Malignancy characterization of hepatocellular carcinoma using hybrid texture and deep features

Abstract

Malignancy of hepatocellular carcinoma (HCC) is significant to establish a therapeutic strategy preoperatively for liver cancer and is one of critical issues that influence recurrence and patient survival. Recently, quantitative texture feature of HCC in arterial phase of Contrast-enhanced MR has been shown to be promising for malignancy characterization of HCC. However, such texture feature is low-level, which is usually insufficient to capture the complicated characteristics of HCC. In this work, we propose a systematic method to automatically extract deep feature from the arterial phase of Contrast-enhanced MR using convolution neural network (CNN) in order to characterize malignancy of HCC. Specifically, we resample each 3D tumor in three orthogonal views (Axial, Coronal and Sagittal) independently to increase training sets, and train one CNN for each view to generate its corresponding deep feature. We investigate a multi-kernel feature fusion method that can fuse deep features derived from three views or fuse deep feature and texture feature in a kernel space. Our experimental results demonstrate several interesting conclusions: (1) deep feature significantly outperforms previous texture feature for malignancy characterization of HCC, (2) fusion of deep feature and texture feature yields best results for malignancy characterization of HCC.