Support vector machine (SVM) based liver classification: fibrosis, steatosis, and inflammation

An SVM based liver classifier was developed to differentiate liver conditions, including normal, fibrosis with low fat, fibrosis with high fat, and inflammation. An in-vivo study was performed with 35 rats under normal conditions or after carbon tetrachloride (CCl4) or concanavalin A (ConA) dosing to induce fibrosis with varying degrees of steatosis, and inflammation, respectively. These livers were imaged in-vivo by an ultrasound, and approximately 30 frames for each rat were acquired. Therefore, a total of 998 ultrasound images were analyzed and used for training a SVM classifier. Each image has three measured parameters: H-scan scattering classification, estimated attenuation coefficient, and B-scan intensity. These parameters were assigned as inputs to the SVM. A liver diagnosis system based on the SVM and H-scan was produced. The clusters representing each state of liver are provided in two- and three- parameter space. From these, the SVM generates decision planes to classify the liver conditions. The classification accuracy is 92.2% with the three features. Therefore, these results provide the beginning of a coherent framework for determining the scattering signatures or clustering in multi-parametric space, of the normal liver compared with diseased livers.