Leveraging Unlabeled Data Using Semi-Supervised Generative Adversarial Network for Medical Image Segmentation

Abstract

Medical image analysis has experienced different stages of development, especially with the emergence of deep learning. However, it is difficult to acquire large-scale, highquality labeled data to train the model when using deep learning. This paper proposes a semi-supervised learning method to achieve medical image segmentation using limited labeled data and large-scale unlabeled data. Inspired by the classic Generative Adversarial Network (GAN), we proposed semi-supervised learning based on GAN (semi-GAN) to implement medical image segmentation. In the proposed semiGAN, adversarial training between the generator and discriminator has achieved higher segmentation accuracy. The dataset used was hippocampus data in Medical Segmentation Decathlon (MSD), and there are four training data settings: 25 labeled slices/3,374 unlabeled slices; 50 labeled slices/3,349 unlabeled slices; 100 labeled slices/3,299 unlabeled slices; 200 labeled slices/3,199 unlabeled slices. For each data setting, there are two experiments conducted: fully-supervised learning based on a generator network using only labeled data (F-Generator), and semi-GAN. The experiments showed that semi-GAN can improve segmentation accuracy by an average of 0.4% using unlabeled data compared to F-Generator using labeled data. Further study will be conducted to improve the semi-GAN architecture.