Research On Multimodal Brain Image Lesion Segmentation Based On Neural Network

Artificial Intelligence(AI)and deep learning originated from the knowledge of human brain and the intention of simulating the running process of human brain,which achieve.the goal of digital intelligence.In recent years,with the development of the theory and practice of deep learning technology,research interest has gradually shifted from mathematical theories such as statistics and optimization to mathematical modeling methods based on data.Affected by this,Computer Vision(CV)has achieved great process.In the three basic visual tasks,image classification,object detection and image segmentation,deep-learningbased methods surpass the traditional digital image processing algorithm.Medical image processing,one of the applications of CV,is also benefit from CV and march into intelligent medical image analysis.However,human brain is extremely complex and delicate,which makes the diagnosis base on three-dimensional(3D)brain images difficult and labor exhausted.Long time work exhausts expert,leading to the increase of misdiagnosis rate.Therefore,it is of crucial importance to use deep learning technology to construct system for automatic or semi-automatic auxiliary diagnosis.Multimodality,i.e.,multiple images reflecting different physical information from multiple aspect for the same region and same subject,is one of the important characteristics of 3D medical image.Such as Magnetic Resonance Imaging(MRI),Positron Emission Tomography(PET),etc.In clinical application,the comprehensive analysis of the multimodal images can effectively locate lesions and diagnose their types,which lays a medical theoretical foundation for the use of deep learning to achieve the automatic lesion segmentation and classification.Therefore,based on the multimodal semantics of brain image and the composition characteristics of 3D brain image,this thesis is mainly focus on the feature extraction for lesion segmentation,background noise filtering,relevance information capture in depth axis and reuse of label information,which realizes the automatic and accurate segmentation of lesions,and provides the category of lesions.Specifically,this paper mainly includes the following two parts:Firstly,to resolve the problem of inaccurate segmentation caused by complex background noise and strong interference,the information fusion module between multimodal features is studied.Based on the semantic information of different modes of 3D brain images,research designs a mask-guided multimodal feature complementary fusion module to filter the features of hidden layer in the model,which is motivated by the visual expression of the image,so as to solve the problem of False-Positive(FP)prediction.Then,based on the idea of Co-Related Attention,research introduces a grouped multimodal feature communication module.While retaining the class specific features,it enhances the information exchange and fusion between multimodal features,captures the relationship between features,and provides more accurate information related to the lesion region for the model.Finally,combined with the structure of autoencoder and the Dice Loss and Focal Loss,modality feature fusion U-Net(MFU-Net)for 3D multimodal brain image segmentation is proposed.Then,considering 3D brain image provides correlation information more than spatial information in depth axis,the voxel continuity along depth changes due to thin-scan and thick-scan,resulting lower continuity of information in depth compared with adjacent pixels in 2D space,which means anisotropy.In order to avoid the negative impact caused by this feature and capture the correlation of lesions in depth dimension more effectively,based on the previous part,research first proposes imageslice-feature aligned model structure,which transforms 3D segmentation problem into sequence segmentation problem.Then,in order to enhance the ability to capture the dynamic change of lesion,the k-neighbor contrast perception module and the corresponding supervised label generation method are proposed to explicitly promote the model to learn the change information along the depth direction.At last,based on the design shown above,Feature-Align Relevance Enhanced U-Net(FAREUNet)is proposed.The experiments in this research confirms that the multimodal feature fusion module,image-slice-feature aligned model structure and kneighbor contrast perception module designed are all effective.The modules leverage the semantic characteristics,depth correlation and morphological change of 3D medical brain images.The context features complementary module with mask guidance plays crucial role in models.This research provides a reliable framework for the multimodal brain images fusion in the hidden layer of neural network,proposes a new method for feature correlation extraction,and construct an effective and high-performance lesion segmentation model under this framework.