Discrepancy Steered Conditional Adversarial Network For Deep Feature Based Malignancy Characterization of Hepatocellular Carcinoma

Preoperative knowledge of the malignancy of hepatocellular carcinoma (HCC) based on medical images plays a significant role in deciding therapy strategies and patient management in clinical prac-tice. Deep learning with Convolutional Neural Network (CNN) has shown high diagnostic performance for lesion characterization with medical images. However, it is very challenging to train robust deep learning system for lesion characterization, especially for HCC, because there are often limited samples in clinical practice. In this work, we propose an efficient end-to-end framework that flexibly combines Conditional Adversarial network (CAN) and CNN to characterize the malignancy of HCC. Specifically, we introduce a similarity discriminative network to make the CAN efficiently generate more discrepant samples and devise hybrid loss functions to embed the proposed similarity discriminative network to the end-to-end framework with CAN and CNN. Experimental results of 115 clinical HCCs with pathologically confirmed malignancy demonstrate that the proposed end-to-end framework with similarity discriminative network can significantly improve the performance of deep feature based malignancy characterization of HCC and remarkably reduce the risk of overfitting with limited samples for the deep learning model in clinical practice.