Brain MRI Image Segmentation Based On Multi-modal Fusion Learning

The brain tumor is still a difficult disease to cure in today’s society.Glioma is caused by severe damage to many glial cells,accounting for more than half of malignant brain tumors,seriously endangering human life.Traditional segmentation methods based on machine learning cannot handle complex scenes.Brain tumors in normal brain tissue are irregular in shape and occupy a small proportion of brain tissue,so the traditional methods cannot achieve high segmentation accuracy.Now,a large number of theses have proved the effectiveness of deep learning-based brain tumor segmentation methods.At the same time,multi-modal magnetic resonance imaging(MRI)will be combined with the current imaging department to analyze brain tumors because the complementary information between multi-modes can provide doctors with more brain tumor image information.In the context of the above,on the premise that feature information between multi-modal images can be complementary,this thesis uses deep learning as a research tool to carry out the following analysis:(1)A two-dimensional segmentation algorithm based on an improved U-Net is proposed.By integrating the improved residual module into U-Net,the advantages of the residual module can alleviate the problem of gradient disappearance in model training when the number of network layers increases.At the same time,the residual module will overlay the features of different levels in the operation process;the purpose is to enrich the characteristic information of the image.At the same time,the network is also integrated into the attention module,whose function is to enhance the attention level of helpful information in the network.In contrast,unimportant information will be suppressed in the network.The improved network was tested on two-dimensional slices of the Bra TS2018 dataset.Experimental results show that compared with the original U-Net method,the improved network can obtain more accurate segmentation results of brain tumors.(2)A lightweight model based on grouped convolution is proposed for automatic segmentation of MRI images of brain tumors.MRI images are three-dimensional images.If they are converted into two-dimensional slices,part of the spatial information in the data will be lost.Therefore,three-dimensional convolution is adopted to form the network.Although 3D convolution can achieve better segmentation performance than 2D convolution,it requires higher hardware requirements.Aiming at these problems,this thesis designs a lightweight segmentation network.In this network,the number of parameters of the model is reduced by grouped convolution,the training time of the model is shortened,and the computational cost is reduced.The receptive field was enlarged by weighted dilated convolution,and the information communication between groups was enhanced by channel shuffle.The network adopts an encoder-decoder structure,and the deep semantic information will be fused with the shallow semantic information to improve the segmentation accuracy of the network.A series of comparative experiments are conducted on the Bra TS2018 dataset,and the experimental results show that the proposed model can significantly reduce the computational cost while maintaining good segmentation results.