer was used for image acquisition, as this contributed minimal processing to images. Ultrasound images (n = 195) from 24 normal subjects and 24 patients suffering histologically confirmed diffuse liver disease were acquired under consistent conditions. Needle biopsy was performed on all patients within 5 hours of the ultrasound scans for assessment of liver fat and fibrosis content. All histological analysis was conducted by an experienced histopathologist at the same time under consistent conditions to minimise subjective variation. 1st order statistical features, the slope of ultrasound beam attenuation [alpha], and individual echo features generated using an original region-growing segmentation technique, were extracted from images for analysis. MD and ANN classifiers correctly identified all subjects as normal or diseased. 1st order median image intensity was selected by MD as the most discriminating feature for detection of disease, and ANN identified consistent 1st order features. MD identified [alpha] for estimating liver fat content (r = 0.89, p = 0.001). Using individual echo features, ANN identified the fibrosis severity of 73% of patient images (r[sub]s = 0.70, p < 0.005) and 79% of cases (r[sub]s = 0.90, p = 0.001), with 88% of images and 96% of cases correct within 1 category. Features selected for fibrosis estimation, echo count and mean echo area, are largely independent of ultrasound beam attenuation. An original echo morphology characterisation technique is presented.