Research On Technologies Of Brain Tumor Image Segmentation Based On Deep Learning And Implemention

The segmentation of brain tumor MRImages is vital step of a success tumor diagnosis and treatment planning.The medical image segmentation techniques based on deep learning models,represented by 3D U-net,could achieves the goal of segmentation,though,it usually comes with several issues: Firstly,the features of brain tumor images are not fully utilized,the accuracy of segmentation could still be improved;Secondly,high requirement of computing resource is inevitable,and U-net has potential structural problems.Based on the reason above,the work of this thesis aiming to achieving higher accuracy and moderate 3D deep learning models’ huge demand on computing resource.Through introducing attention blocks and the idea of group convolution to U-net,presented model achieves the task of segmentation with lower computing resource.The main works of this thesis are as followed:(1)A dual-way attention mechanism is proposed.This attention mechanism is composed of spatial attention branch and channel attention branch.When low-level features transmit through skip connections,the dual-way attention block plays the roles of buffer,the channel attention branch squeezes the spatial features in one voxel and the spatial attention branch could focus on feature-rich regions of input voxels though connected CCA modules.Two branches work together to alleviate the segment gap between low-and high-level feature maps and improve the utilization of low-level feature map.U-net embedded with dual-way attention block improves the dice score by 2%.(2)Multi-path dilated convolution block is proposed.When images input,this block could reduce the parameters consumption saving computing resource through grouped dilated convolution.Meantime,convolution paths with different dilated rate could capture different features of input images and offering better performance on segmentation.A residual branch is introduced to the block to avoid the potential issues of deep learning training.U-net embedded this block saves 2M parameters in total comparing to the 3D U-net and about 18% GPU memory as well.(3)Based on the integration of the two modules above,this thesis proposed a MP(multi-path)U-net,and conducts experiments on it using Bra TS2020 dataset,and the results reveals that proposed model lower parameter consumption by 10% comparing to3 D U-net,and dice score reaches 76.61%,89.05%,78.42% on ET,WT and TC region respectively.(4)Based on demand analysis of models’ target user,a brain tumor segmentation assistance system is developed using springboot framework.This further attempts the application of medical image segmentation networks.