Research On Glioma Segmentation Strategy Based On Deep Learning

Glioma is a common primary brain tumor that seriousl y affects human health.In the field of glioma treatment,magnetic resonance imaging(MR I)has become an important tool for the identification and classification of brain lesions due to its strong contrast,high resolution and sensitivit y to lesions.By segmenting the glioma,t he doctor can locate and obtain the size of the glioma,and then formulate an accurate treatment plan.However,the complex lesion characteristics and abstract irregular expression of gliomas make precise segmentation a challenge.The traditional segmentat ion method not onl y relies too much on the clinical experience of doctors,but also consume time and effort.For different doctors,the MRI images of the same patient may draw different conclusions.How to effectivel y obtain key lesion features from glioma image information and achieve precise semantic segmentation of the target area is an urgent problem that needs to be solved in the field of medical imaging and deep learning.In order to further improve the accuracy of glioma segmentation,this paper studies the effects of data augmentation,algorithm structure and generative adversarial networks(GAN)on glioma segmentation.The innovation of this paper is as follows:(1)To solve the problem of how data augmentation affects the result of glioma segmentation,this paper defines the ratio of the number of pictures generated by data augmentation to the total number of training sets in the data preprocessing stage as the data augmentation ratio(DAR).Through the traditional data augmentation,data a ugmentation based on single data set and data augmentation based on image translation,the effects of different algorithms on glioma segmentation results under different DARs are studied.Based on this,a glioma segmentati on strategy based on optimal DAR is proposed.(2)Aiming at the problem that neural networks are easil y degraded,this paper introduces a residual-dense block(RDB)in the segmentation network based on the encoder-decoder structure.The RDB extracts the detailed features to the greatest extent and concatenates the features of different levels.In order to extract more abundant target feature information,this paper introduces a dual attention mechanism.During the downsampling process,the feature channel attention(CA)mechanism is used to adaptivel y learn the importance of each channel at the same scale.During the upsampling process,the spatial attention(SA)mechanism is used to obtain the weight of the target feature on the same feature map.This mechanism can deeply extract the usef ul information related to the target and integrate the information of the two dimensions.Finally,a new segmented network(generated network)RD2A-Net is proposed.(3)In order to further refine the segmentation effect of gliomas,the idea of GAN is intro duced into the glioma segmentation framework,and a semantic segmentation network based on the generator-discriminator structure is constructed.This method takes the segmentation network proposed in this paper that introduces RDB and dual attention mechan ism as a generator,and constructs a discriminator with the abilit y to extract the deep feature information of the image separatel y.Finall y,a segmentation network RD2A-GAN based on generative adversarial network is proposed,and the effectiveness of the method is proved through experiments.