3D deep feature fusion in contrast-enhanced MR for malignancy characterization of hepatocellular carcinoma

Abstract

The malignancy of hepatocellular carcinoma (HCC) is of great significance to prognosis. Recently, deep feature in the arterial phase of Contrast-enhanced MR has been shown to be superior to texture features for malignancy characterization of HCCs. However, such extracted deep feature of HCCs was limited to be 2D, which apparently disregards the contextual information of the third dimension in volumes. Furthermore, only arterial phase was used for deep feature extraction, ignoring the impact of other phases in Contrast-enhanced MR for malignancy characterization. To further enhance volumetric spatial information and take full advantage of multi-phasic information in Contrast-enhanced MR for malignancy characterization of HCC, this study proposes a systematic method to automatically extract 3D deep feature of HCCs by utilizing 3D convolution neural network (CNN), followed by deep feature fusion from multiple phases of Contrast-enhanced MR for more accurate malignancy characterization. Experimental results on 46 clinical patients with Contrast-enhanced MR images show several intriguing conclusions as follows: (1) 3D deep feature outperforms 2D deep feature for malignancy characterization; (2) fusion of 3D deep features from multiple phases of Contrast-enhanced MR yields better performance for malignancy characterization; (3) 3D deep features in Arterial phase is best, followed by portal vein phase and pre-contrast phase for malignancy characterization.