Research On The Segmentation Network Of White Matter Nerve Fiber Bundles In Diffusion Magnetic Resonance Image

Diffusion magnetic resonance imaging(DMRI)is a non-invasive technique that can be used to study the microstructural properties of brain white matter fiber bundles in vivo.DMRI has played an important role in the study of a variety of neurological diseases,such as stroke,Parkinson’s disease,Alzheimer’s disease,and multiple sclerosis.By providing a unique,non-invasive view,DMRI enables researchers to study the impact of brain white matter microstructure on neurological diseases in greater depth,while also contributing to the early diagnosis of these diseases.Deep learning,which automatically learns the features contained in data,has been remarkably successful in areas such as computer vision and medical image analysis.Brain white matter fiber bundle segmentation is an important task in brain microstructure analysis,which can accurately portray the nerve fibers in the human brain and assist doctors in making accurate diagnoses and treatments.However,due to the large number and complex shape of white matter fiber tracts in the brain,traditional manual feature extraction methods are often difficult to achieve ideal segmentation results.Therefore,the introduction of deep learning technology provides a new idea and method for brain white matter fiber bundle segmentation,which can automatically extract features and learn to the complex features of brain white matter fiber bundles to achieve more accurate and robust segmentation.Despite the tremendous progress in deep learning-based approaches for brain white matter nerve bundle segmentation,the following problems remain:(1)Existing work typically uses simple full convolutional networks(FCNs),and while FCNs are effective in capturing local contextual information,they cannot fully localize remote dependencies.However,for accurate fiber bundle segmentation,global contextual information provides valuable high-level semantic cues that are key to identifying fiber bundles.(2)Due to structural factors such as complex morphology,dense distribution,crossover and bifurcation of brain nerve fiber bundles,existing methods have limitations in the completeness and accuracy of segmentation.To address the mentioned problems,an effective full convolutional neural network(Dense Criss-Cross U-Net,DC~2U-Net)based on dense cross-attention is proposed in this paper to optimize the shortcomings of existing methods.Also,this paper considers the effect of processing data from different acquisition protocols on the generalizability of the model,and the optimization is as follows:Firstly,due to the high-dimensional nature of the raw DMRI data,its direct use as network input leads to a large amount of additional computation and graphics card resource usage.Also,the generalizability of the network is limited by using DMRI data obtained with different sampling schemes in q-space.To alleviate these problems,this paper adopts a multi-shell multi-organization constrained spherical deconvolution(CSD)method to process the data,extracting the three main fiber directions of each voxel to represent the original data in a concise manner while keeping the loss of original data accuracy to a minimum.Moreover,during the training process,two-dimensional slices of one direction are randomly selected to predict the probability map of each fiber bundle,and the probability maps of different directions are combined during the testing process to produce the final three-dimensional segmentation results.Secondly,to improve the brain white matter fiber bundle segmentation accuracy,a new fiber bundle segmentation algorithm,DC~2U-Net,is proposed in this paper.This algorithm innovatively integrates an efficient Dense Cross-Cross Attention(DCCA)component into the U-Net model.DCCA is able to capture richer global contextual information and uses this generated information to improve the accuracy of bundle segmentation.In addition,it is difficult to correctly represent white matter fiber bundles by simply connecting encoder features and decoder features because they are closely adjacent to each other and the boundary regions are difficult to segment accurately.Therefore,this paper proposes to use a low-level feature activation map(Low-Level Feature Activation Map,T＿FAM)to assist feature learning in order to better recover the details and edge information of the feature map during upsampling.Finally,to better train the model,this paper also proposes the use of a loss function(Depth Supervision,DS)designed based on a depth supervision strategy,which provides more effective network training.This paper is evaluated using a public Human Connectome Project(HCP)high-quality dataset as well as clinical data from low-quality cocaine addicted patients.Quantitative and qualitative results indicate that the proposed method is effective in improving the accuracy of white matter fiber bundle segmentation,especially for topologically complex fiber bundles.The analysis of ablative experiments also demonstrated the positive effect of each proposed module on the improvement of experimental results.