Brain Tumor Semantic Segmentation from MRI Image Using Deep Generative Adversarial Segmentation Network

摘要

The accuracy of brain tumor segmentation will be closely related to subsequent disease diagnosis, monitoring and treatment planning. In order to further improve the segmentation accuracy of multi-labels brain tumor images, we propose a Deep Generative Adversarial Network (DGAN) based on convolutional neural network. The DGAN consists of the generator and the discriminator. In our experiment, the generator is built by fully convolutional network called tumor segmentation network and it competes with the discriminator using the adversarial idea of zero-sum game. The segmentation maps outputted end-to-end by the Tumor Segmentation Network replace the commonly used patches of image to train the network, which can save lots of computer hardware resources and improve computing power. Faced with the major problem of label imbalances often encountered in medical MRI brain tumor image segmentation, we propose an innovative loss function to mitigate the impact of label imbalance on the experiment. The experimental results prove that the proposed network structure and innovative loss function are effective in improving the segmentation accuracy of brain tumor MRI images.