Conditional Generative Adversarial Networks with Multi-scale Discriminators for Prostate MRI Segmentation

Abstract

Accurate prostate MR image segmentation is a necessary preprocessing stage for computer-assisted diagnostic algorithms. Convolutional neural network, as a research focus in recent years, has been proven to be powerful in computer vision field. Recently, the most effective prostate MRI segmentation technology mainly relies on full convolutional network which has been widely used in semantic segmentation task. However, it’s independent and identically distributed assumption neglect the structural regularity present in MR images and miss information between pixels. In this paper, we propose an MRI-conditional generative adversarial networks for prostate segmentation. Our adversarial training make it context aware and the use of adversarial loss functions learn high-level structural information. The network consist of a generator and a discriminator. The generator consists of a contraction channel and an expansion channel like U-Net. The method we proposed uses a multi-scale discriminator which consist of two discriminators with the same structure but different input sizes. The objective function has two parts: one is the adversarial loss, the other is feature matching loss which stabilizes the training and get better convergence. The experiment show that our network can accurately segment the prostate MRI and outperforms most existing methods.