Brain Tumor Segmentation Method Based On Deep Learning And Multi-modal MRI Images

The multi-modal MRI brain tumor segmentation aims at segmenting the “whole tumor”(WT),the “tumor core”(TC)and the “enhancing tumor”(ET)from the image by taking full advantages of the characteristics of each mode in the MRI images.Computeraided diagnosis of multi-modal MRI brain tumor image segmentation have always been an important topic in the field of medical image processing.However,due to grayscale similarity between brain tissues and a lot of differences between patients,it is difficult to deal with the segmentation of brain tumor using traditional algorithms.Aiming at solving this problem,this thesis have done some researches as follows:1.Based on the fully convolutional neural network(FCN),this thesis constructed a model which is suitable for the brain tumor segmentation tasks.The model is trained endto-end on MRI two-dimensional slices.This thesis provides a scheme for pre-processing which is exploiting N4 ITK algorithm to correct the bias field and using intensity normalization algorithm to balance the grayscale in different cases.To solve the problem of unbalanced categories in brain tumor images,the Dice loss function was introduced into the network.Meanwhile,this thesis proposed a parallel Dice loss function structure to further improve the effect of sub-regions segmentation.A cascade network model based on FCN was proposed,and it achieved good prediction results in the segmentation of sub-regions in the BraTs2017 dataset.2.The MRI image data is essentially a three-dimensional image.Extracting the twodimensional slices from MRI images will lose the inherent structural information.To explore the reliability of applying 3D convolutional networks to segment the brain tumor images,this thesis built a network based on C3 D and FCN.The residual structure and dense connection unit have also been introduced to accelerate the convergence of the network.The detailed information of the images is preserved by using the context connection module.In the segmentation training process,this thesis concatenated the multi-modal MRI images and combined the multi-scale information to improve accuracy of sub-regions segmentation.