Research On Deep Learning Based Mr Brain Image Segmentation And Unsupervised Multimodal Registration

Stroke is the most common cerebrovascular disease and is one of the most common causes of death and disability worldwide.It is of potential research significance to mark specific brain regions.These markers can potentially provide additional information,including within behavioral assessments,to predict recovery in the patient’s functional area(eg,exercise,speech)and the likelihood of responding to treatment.In addition,the images obtained by combining multiple imaging modes can provide complementary multimodal information,but the direct use of unregistered multimodal images may not only limit the performance of the computer automatic segmentation method,but also increase the burden on clinicians.Multi-modal registration of medical images can help people to establish various relationships of the same organization,and promote complementary information to make more accurate judgments.However,1)manual specific brain region labeling may no longer be suitable for actual needs;2)currently multimodal registration-based methods are not mature.This paper first focuses on the current automatic segmentation of MR imaging strokes,and proposes two main solutions for the insufficient utilization of multi-scale information and the inherent limitations of two-dimensional and three-dimensional networks,called CLCI-Net and D-UNet.First,1)CLCI-Net is proposed to cope with the insufficient use of multi-scale features and contextual information in current research.We have developed a new cross-layer feature fusion strategy(CLF)to take full advantage of different levels of features;CLF is further used to extend ASPP to alleviate the challenges of different lesion scales;Conv LSTM is used to replace traditional channel stacking operations,Capturing more fine-grained structures to distinguish visually similar tissues.The proposed method was evaluated on the open source dataset ATLAS and compared with the five most popular methods,ranking first with a DSC score of 57.8%.Secondly,2)After CLCI-Net has achieved good results,we believe that the 2D structure ignores the 3D information of MR medical images.However,using pure 3D CNN requires too much computing resources and is not conducive to clinical deployment.To this end,we propose a dimensional fusion architecture,D-UNet,to address this challenge.This architecture innovatively combines 2D and 3D convolutions during the coding phase;we also propose a method called enhanced mixed loss(EML)New loss function,which can significantly accelerate the convergence speed of the network.The proposed architecture DSC achieves 59.2% on the ATLAS dataset,has better segmentation performance than 2D networks,and significantly reduces the parameters required for 3D networks.Then,in order to solve the clinical multi-modal registration problem,a new deep learning-based unsupervised registration method called PGDC-Net isproposed.This method successfully completes the registration on MR FLAIR and DWI sequence images through affine transformation model guidance and dual consistency recovery strategy.We evaluated the fitting degree of the goldstandard based on two modal outlines after registration,and its DSC score was77.7%,showing promising results for multimodal fusion.In addition,once the registration network training is completed,it can be easily migrated to other modalities.As far as I know,this method is the first registration work on FLAIR and DWI modalities.In summary,this paper proposes new stroke segmentation and unsupervised multimodal registration methods to facilitate the analysis of MRI stroke images.The effectiveness of the method was verified on the open source data set ‘ATLAS’and the collected data of Guizhou People’s Hospital.The results show that our proposed methods have achieved state of art performance.In addition,the registered multi-modality images have the potential to improve the performance of stroke segmentation.By fusing multi-modality images,it is hoped that more robust segmentation results will be obtained in the future.These results fully demonstrate the prospect of our method being competitive in stroke analysis tasks.