GLM-Net: A multi-scale image segmentation network for brain abnormalities based on GLCM

Abstract

In medical image processing, robust brain Magnetic Resonance (MR) images segmentation algorithm is one of the most concerned research fields. It plays an important role in distinguishing healthy tissues from diseased tissues. Because of the complex structure and unpredictable appearance of the brain, it is a complex task to segment tissue parts from brain MR images. The current brain segmentation methods are mostly based on the deep convolution network model, which has the problem of large loss of information in the encoding and decoding process. In order to solve this problem, we propose a brain MR images tissue segmentation method, Gray Level Multiscale Network (GLM-Net), based on three-dimensional U-Net network. The input of the network is constructed of the original image and four characteristic images. The original image is enhanced by Contrast-Limited Adaptive Histogram Equalization (CLAHE) algorithm, and the four characteristic images are generated by Gray Level Co-occurrence Matrix (GLCM) method. The network fused the residual module and dilated convolution for multi-scale feature restoration in the upsampling process of the network decoder to segment the brain MR images into white matter (WM), gray matter (GM), cerebrospinal fluid (CSF) and tumor. The skip connection is used to transmit each set of feature maps generated on the encoder path to the corresponding feature map on the decoder path. Tested and trained on the BraTS 2020 dataset, the average Dice coefficients of WM, GM, CSF and tumor in the segmentation results of the model are about 0.92, 0.91, 0.92 and 0.82 respectively.