GAN‐segNet: A deep generative adversarial segmentation network for brain tumor semantic segmentation

Abstract In this study, we present a novel automatic segmentation method using a neural network model named GAN‐segNet, which can not only identify brain tumors from MRI images but also accurately delineate intratumor regions. Since brain tumors with varying shapes and sizes can appear anywhere in the brain and image quality and contrast of MRI could be inadequate, automatic segmentation remains challenging despite its importance in the clinical workflow. The proposed GAN‐segNet is an innovative modification of the Generative Adversarial Network (GAN) and can efficiently and accurately segment brain tumors. One key innovation of our GAN model is an autoencoder learning representation of input data that were added to the generative network of the above‐mentioned GAN. By doing so, information extracted through convolution operations can be meaningfully regularized. As a result, the scales of extracted features can be controlled by the added autoencoder to preserve detail. Additionally, we propose an innovative loss function based on the concept of focal loss to effectively mitigate the impact of label imbalance. The above‐mentioned combination enables the proposed GAN‐segNet model to improve the segmentation of small intratumor region(s). We demonstrate the proposed method using MRI data available from a public database, that is, Brain Tumor Segmentation Challenge 2018 database (BRATS 2018). Using the proposed GAN‐segNet model, the average Dice scores were 0.8280, 0.9022, and 0.814 for segmenting enhancing tumor core, whole tumor, and tumor core, respectively. Furthermore, positive predictive values for segmenting enhanced tumor core, whole tumor, and tumor core were 0.8496, 0.9270, and 0.8610, respectively.