MOS-GAN: A U-Net++ based GAN for multi-organ segmentation

Abstract

Radiotherapy, a standard and common form of cancer treatment, is used in about half of all cancer cases. It is critical to segment organs at risk (OARs) accurately and promptly for radiation treatment planning to be efficient and high-quality. In this paper, an adversarial training strategy is used in our deep learning network. This model is called the Multiple Organ Segmentation Generative Adversarial Network (MOS-GAN). The U-Net++ is our generator to create a multi-organ segmentation image by learning the end-to-end mapping from CT images to OARs segmentation. A convolutional neural network (CNN) acts as the discriminator to identify the difference between the manual contour and the segmented images created by the generator. A 2.5D method is employed by us to extract the spatial information among slices. We applied the proposed method to segment the five organs from the 2017 AAPM segmentation challenge dataset. Segmentation of the bilateral lungs, heart, and spinal cord were successfully achieved, as well as the outline of the esophagus. In terms of dice similarity coefficients, the aforementioned five OARs have average values of 0.97, 0.97, 0.92, 0.91 and 0.77, respectively.