Research On Fast Segmentation Of White Matter Fibers Based On Deep Convolutional Networks

The white matter of the brain is composed of nerve fibers with different functions in the central nervous system,which enables doctors to understand the brain development,the aging,and the disease.The segmentation of the white matter fiber enables detailed analysis on the white matter fiber bundles,that benefits to characterize the healthy brains and recognize the abnormally shaped area.In the meantime,the segmentation of the white matter fiber can simulate the connectivity of the brain white matter better.The traditional method of the automatic white matter fiber segmentation involves a series of steps such as tractography,image registration and parcellation.Pipelines derives from this way are too complex and computationally intensive to finetuning.In recent years,the performance of the deep learning in image processing have been widely recognized by publics.Among them,some extending applications related to deep convolutional neural networks have also begun to appear in the field of the medical image segmentation.The paper utilizes the deep convolutional neural networks to segment the white matter fibers of the brain,which can reduce the traces of manual operations,skip the tedious operation links,and achieve faster,complete,and accurate segmentation.This paper firstly lists some applications on the deep convolutional neural network in white matter fiber segmentation.Then,giving a brief introduction on the origin of deep convolutional neural networks,and analyzes the network structure detailly.Based on the characteristics of the diffusion weighted MRI of the white matter fibers,a segmentation model based on deep convolutional networks is proposed to realize a fast segmentation of the white matter fibers.In addition,this paper improves the white matter fiber model based on the deep convolutional network from two dimensions of segmentation efficiency and segmentation accuracy.Some experiments are designed to compare different white matter fiber segmentation models.The main research contents of this paper are summarized as the following two parts:(1)Construct a fast segmentation model of white matter fibers based on deep convolutional neural networks: In this paper,HCP dataset has been pre-processed to a uniform suitable size such as distortion correction,motion correction,and registration to MNI space.Then,in three different directions(coronal,axial,sagittal)of each voxel,the principal directions were extracted by using the multi-shell multi-tissue constrained spherical deconvolution(CSD)and peak extraction available in MRtrix with a maximum number of three peaks per voxel respectively.Randomly sample 2D slices in three different directions without tractography,image registration or parcellation,which is based on the encoder-decoder full convolutional neural network.Through the peak of the fiber direction distribution function(fODF),the rapid segmentation of white matter fibers in the brain was achieved.(2)Improved white matter fiber segmentation model of deep convolutional neural network: This paper improves the white matter fiber segmentation model based on the deep convolutional neural network,and constructs network that combined with atrous convolution to improve the model's segmentation efficiency.The network structure of the white matter fiber segmentation model is further improved based on the atrous spatial pyramid pooling and dense connectivity.The result demonstrates that the segmentation accuracy of the model is significantly improved.