Research Of Brain Image Segmentation Technology Based On Deep Learning

With the development of industrialization,people’s living environment and lifestyle changes,brain tumors and cerebrovascular diseases are gradually becoming the leading cause of death in China,and early diagnosis and treatment can help reduce the death rate of patients.The diagnosis,surgical planning and postoperative therapeutic effect evaluation of brain tumors and cerebrovascular diseases often rely on medical imaging technologies such as Magnetic Resonance Imaging(MRI),Computed Tomography(CT).Due to the large amount of image data,the manual extraction of the region of interest is time-consuming and laborious.So the automatic extraction of cerebrovascular and brain tumor regions from the image is the basis of the computer-aided diagnosis and treatment system.Based on the characteristics of medical imaging,this dissertation focuses on the study of cerebrovascular segmentation based on neural network and brain tumor segmentation based on neural network,and proposes a variety of solutions based on deep neural network from different perspectives.1.Proposed a cerebrovascular segmentation algorithm based on multi-modal convolutional neural networkThis dissertation firstly introduces the convolutional neural network to cerebrovascular segmentation and discusses in detail the influence of different network parameters on the results of cerebrovascular segmentation.On this basis,the dissertation uses the Gaussian filter and the Laplace filter to process the original image respectively,and adopts three parallel convolutional neural network to segment these three modal images.Finally,the results of these three modal images are fused to get the final segmentation result.The algorithm can reduce the influence of noise in the image and the influence of non-vascular tissue.Experimental results show that this algorithm is better than the traditional cerebrovascular segmentation algorithm,and multi-mode convolutional can effectively improve the accuracy of cerebrovascular segmentation of network.2.Proposed a brain tumor segmentation algorithm based on generative adversarial networkA brain tumor segmentation algorithm based on the generative adversarial network is proposed in this dissertation.The algorithm consists of two networks,namely the generative network and the discriminative network.The generative network consists of two output branches,namely the normal discriminative branch and the generative branch.The discriminative branch is used to obtain the independent segmentation results of each pixel,while the generating branch together with the discriminative network is used to obtain the higher-order information between the results of different pixels.The final results are achieved by fusing these two results in end-to-end way.In addition,this dissertation proposes multi-perspective patchGAN to obtain high-level information.Experimental results show that this algorithm is superior to the existing optimization method based on full connected conditional random field.3.Proposed a brain tumor segmentation algorithm based on the same feature spaceIn order to solve the problem that the existing brain tumor segmentation algorithm based on convolutional neural network can’t deal with the problem of modal loss and modal increase,this dissertation proposes a brain tumor segmentation algorithm based on the same feature space.The algorithm maps images of different modals to the same feature space through the same convolutional neural network.These features of different modals are fused and classified to achieve the segmentation of brain tumors.Experimental results show that the proposed algorithm can effectively deal with the modal loss and increase.4.Proposed a brain tumor segmentation algorithm based on symmetryExisting brain tumor segmentation algorithms based on convolutional neural network ignore the existing medical knowledge.The common medical knowledge is that the left and right asymmetry area reflects the existence of brain tumor.So,this dissertation proposes a brain tumor segmentation network based on symmetry.This network adopts the features extracted from the convolutional neural network to calculate the left and right similarity map of the brain image,and uses this map as the attention to make the network pay more attention to the asymmetry region.Experimental results show that the performance of the network in tumor segmentation can be effectively improved by the incorporation of symmetry into the network.5.Proposed a two-stage segmentation algorithm for brain tumors from coarse to fineThis dissertation presents a two-stage segmentation algorithm for brain tumors from coarse to fine.The algorithm consists of two networks,one is a rough segmentation network composed of conditional generative adversarial network,which is used to obtain rough segmentation results of each tumor tissue.The other is a fine network based on attention.The algorithm takes the rough segmentation results as the attention of feature layers in fine network.In addition,in order to reduce the impact of data imbalance,the fine-segmentation network adopts the cross-entropy loss function with mask and the double-branch output to balance the brain tumor tissue and the normal brain tissue.Experimental results show that the algorithm has a good performance in the segmentation of brain tumors.With the development of computer-aided diagnosis and treatment system,the accurate segmentation of cerebrovascular and brain tumors is becoming more and more important.Although some studies have been carried out in this dissertation on the segmentation of cerebrovascular and brain tumors based on deep neural network.more attention is needed on how to extract cerebrovascular and brain tumors more efficiently and accurately.