Semi-Supervised Learning Using Co-Generative Adversarial Network (Co-GAN) for Medical Image Segmentation

Medical image analysis has experienced different stages of development, especially with the emergence of deep learning. However, acquiring large-scale, high-quality labeled data to train a deep learning model takes time and effort. This paper proposes a semi-supervised learning method for medical image segmentation using limited labeled data and large-scale unlabeled data. Inspired by the classic Generative Adversarial Network (GAN) and co-training strategy, we proposed a new Co-GAN framework to implement medical image segmentation. The proposed Co-GAN comprises two generators and one discriminator, in which two generators can provide mutual segmentation information to each other. Through adversarial training between generators and discriminators, Co-GAN achieved higher segmentation accuracy. The dataset used was the hippocampus in Medical Segmentation Decathlon (MSD). There were 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; and 200 labeled slices/3,199 unlabeled slices. Three experiments were conducted for each data set: fully supervised learning based on a generator network using only labeled data (F-Generator), semi-supervised learning based on GAN (Semi-GAN), and semi-supervised learning based on Co-GAN. The experiments showed that Co-GAN improved the segmentation accuracy by (1.9%, 2.6%, 1.1%, and 0.1%) compared to F-Generator and (2.2%, 0.8%, 0.5%, 0.7%) to Semi-GAN.