Research On Automatic Segmentation Method Of Multimodal Image Based On Convolutional Neural Network

Multimodal brain tumor segmentation method is a technology to segment brain tumor automatically and accurately.It can be applied to the clinical diagnosis of brain tumors and can provide support for the follow-up treatment plan,patient survival prediction and other key tasks.In addition,the patient's prognosis effect would be affected by the accuracy of brain tumor segmentation results.In recent years,convolutional neural network has demonstrated its powerful image processing ability in the field of medical image segmentation.Various brain tumor segmentation methods based on convolution neural network are constructed to improve the accuracy of brain tumor segmentation.Therefore,the brain tumor segmentation methods based on convolutional neural network has achieved rapid development.However,existing network still have three problems.The mode of medical images used in the segmentation task is relatively single.The network structure can't effectively use the complementary information in multimodal medical images to enhance the ability of feature representation.The brain tumor edge segmentation results are poor.Therefore,it is necessary to develop a reliable automatic segmentation algorithm for improving the accuracy of brain tumor segmentation.In this dissertation,we began with the basic concepts and algorithms of convolution neural network.The classical convolution neural network was studied deeply.And based on classical model,the multimodal brain tumor segmentation task was further researched.The main work are as follows:(1)The thesis summarized the latest research on brain tumor segmentation,introduced the basic concepts of convolution neural network.Furthermore,the basic structure of classical semantic segmentation methods were described and multimodal brain tumor datasets were established.(2)Based on the multi-feature extraction structure,Residual X-Net(RX-Net)was proposed to optimize the network structure and improve the information utilization in multimodal medical images.The excellent performance of RX-Net in multimodal brain tumor segmentation task is attributed to three factors.Firstly,the multi-feature extraction structure can effectively extract information from multimodal medical images and avoid semantic conflict between medical images.Secondly,the feature fusion block in the skipconnection can fully integrate the semantic information in the multimodal medical images.Thirdly,different methods of up-sampling can regularize the shared encoders to improve the accuracy of segmentation.Experimental results on multimodal brain tumor dataset showed that RX-Net could accurately segment brain tumor from multimodal medical images.(3)In order to overcome the influence of noise on brain tumor segmentation task,Residual Multi-Scale Attention X-Net(RAX-Net)based on multi-scale attention block was proposed by combining RX-Net,multi-scale features and attention mechanism.The multi-scale attention block was constructed by using multi-scale features and attention mechanism.RAX-Net was built by introducing multi-scale attention block into RX-Net.The categories of pixels could be accurately determined by multi-scale attention block with different medical images.In addition,the multi-scale attention block could efficiently allocate the weight of attention,suppressed the interference of noise,and realized the rapid convergence of the network.The experimental results in multimodal brain tumor dataset showed that RAX-Net had the fastest learning speed and the best segmentation performance on multimodal brain tumor datasets.