Brain-spinal DTI fiber bundle extraction system and method based on clustering and overlap rate denoising

Abstract

The present invention discloses a brain-spinal DTI fiber bundle extraction system and method based on clustering and overlap rate denoising, and performs preprocessing including noise reduction, Gibbs ringing correction, eddy current correction and normalization on the brain-spinal DTI image; performs normalization processing of the N4 deviation field on the individual T1 structural image, and performs joint registration between the spinal cord T1 image and the standard template according to the anatomical site segmentation; tracks the fiber bundles of interest in the brain-spinal joint for the individual; clusters the fiber bundles of interest in the individual space; transforms the mapping image of the fiber bundle in the individual space to the standard space, takes the mean of all individual images to obtain the overlap map, and then removes the noise. The present invention provides a standardized preprocessing and fiber bundle extraction and analysis process for brain-spinal synchronous DTI image data, which can stably and truly extract the fiber bundles of interest in the individual DTI image, and has significant advantages such as sufficient denoising and stable and reliable effect.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and further relates to a brain-spinal cord synchronous diffusion tensor imaging DTI (diffusion tensor imaging) extraction system and method based on clustering and overlap rate denoising in the field of medical image processing technology. The present invention can be used to perform preprocessing operations including denoising and Gibbs ringing removal on brain-spinal cord synchronous DTI images, and can also be used to extract individualized fiber bundles of interest, perform clustering and other individualized fiber bundle analysis. Background Art

[0002] Brain-spinal synchronized diffusion tensor imaging (DTI) is a human tissue image obtained by magnetic resonance imaging technology. It mainly achieves imaging by capturing the Brownian motion of water molecules in the human central nervous system. In the central nervous system, water molecules are subjected to different forces during Brownian motion, resulting in different deformations, and this deformation is significantly different in white matter, gray matter and cerebrospinal fluid. Therefore, DTI images can effectively extract white matter information in the central nervous system. The fiber bundle tracking method combined with brain-spinal synchronized DTI images can effectively construct the fiber bundles of the brain and spinal cord, thereby effectively characterizing the positioning and morphological characteristics of the white matter fiber bundles of the central nervous system. In view of this, when analyzing the fiber bundles of the brain and spinal cord, a method for extracting individualized fiber bundles of interest is needed, so as to extract the brain-spinal cord joint ROI (Region of Interest) fiber bundles in individuals for subsequent analysis.

[0003] Xi'an Children's Hospital and Xi'an University of Electronic Science and Technology disclosed a system and method for extracting fiber bundles of interest based on clustering denoising in their jointly applied patent document "System and method for extracting fiber bundles of interest based on clustering denoising" (application number: CN 202210258177.6, application publication number: CN114627283A). The system disclosed in the patent document includes a fiber bundle extraction module that tracks the whole brain fiber bundles on a standard template, and then extracts the set of fiber bundles of interest passing through the ROI by calculating the distance from each fiber bundle to the ROI; the clustering denoising module clusters the set of fiber bundles of interest using the fiber bundle clustering method, and then removes the noise fiber bundles in the clustered set of fiber bundles of interest; the fiber bundle inverse transformation module inversely transforms the denoised set of fiber bundles of interest to individual images to obtain the set of fiber bundles of interest for each individual. There are two shortcomings in this system: First, it is difficult to obtain the whole brain tracking fiber bundles for the standard template of the brain-spinal joint. The fiber bundle extraction module used subsequently lacks a widely recognized brain-spine combined DTI standard template, making it difficult to construct an effective fiber bundle template based on the standard template, which makes the system face difficulties in fiber bundle analysis of brain-spine combined DTI images. Secondly, it is very difficult to transform the extracted fiber bundles to individual space. The fiber bundle inverse transformation module used subsequently aims to transform the standard space fiber bundles to individual space, but there is currently a lack of effective, convenient and reliable registration tools for brain-spine synchronized DTI data. It is difficult to analyze individual fiber streamlines using existing registration tools and processes, which makes it difficult for the system to obtain continuous and accurate individual fibers of interest.

[0004] Nanjing Huinao Cloud Computing Co., Ltd. discloses a method for extracting fiber bundles by using ROI as a seed point for fiber bundle tracing in its patented technology "A method and system for tracing and analyzing white matter fiber bundles based on magnetic resonance imaging" (application number: CN 202010029910.8, authorization announcement number: CN 110827282B). The implementation steps of this method are: through the white matter fiber bundle reconstruction module, the ROI of each individual is used as a seed point, and multiple streamlines are used to start from the seed point and extend to all voxels in the whole brain in the direction of fiber bundle distribution probability, and finally all fiber bundles with a connection probability greater than the threshold starting from the ROI are regarded as extracted fiber bundles passing through the ROI. The shortcomings of this method are: the extracted fiber bundles are of various types and relatively complex. When the noise fibers are removed by overlapping all individual fiber bundles in the subsequent process, it is difficult to construct an effective overlapping area due to the excessive number of noise fibers, which makes it difficult for this method to achieve a good denoising effect. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a DTI fiber bundle extraction system and method based on clustering and overlap rate denoising, aiming to solve the problems of poor authenticity, weak stability and significant individual differences in the results presented by brain and spinal cord fiber bundles.

[0006] The idea of ​​achieving the purpose of the present invention is: in the individual space, the coordinates of the starting point, midpoint and end point of each fiber bundle of interest are extracted to form a set of starting point, midpoint and end point coordinates, and the density-based spatial clustering and noise algorithm are applied respectively to generate respective clustering labels, and the clustering labels are assembled into a three-dimensional array; the distance between each two coordinates is calculated, and when the distance between the coordinates is 0, it means that the two fiber bundles belong to the same category, and this process is iterated repeatedly until all fiber bundles have unique classification labels, and finally the set of fiber bundles of interest after denoising and classification is obtained; a mapping image is generated for the individual fiber bundle according to its spatial distribution and quantity, and it is registered to the standard space template space with the help of the registration parameters of the T1 structural image; all the fiber bundles registered to the standard space are averaged to obtain the fiber bundle mapping image, and a new fiber bundle mask image is generated with a 50% overlap probability of the mapping image; the new fiber bundle mask image is inversely transformed back to the individual space using the registration parameters of the T1 structural image in the individual space, and the intersection is taken with the classified fiber bundle set to obtain the individual fiber bundle set of interest after the overlap rate is denoised.

[0007] The system of the present invention comprises an individual brain-spine synchronous DTI image data preprocessing module, an individual spatial T1 structural image preprocessing module, an individual spatial brain-spine joint fiber bundle tracking module, an individual spatial brain-spine joint fiber bundle of interest clustering module, and an overlap rate denoising module for brain-spine joint fiber bundles of interest; wherein:

[0008] The individual brain-spine synchronous DTI image data preprocessing module is used to preprocess the brain-spine DTI images including noise reduction, Gibbs ringing correction and eddy current correction;

[0009] The individual spatial T1 structural image preprocessing module is used to perform N4 deviation field normalization processing on the individual T1 structural image, segment the spinal cord T1 image according to anatomical sites, and jointly register the spinal cord T1 image with the standard template;

[0010] The individual spatial brain-spinal combined fiber bundle tracking module is used to track the fiber bundles of interest of the brain-spinal combined for the individual;

[0011] The individual space brain-spinal joint fiber bundle clustering module of interest is used to cluster the fiber bundles of interest in the individual space;

[0012] The overlap rate denoising module of the brain-spinal joint fiber bundle of interest is used to transform the mapping image of the individual spatial fiber bundle into the standard space, take the mean of all individual images to obtain the overlap map and then denoise it.

[0013] The steps of the extraction method of the present invention include the following:

[0014] Step 1: The individual brain-spine synchronous DTI image data preprocessing module performs preprocessing on the brain-spine DTI images including noise reduction, Gibbs ringing correction, eddy current correction and normalization;

[0015] Step 2: The individual spatial T1 structural image preprocessing module performs normalization processing of the N4 deviation field on the individual T1 structural image, segments the image according to anatomical sites, and jointly registers the spinal cord T1 image with the standard template;

[0016] Step 3, the individual spatial brain-spinal joint fiber tract tracing module tracks the fiber tracts of interest of the brain-spinal joint for the individual;

[0017] Step 4: Individual space brain-spinal joint fiber bundle clustering module of interest: clustering the fiber bundles of interest in the individual space;

[0018] Step 5: The overlap rate denoising module of the brain-spinal joint fiber bundles of interest transforms the mapping image of the individual space fiber bundles into the standard space, takes the mean of all individual images to obtain the overlap map, and then denoises it.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] First, because the system of the present invention adopts an individual space brain-spinal joint fiber bundle clustering module of interest and an overlap rate denoising module of brain-spinal joint fiber bundles of interest, it overcomes the problems of poor brain-spinal joint DTI image quality, low template recognition and poor fiber bundle specificity extraction in the prior art, so that the system of the present invention improves the accuracy and authenticity of fiber bundle extraction.

[0021] Second, since the method of the present invention performs double denoising of clustering denoising and group overlap rate denoising on the fiber bundles of interest, it effectively reduces the problem of high false positive rate in the individual tracking process and improves the authenticity and accuracy of the fiber bundles. At the same time, it also overcomes the shortcomings of the single denoising method of the prior art, so that the method of the present invention achieves a stable and efficient fiber bundle optimization effect.

[0022] Third, since the method of the present invention uses the T1 structural image of the combined brain-spinal scan to provide a set of standardized processes for brain and spinal cord images using different tools, it solves the problem that existing brain alignment tools are unable to standardize brain-spinal combined DTI images, making the present invention more feasible and reliable in terms of standardization. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a structural schematic diagram of the system of the present invention;

[0024] Figure 2 is a flow chart of the method of the present invention;

[0025] Figure 3 is a flowchart of preprocessing of brain-spinal synchronous DTI images of the method of the present invention;

[0026] Figure 4 is a preprocessing flow chart of T1 structural image of the present invention;

[0027] Figure 5 : is a schematic diagram of tracking, group overlap, and de-anxiety of individual brain-spinal fiber bundles of interest of the present invention; wherein, Figure 5 (a) is a flow chart of tracking an individual brain-spinal fiber bundle of interest of the present invention; Figure 5 (b) is a schematic diagram of the results of extracting individual brainstem-spinal cord fiber bundles according to the present invention; Figure 5 (c) is a flow chart of the group overlap rate and de-noising of the present invention. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0029] Reference Figure 1 , the system of the present invention is further described in detail.

[0030] The system of the present invention comprises an individual brain-spine synchronous DTI image data preprocessing module, an individual spatial T1 structural image preprocessing module, an individual spatial brain-spine joint fiber bundle tracking module, an individual spatial brain-spine joint fiber bundle of interest clustering module, and an overlap rate denoising module for brain-spine joint fiber bundles of interest; wherein:

[0031] The individual brain-spine synchronous DTI image data preprocessing module is used to preprocess the brain-spine DTI images including noise reduction, Gibbs ringing correction, eddy current correction and normalization;

[0032] The individual spatial T1 structural image preprocessing module is used to perform N4 deviation field normalization processing on the individual T1 structural image, segment the spinal cord T1 image according to anatomical sites, and jointly register the spinal cord T1 image with the standard template;

[0033] The individual spatial brain-spinal combined fiber bundle tracking module is used to track the fiber bundles of interest of the brain-spinal combined for the individual;

[0034] The individual space brain-spinal joint fiber bundle clustering module of interest is used to cluster the fiber bundles of interest in the individual space;

[0035] The overlap rate denoising module of the brain-spinal joint fiber bundle of interest is used to transform the mapping image of the individual spatial fiber bundle into the standard space, take the mean of all individual images to obtain the overlap map and then denoise it.

[0036] Reference Figure 2 , the implementation steps of the method of the present invention are further described.

[0037] Step 1: The individual brain-spine synchronous DTI image data preprocessing module performs preprocessing on the brain-spine DTI images including noise reduction, Gibbs ringing correction, eddy current correction and normalization.

[0038] The steps of performing noise reduction, Gibbs ringing correction, eddy current correction and normalization preprocessing on the brain-spinal DTI image are as follows:

[0039] In the first step, the complete sequence of brain-spinal synchronization DTI images obtained by scanning along the A→P direction of the individual and the brain-spinal synchronization b=0s / mm obtained by scanning along the P→A direction were compared. 2 The images are fused into a complete 4D brain-spinal synchronized DTI image.

[0040] In the second step, the 4D brain-spine synchronized DTI images were subjected to noise level estimation, denoising, and Gibbs ringing artifact removal based on Marchenko-Pastur PCA.

[0041] The third step is to extract a b = 0s / mm scanned along the A→P direction from the 4D brain-spine synchronized DTI image after noise removal and artifact removal. 2 Image and a scan along the P→A direction b=0s / mm 2 The images are fused to obtain a 4D image with b=0s / mm 2 image; and sequentially estimating and correcting the image distortion caused by the magnetic susceptibility effect on the 4D image to obtain a corrected output image and a parameter file during the correction process.

[0042] In the fourth step, the average image of the corrected output images is used as the mask image, and the parameter file in the correction process is used as the motion correction parameter to correct the distortion caused by eddy current and the subject motion of the 4D brain-spine synchronized DTI images after denoising and artifact removal.

[0043] The fifth step is to segment the corrected 4D brain-spinal synchronous DTI image into brain DTI image and spinal cord DTI image along the z direction with the lower edge of the cerebellum as the anatomical point, and generate a brain mask image and a spinal cord segmentation mask image respectively.

[0044] In the sixth step, the brain mask image and the spinal cord segmentation mask image are spliced ​​along the z direction with the lower edge of the cerebellum as the anatomical point to obtain a joint mask image of the 4D brain-spinal synchronized DTI image.

[0045] In the seventh step, the joint mask image of the DTI image is used as a mask image to perform B1 field non-uniformity correction on the 4D brain-spine synchronized DTI image after the distortion caused by eddy current and subject motion correction, so as to obtain the preprocessed individual brain-spine synchronized DTI image.

[0046] Reference Figure 3 , the process of preprocessing of individual brain-spinal synchronous DTI images of the method of the present invention is further described.

[0047] In the embodiment of the present invention, the fslmerge function of the FSL software is used to merge the complete sequence of brain-spinal synchronous DTI images obtained by scanning along the A→P direction of the individual with the b=0s / mm scanned along the P→A direction. 2 The images were fused to obtain 4D brain-spine synchronized DTI images. The denoise function of the MRtrix3 software was used to perform noise level assessment and denoising based on Marchenko-Pastur PCA on the individual original brain-spine synchronized DTI image data. The mrdegibbs function of the MRtrix3 software was used to effectively remove Gibbs ringing artifacts. The fslroi function of the FSL software was used to extract b=0s / mm from one scan along the A→P direction and one scan along the P→A direction. 2 Image, use fslmerge function to merge these two images into a 4D image with b=0s / mm 2 Image. The topup function was used to estimate and correct the image distortion caused by the magnetic susceptibility effect, and the output image and parameter files during the correction process were obtained. The eddy function of the FSL software was used to correct the distortion caused by eddy current and subject motion for the input image (4D brain-spine synchronized DTI image after denoising), the input mask image (the average image of the corrected output image) and the motion correction parameters (parameter files during the correction process). The fslroi function of the FSL software was used to divide the corrected 4D brain-spine synchronized DTI image into brain DTI image data and spinal cord DTI image data based on the lower edge of the cerebellum as the anatomical point. The bet function was used to generate the brain mask image for the brain DTI image. The sct_maths function of the Spinal cord toolbox software was used to generate the spinal cord segmentation mask image for the spinal cord DTI image data. Afterwards, the fslmerge function was used to reassemble the anatomical points along the 4D brain-spine synchronized DTI image into a joint mask image. The dwibiascorrect function of MRtrix3 software was used to perform B1 field non-uniformity correction on the input image (corrected 4D brain-spine synchronized DTI image) and the input mask image (combined mask image).

[0048] Step 2: The individual spatial T1 structural image preprocessing module performs normalization processing on the N4 deviation field of the individual T1 structural image, segments the spinal cord T1 image according to anatomical sites, and jointly aligns the spinal cord T1 image with the standard template.

[0049] Reference Figure 4 , the steps of normalizing the N4 deviation field of the individual T1 structural image, segmenting the spinal cord T1 image according to anatomical sites and co-registering the spinal cord T1 image with the standard template are further described.

[0050] In the first step, the lower edge of the cerebellum is used as the anatomical point along the z direction to segment the individual T1 structural image data into the brain T1 structural image and the spinal cord T1 structural image, and generate the brain mask image and the spinal cord segmentation mask image respectively.

[0051] In the second step, the brain mask image and the spinal cord segmentation mask image are spliced ​​along the z direction with the lower edge of the cerebellum as the anatomical point to generate a joint mask image of the T1 structural image.

[0052] In the third step, the joint mask image of the T1 structural image is used as the mask image, and the individual T1 structural image is normalized based on the N4 deviation field to obtain the normalized T1 structural image.

[0053] In the fourth step, the normalized T1 structural image was segmented into the brain T1 structural image and the spinal cord T1 structural image along the z direction with the lower edge of the cerebellum as the anatomical point; the brain T1 structural image and the spinal cord T1 structural image were co-registered with the brain template MNI152 and the spinal cord template PAM50 respectively in the order of linear registration and then nonlinear registration to obtain the transformed warp matrix image in the registration process.

[0054] Step 3: The individual spatial brain-spinal joint fiber bundle tracking module tracks the fiber bundles of interest of the brain-spinal joint for the individual.

[0055] The steps for tracing the fiber tract of interest at the brain-spinal junction are as follows:

[0056] In the first step, a ROI set is formed for all brain regions where the fiber bundles of interest pass through in the brain and spinal cord;

[0057] In the second step, the two ROIs closest to the two ends of the fiber bundle of interest are selected from the ROI set as the seed areas for fiber bundle tracking, and the remaining ROIs are used as the areas through which the fiber bundle passes; the preprocessed individual brain-spinal synchronized DTI image is used as the input image, and the joint mask image of the 4D brain-spinal synchronized DTI image is used as the mask image.

[0058] In the third step, the preprocessed individual brain-spine synchronized DTI images are used as input images, and the joint mask image of the 4D brain-spine synchronized DTI images is used as the mask image;

[0059] In the fourth step, the tracking angle was set to 15 to 30°, the tracking step was set to 2 mm, the minimum number of tracking lines was set to 5000, the maximum FA value was set to 0.2, and the probabilistic fiber tracking algorithm was used to obtain the three-dimensional spatial coordinates of the fiber bundles of interest and all sampling points on each fiber bundle.

[0060] Reference Figure 5 (a) Further description of the method of the present invention for obtaining fiber bundles of interest in individual space.

[0061] In an embodiment of the method of the present invention, the C1 to C7 vertebrae of the brainstem and spinal cord were selected as the set of interest using MRtrix3 software; wherein the C7 segment of the brainstem and spinal cord was the seed region, and the remaining ROIs were the regions through which the fiber bundles were determined; in terms of specific parameter settings, the maximum tracking angle was set to 30°, the tracking step was 2 mm, the number of tracking bars was 5000, the FA cutoff value was 0.1, the minimum length was 140× the voxel size, and the maximum length was 145× the voxel size; the TensorProb probabilistic fiber tracking algorithm was used to obtain the fiber bundles of interest in the individual space.

[0062] Step 4: The individual space brain-spinal joint fiber bundle clustering module clusters the fiber bundles of interest in the individual space.

[0063] The steps of clustering the fiber bundles of interest in the individual space are as follows:

[0064] In the first step, the coordinates of the starting point, midpoint and end point of each fiber bundle of interest in the set of fiber bundles of interest tracked in the individual space are extracted respectively to form the starting point coordinate set, midpoint coordinate set and end point coordinate set of the fiber bundle of interest;

[0065] The second step is to implement the clustering operation of density-based spatial clustering using noise algorithm on the starting point coordinate set, the midpoint coordinate set and the end point coordinate set, and obtain the cluster labels corresponding to each of the three coordinate sets;

[0066] In the third step, the clustering labels of the three coordinate sets are combined into an array containing three columns. Each row of the three-dimensional array is regarded as a three-dimensional coordinate, and the distance between every two coordinates is calculated. When the distance between the coordinates is 0, it means that the two fiber bundles belong to the same category. This process is continuously repeated until all fiber bundles have unique labels, and finally the denoised and classified fiber bundle set of interest is obtained.

[0067] The following is a special example to further describe the clustering of fiber bundles of interest in the individual space in this step.

[0068] The special case of the fiber bundle set F of interest contains 10 fiber bundles F = {F 1 ,F 2 ,…,F 10}, each fiber bundle has several different three-dimensional coordinates. Due to the changeable shape of the fiber bundle, it is difficult to cluster the fiber bundles well by relying only on the position information of a single position coordinate point. Therefore, the fiber bundle clustering algorithm of the present invention uses the coordinate points of the starting point, midpoint, and end point of the fiber bundle to cluster the fiber bundles of interest. The steps are as follows:

[0069] In the first step, the first coordinate of each fiber bundle is recorded as the starting point of the fiber bundle, the middle coordinate is the midpoint of the fiber bundle, and the last coordinate point is the end point of the fiber bundle. The first coordinate point, the middle coordinate point, and the last coordinate point of all fiber bundles in F are extracted to form the starting point coordinate set S = {S 1 ,S 2 ,…,S 10}、Midpoint coordinate set M={M 1 ,M 2 ,…,M 10} and the midpoint coordinate set E = {E 1 ,E 2 ,…,E 10}.

[0070] In the second step, the starting point coordinate set S, the midpoint coordinate set M and the end point coordinate set E are clustered by using the density-based spatial clustering and noise algorithm. The clustering parameters are set as follows: the minimum number of adjacent points around the core point MinPts = 2, and the minimum search radius of the core point as the center of the circle ε = 2mm. After clustering, the cluster label of the starting point coordinate set S is s = {s 1 ,s 2 ,…,s 10}, the cluster label of the midpoint coordinate set M is m = {m 1 ,m 2 ,…,m 10} and the cluster label of the endpoint coordinate set E is e={e 1 ,e 2 ,…,e 10}.

[0071] The third step is to cluster the three coordinate sets and combine the cluster labels into an array containing three columns:

[0072]

[0073] Among them, s is the clustering label of the starting point coordinate set S after clustering, m is the clustering label of the midpoint coordinate set M, and e is the clustering label of the end point coordinate set E.

[0074] The fourth step is to calculate the Euclidean distance from the first fiber bundle to other fiber bundles, denoted as D 1 ={D 1,2 ,D 1,3 ,…,D 1,10}. When D 1 When there is a value of zero in , it means that the first fiber bundle and the other fiber bundle with the subscript are in the same category; then, the fiber bundles that have not been assigned a label are taken out and clustering is continued according to this strategy until all fiber bundles are assigned a unique clustering label. Finally, the fiber bundles in the set of fiber bundles of interest are divided into 23 categories. Figure 5 (b) shows some fiber bundle categories among the 23 clustering results.

[0075] Step 5: The overlap rate denoising module of the brain-spinal joint fiber bundles of interest transforms the mapping image of the individual space fiber bundles into the standard space, takes the mean of all individual images to obtain the overlap map, and then denoises it.

[0076] The overlapping image obtained by taking the mean of all individual images is obtained by the following formula:

[0077]

[0078] Where I represents the average fiber bundle mapping image of all individuals registered to the standard space, I i represents the fiber bundle mapping image of the ith individual in the standard space, n represents the total number of individuals, I min Represents the minimum gray value in I, I max Indicates the maximum grayscale value in I.

[0079] Reference Figure 5 (c) further describes the present invention's de-mania.

[0080] The re-denoising refers to obtaining the overlap rate image of the group of fiber bundles of interest in the standard space using the calculation formula of the overlap rate; selecting the overlap rate under the probability of 50% as the basis for constructing a new fiber bundle mask image. This selection ensures the uniformity and continuity of the constructed mask image, avoids the image breakage that may be caused by using too high a probability, and also reduces the false positive results that may be introduced by using a lower probability, thereby improving the accuracy and reliability of fiber bundle extraction; using the alignment parameters generated in the step of preprocessing the original T1 structural image, the new fiber bundle mask image in the standard space is inversely transformed to the individual brain-spine synchronous DTI space, and an intersection operation is performed with the set of fiber bundles of interest after individual clustering to obtain the set of fiber bundles of interest after re-denoising.

Claims

1. A brain-spinal DTI fiber bundle extraction system based on clustering and overlap rate denoising, comprising an individual brain-spinal synchronous DTI image data preprocessing module, an individual spatial T1 structural image preprocessing module, and an individual spatial brain-spinal joint fiber bundle tracking module; characterized in that: It also includes a clustering module of the brain-spinal joint fiber bundles of interest in individual space and an overlap rate de-noising module of the brain-spinal joint fiber bundles of interest; wherein: The individual brain-spine synchronous DTI image data preprocessing module is used to preprocess the brain-spine DTI images including noise reduction, Gibbs ringing correction, eddy current correction and normalization; The individual spatial T1 structural image preprocessing module performs N4 deviation field normalization processing on the individual T1 structural image, performs segmentation based on anatomical sites, and performs joint registration between the spinal cord T1 image and the standard template; The individual spatial brain-spinal combined fiber bundle tracking module is used to track the fiber bundles of interest of the brain-spinal combined for the individual; The individual space brain-spinal joint fiber bundle clustering module of interest is used to cluster the fiber bundles of interest in the individual space; the clustering steps are as follows: In the first step, the coordinates of the starting point, midpoint and end point of each fiber bundle of interest in the set of fiber bundles of interest tracked in the individual space are extracted respectively to form the starting point coordinate set, midpoint coordinate set and end point coordinate set of the fiber bundle of interest; The second step is to implement the clustering operation of density-based spatial clustering using noise algorithm on the starting point coordinate set, the midpoint coordinate set and the end point coordinate set, and obtain the cluster labels corresponding to each of the three coordinate sets; The third step is to combine the cluster labels of the three coordinate sets into an array containing three columns, and regard each row of the three-dimensional array as a three-dimensional coordinate, and calculate the distance between every two coordinates; when the distance between the coordinates is 0, it means that the two fiber bundles belong to the same category; by continuously looping this process until all fiber bundles have unique labels, we finally get the denoised and classified fiber bundle set of interest; The overlap rate noise reduction module of the brain-spinal joint fiber bundle of interest transforms the mapping image of the individual space fiber bundle into the standard space according to the following formula, takes the mean of all individual images to obtain the overlap map and then removes the noise; Wherein, I represents the average fiber bundle map image of all individuals registered to the standard space, Ii represents the fiber bundle map image of the ith individual in the standard space, n represents the total number of individuals, Imin represents the minimum grayscale value in I, and Imax represents the maximum grayscale value in I.

2. A brain-spinal DTI fiber bundle extraction method based on clustering and overlap rate denoising according to the extraction system of claim 1, characterized in that: Clustering the set of fiber bundles of interest extracted from the individual brain-spinal joint fiber bundles, and the overlap rate of the brain-spinal fiber bundles of interest in the individual space; the steps of the extraction method include the following: Step 1, the individual brain-spine synchronous DTI image data preprocessing module performs preprocessing on the brain-spine DTI image including noise reduction, Gibbs ringing correction, eddy current correction and normalization; Step 2: The individual spatial T1 structural image preprocessing module performs normalization processing of the N4 deviation field on the individual T1 structural image, segments the image according to anatomical sites, and jointly registers the spinal cord T1 image with the standard template; Step 3, the individual spatial brain-spinal joint fiber tract tracing module traces the fiber tracts of interest of the brain-spinal joint for the individual; Step 4: The individual space brain-spinal joint fiber bundle clustering module clusters the fiber bundles of interest in the individual space; the clustering steps are as follows: In the first step, the coordinates of the starting point, midpoint and end point of each fiber bundle of interest in the set of fiber bundles of interest tracked in the individual space are extracted respectively to form the starting point coordinate set, midpoint coordinate set and end point coordinate set of the fiber bundle of interest; The second step is to implement the clustering operation of density-based spatial clustering using noise algorithm on the starting point coordinate set, the midpoint coordinate set and the end point coordinate set, and obtain the cluster labels corresponding to each of the three coordinate sets; The third step is to combine the cluster labels of the three coordinate sets into an array containing three columns, and regard each row of the three-dimensional array as a three-dimensional coordinate, and calculate the distance between every two coordinates; when the distance between the coordinates is 0, it means that the two fiber bundles belong to the same category; by continuously looping this process until all fiber bundles have unique labels, we finally get the denoised and classified fiber bundle set of interest; Step 5: The overlap rate denoising module of the brain-spinal joint fiber bundles of interest transforms the mapping image of the individual space fiber bundles into the standard space, takes the mean of all individual images to obtain the overlap map, and then denoises it; The overlap ratio of the fiber bundles of interest in the group space is obtained by the following formula: Wherein, I represents the average fiber bundle map image of all individuals registered to the standard space, Ii represents the fiber bundle map image of the ith individual in the standard space, n represents the total number of individuals, Imin represents the minimum grayscale value in I, and Imax represents the maximum grayscale value in I.

3. The method for extracting brain-spinal DTI fiber bundles based on clustering and overlap rate denoising according to claim 2, characterized in that: The steps of preprocessing the brain-spinal DTI images described in step 1, including noise reduction, Gibbs ringing correction, eddy current correction and normalization, are as follows: In the first step, the complete sequence of brain-spinal synchronization DTI images obtained by scanning along the A→P direction of the individual and the brain-spinal synchronization b=0s / mm obtained by scanning along the P→A direction were compared. 2 The images are fused into a complete 4D brain-spine synchronized DTI image; In the second step, the noise level estimation, denoising and Gibbs ringing artifact removal were performed on the 4D brain-spine synchronized DTI images based on Marchenko-Pastur PCA. The third step is to extract a b = 0s / mm scanned along the A→P direction from the 4D brain-spine synchronized DTI image after noise removal and artifact removal. 2 Image and a scan along the P→A direction b=0s / mm 2 The images are fused to obtain a 4D image with b=0s / mm 2 image; sequentially estimating and correcting the image distortion caused by the magnetic susceptibility effect on the 4D image, and obtaining a corrected output image and a parameter file in the correction process; The fourth step is to use the average image of the corrected output images as the mask image and the parameter file in the correction process as the motion correction parameters to correct the distortion caused by eddy current and the subject motion of the 4D brain-spine synchronized DTI images after noise removal and artifact removal. The fifth step is to segment the corrected 4D brain-spine synchronous DTI image into a brain DTI image and a spinal cord DTI image along the z direction with the lower edge of the cerebellum as the anatomical point, and generate a brain mask image and a spinal cord segmentation mask image respectively; Step 6: Using the lower edge of the cerebellum as the anatomical point along the z direction, the brain mask image and the spinal cord segmentation mask image are spliced ​​to obtain a joint mask image of the 4D brain-spinal synchronized DTI image. In the seventh step, the joint mask image of the DTI image is used as a mask image to perform B1 field non-uniformity correction on the 4D brain-spine synchronized DTI image after the distortion caused by eddy current and subject motion correction, and obtain the preprocessed individual brain-spine synchronized DTI image.

4. The method for extracting brain-spinal DTI fiber bundles based on clustering and overlap rate denoising according to claim 2, characterized in that: The steps of normalizing the N4 deviation field of the individual T1 structural image in step 2, segmenting the spinal cord T1 image according to anatomical sites and co-registering the spinal cord T1 image with the standard template are as follows: In the first step, the lower edge of the cerebellum is used as the anatomical point along the z direction to segment the individual T1 structural image data to obtain the brain T1 structural image and the spinal cord T1 structural image, and generate the brain mask image and spinal cord segmentation mask image respectively; The second step is to combine the brain mask image and the spinal cord segmentation mask image along the z direction with the lower edge of the cerebellum as the anatomical point to generate a joint mask image of the T1 structural image. The third step is to use the joint mask image of the T1 structural image as the mask image, and perform normalization processing on the individual T1 structural image based on the N4 deviation field to obtain the normalized T1 structural image; In the fourth step, the normalized T1 structural image was segmented into the brain T1 structural image and the spinal cord T1 structural image along the z direction with the lower edge of the cerebellum as the anatomical point; the brain T1 structural image and the spinal cord T1 structural image were co-registered with the brain template MNI152 and the spinal cord template PAM50 respectively in the order of linear registration and then nonlinear registration to obtain the transformed warp matrix image in the registration process.

5. The method for extracting brain-spinal DTI fiber bundles based on clustering and overlap rate denoising according to claim 2, characterized in that: The steps for tracing fiber tracts of interest at the brain-spinal junction for an individual as described in step 3 are as follows: In the first step, a set of ROIs is formed for all brain regions through which the fiber bundles of interest pass in the brain and spinal cord; The second step is to select the two ROIs closest to the two ends of the fiber bundle of interest from the ROI set as the seed regions for fiber bundle tracking, and the remaining ROIs are the regions where the fiber bundle passes; In the third step, the preprocessed individual brain-spine synchronized DTI images are used as input images, and the joint mask image of the 4D brain-spine synchronized DTI images is used as the mask image; In the fourth step, the tracking angle was set to 15 to 30°, the tracking step was set to 2 mm, the minimum number of tracking lines was set to 5000, the maximum FA value was set to 0.2, and the probabilistic fiber tracking algorithm was used to obtain the three-dimensional spatial coordinates of the fiber bundles of interest and all sampling points on each fiber bundle.

6. The method for extracting brain-spinal DTI fiber bundles based on clustering and overlap rate denoising according to claim 2, characterized in that: The re-denoising described in step 5 refers to obtaining the overlap rate image of the group of fiber bundles of interest in the standard space using the calculation formula of the overlap rate; selecting the overlap rate under a probability of 50% as the basis for constructing a new fiber bundle mask image. This choice ensures the uniformity and continuity of the constructed mask image, avoids image breakage that may be caused by using too high a probability, and also reduces the false positive results that may be introduced by using a lower probability, thereby improving the accuracy and reliability of fiber bundle extraction; using the alignment parameters generated in the step of preprocessing the original T1 structural image, the new fiber bundle mask image in the standard space is inversely transformed to the individual brain-spine synchronous DTI space, and an intersection operation is performed with the set of fiber bundles of interest after individual clustering to obtain the set of fiber bundles of interest after re-denoising.

Images