Dissimilarity Corrective Generative Adversarial Network for Brain Image Segmentation

Abstract

More accurate diagnosis of brain disorders can be achieved by properly analyzing structural changes in the brain. For the quantification of change in brain structure, the segmentation task is crucial. Recently, generative adversarial networks (GAN) have been rapidly developed and used in many fields. Segmentation of medical images with these networks will greatly improve performance. However, segmentation accuracy improvement is a challenging task. In this paper, we propose a novel corrective algorithm for updating the accuracy and a novel loss function based on dissimilarity. First, we update the generator using the typical dice similarity coefficient (DSC) as a loss function only. For the next update, we use the same image as input and obtain the output; this time, we calculate dissimilarity and update the generator again. In this way, false prediction, due to the first weight update, can be updated again to minimize the dissimilarity. Our proposed algorithm can correct the weights to minimize the error. The DSC scores obtained with the proposed algorithm and the loss function are higher, and clearly outperformed the model with only DSC as the loss function for the generator.