Deep learning classification of focal liver lesions with contrast-enhanced ultrasound from arterial phase recordings

Abstract

Contrast-enhanced ultrasound (CEUS) has been known as a safe, robust, and cost-effective image modality to diagnose an early sign of hepatocellular carcinoma (HCC). The enhancement patterns on CEUS are composed of arterial, portal venous, and late phases, where the hepatic arterial phase provides information on the degree and pattern of vascularity, and the portal venous and late phases provide important information on the differentiation between benign and malignant liver lesions. The enhancement patterns of HCC on CEUS are hyper-enhanced in the arterial phase. Therefore, we propose learning-based frameworks to differentiate between hepatocellular carcinoma (HCC) and focal nodular hyperplasia (FNH) during the arterial phase. We design artificial neural networks to learn the change of characteristics over time for the differentiation of HCC from FNH. We had gathered CEUS videos during the arterial phase for 4 years in Samsung Medical Center (SMC) and picked out only small hepatic lesions under 3 centimeters. From these datasets, the proposed novel 3D-CNN and CNN-LSTM networks show accuracy rates of 100% and 98% for 10-fold and 5-fold cross-validations. In the end, the proposed models are proved to be feasible for accurate automatic classification between HCC and FNH in livers.