Semi-supervised Cardiac MRI Segmentation Based on Generative Adversarial Network and Variational Auto-Encoder
Shaojie Li, Yifan Zhang, Xuan Yang
Abstract
Segmentation of the heart structure plays an essential role in cardiac diseases diagnosis and treatment planning. This paper proposes a semi-supervised learning network for multi-objective segmentation of cardiac MRI data, which comprises a U-Net as encoder and a conditional generative adversarial network (GAN) as the decoder. The process of segmentation takes the predictive result as latent variables and aims to estimate the distribution of the latent. We pre-align cardiac images to each other in an affine way and pre-train the GAN using a small amount of annotated data. A loss function focusing on the region of interest (ROI) is proposed. Experimental results on MICCAI 2020 Multi-Centre, Multi-Vendor & MultiDisease Cardiac Image Segmentation Challenge (M & Ms) dataset show that the performance of our semi-supervised learning network not only reaches or exceeds the supervised methods in multi-object segmentation but also has the generalization ability of supervised learning-based networks.