Cycle-consistent adversarial network with polyphase U-Nets for liver lesion segmentation

Boah Kim, Jong Chul Ye

Abstract: We proposed a novel deep learning algorithm for liver lesion segmentation using a cycle-consistent generative adversarial network (cycleGAN) architecture. In order to overcome the mode collapsing phenomenon from many-to-one mapping nature of segmentation, our method discovers relationships between the computed tomography (CT) images and segmentation-augmented CT images through a cyclical constraint. Moreover, to retain the accurate boundary information, we employ an improved U-Net architecture called the polyphase U-Net as a generator, inspired by the recent theory of deep convolutional framelets. The performance improvement by the proposed method was evaluated on the Liver Tumor Segmentation Challenge 2017 datasets.

Author Affiliation: Korea Advanced Institute of Science and Technology

Keywords: liver lesion segmentation, cycleGAN, polyphase U-Net