ISDNet: Importance Guided Semi-supervised Adversarial Learning for Medical Image Segmentation

Recent deep neural networks have achieved great success in medical image segmentation. However, massive labeled training data should be provided during network training, which is time consuming with intensive labor work and even requires expertise knowledge. To address such challenge, inspired by typical GANs, we propose a novel end-to-end semi-supervised adversarial learning framework for medical image segmentation, called "Importance guided Semi-supervised Deep Networks" (ISDNet). While most existing works based on GANs use a classifier discriminator to achieve adversarial learning, we combine a fully convolutional discriminator and a classifier discriminator to fulfill better adversarial learning and self-taught learning. Specifically, we propose an importance weight network combined with our FCN-based confidence network, which can assist segmentation network to learn better local and global information. Extensive experiments are conducted on the LASC 2013 and the LiTS 2017 datasets to demonstrate the effectiveness of our approach.