Brain Tumor Segmentation Using Discriminator Loss

Abstract

The emerging field of Computer Vision has found enormous applications in our day-to-day lives and Medical Image Processing is one of the most prominent fields among them. Brain Tumor Segmentation is an important and challenging task because of the variety in shapes, sizes and texture content of the various types of brain tumors. Specifically, MICCAI BraTS organizes Brain Tumor Segmentation challenge every year. Since the evolution of CNNs it has obtained state-of-the-art results in the majority of computer vision related tasks. On BraTS Challenge 2017, an assemble average of various CNN models (EMMA) holds the state-of-the-art performance. In this paper, we have proposed a model inspired by the classic Generative Adversarial Network (GAN). The proposed network has two models namely, Generator or Segmentor which generates label map of the input image and a Discriminator which helps the Generator model for an optimum solution by taking into account both short as well as long-distance spatial correlations between pixels with the help of a novel multi-scale loss function. The proposed architecture has three GANs in a cascaded fashion, each for Whole Tumor, Tumor Core and Enhancing Tumor, where the former network helps in effective reduction of false positives for the later networks. Our method also employs a multi-scale loss function derived from intermediate layers of Discriminator rather than depending just on a final layer cross-entropy loss. A mutli-scale loss function also reduces unnecessary smoothing on contours. The proposed method performed comparatively better than the state-of-the-art techniques, having Dice scores of 0.820, 0.874 and 0.783 for Enhancing Tumor, Whole Tumor and Tumor Core respectively.