CN111739580A - Brain white matter fiber bundle clustering method based on fiber midpoint and end points - Google Patents

Abstract

The invention provides a brain white matter fiber bundle clustering method based on fiber middle points and end points, aiming at improving the accuracy and efficiency of brain white matter fiber bundle clustering, and the implementation steps are as follows: (1) acquiring a whole brain fiber set; (2) dividing the whole brain fiber set; (3) determining the midpoint of the fiber; (4) midpoint clustering; (5) end point clustering; (6) determining an endpoint clustering result; (7) and acquiring a brain white matter fiber bundle clustering result. According to the invention, position information of the middle point and the end point of the fiber is used as the characteristic, the white matter fiber bundle of the brain is divided into a plurality of fiber classes through DBSCAN density clustering, the fibers in the classes are similar in shape and consistent in shape, the reduction of characteristic sensitivity is reduced, the clustering accuracy is improved, the calculated amount is reduced, and the clustering efficiency is improved.

Description

Brain white matter fiber bundle clustering method based on fiber midpoint and end points

Technical Field

The invention belongs to the technical field of image processing, relates to a method for clustering white matter fiber tracts, in particular to a method for clustering white matter fiber tracts based on fiber midpoints and end points, and can be applied to auxiliary research of the white matter fiber tracts.

Background

White matter, gray matter and cerebrospinal fluid are components of the brain, but the internal components of the brain are different from each other, so that water molecules in different brain tissue structures can have different diffusion behaviors, the water molecules in the cerebrospinal fluid mainly assume a free diffusion state, the water molecules in the gray matter assume a limited diffusion state, and the water molecules in the white matter are restrained by myelin fibers and neuron fibers, so that the resistance in the walking direction parallel to fiber bundles is small, the diffusion speed cannot be inhibited, the resistance in the walking direction perpendicular to the fiber bundles is large, and the diffusion speed is inhibited. Diffusion Tensor Imaging (DTI) is a non-invasive magnetic resonance imaging technology capable of detecting the dispersive motion of water molecules in a living body, and people can simulate and reproduce the diffusion track of water molecules of white matter in the brain by using DTI to form white matter fiber bundles. Generally, there are thousands of fiber bundles in the whole brain, so when analyzing a white fiber bundle, an interested fiber bundle is usually selected first, then a coordinate system is constructed on the interested fiber bundle, points on the fiber bundle are matched point by point, and then subsequent statistical analysis is performed. In order to ensure that the corresponding accuracy of the points is higher as much as possible, the fiber bundles are clustered, and then point matching is carried out on the fiber bundles with similar shapes and consistent shapes.

Clustering is the process of classifying unknown data into different classes or clusters according to degree of similarity. Fiber clustering methods are various, typically, similarity of fibers is calculated as characteristics, and then the characteristics are clustered by using methods such as K-means clustering or density clustering. The accuracy of clustering is an important index for measuring the quality of a clustering method, but is influenced by factors such as feature selection, data distribution and the like, and the improvement of the accuracy of clustering has certain challenges. In order to improve the accuracy of Clustering White Matter Fiber Tracts in the brain, researchers have made many attempts, for example, in the paper "advanced for Clustering White Matter Fiber traces" published in the American Journal of neurobiology in 2006 by the American college of science and artificial intelligence laboratory l.j.o' Donnell et al, a method for Clustering White Matter Fiber Tracts is disclosed, which first calculates the average distance between two fibers to obtain similarity values between fibers, then converts the distance into similarity using a gaussian function, and obtains a similarity matrix of fibers, and then performs feature decomposition on the similarity matrix to form a spectrum Clustering spectrum of the largest feature values, and finally performs Clustering on the spectrum using the Clustering method. For another example, in a paper "replication of superior white matter substrates using disuse-weighted imaging tragragraph" published by doctor gugue Guevara et al of the university institute of technologies of the university of the wisikang celebration in 2016 NeuroImage, a clustering method of white matter fiber bundles is disclosed. The two methods have the following defects: the characteristic value is calculated by calculating the distance matrix and then calculating the affinity diagram or calculating the similarity matrix to serve as the clustering characteristic, the sensitivity of the characteristic can be reduced, the accuracy of the clustering result is insufficient, meanwhile, the distance between each two fibers and each point need to be calculated, the calculation amount is large, and the efficiency is low.

Disclosure of Invention

The invention aims to overcome the defects in the prior art, provides a method for clustering white matter fiber tracts based on fiber midpoints and end points, and aims to improve the accuracy and efficiency of white matter fiber tract clustering.

In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps:

(1) obtaining a whole brain fiber set:

adopting image processing software, and carrying out whole brain determination type fiber tracking through a diffusion tensor image DTI of a brain vector position with a format of NIFTI and a gradient code of DTI scanning, a bval file and a bvec file to obtain a whole brain fiber set Fibers, wherein Fibers are { Fibers } Fibers₁,fiber₂,...,fiber_i,...,fiber_MTherein fiber_iDenotes the ith strip of 3D fibre, fiber_i＝{(x₁,y₁,z₁),(x₂,y₂,z₂),...,(x_j,y_j,z_j),...,(x_m,y_m,z_m)}，M≥1，m≥1，(x_j,y_j,z_j) Denotes the jth discrete point, x_j、y_jAnd z_jRespectively representing the voxel coordinates on the x, y and z axes corresponding to the DTI;

(2) segmenting the whole brain fiber set Fibers:

(2a) 3D region of interest ROI defining respectively the fiber start points in Whole brain fiber sets Fibers₁And 3D region of interest ROI of fiber end-point₂，ROI₁＝{(x′₁,y′₁,z′₁),(x′₂,y′₂,z′₂),...,(x′_k,y′_k,z′_k),...,(x′_p,y′_p,z′_p)}，ROI₂＝{(x″₁,y″₁,z″₁),(x″₂,y″₂,z″₂),...,(x″_l,y″_l,z″_l),...,(x″_q,y″_q,z″_q) Of which is (x'_k,y′_k,z′_k) Denotes the k-th discrete point, x'_k、y′_kAnd z'_kDenotes the voxel coordinates on the x, y and z axes corresponding to DTI, respectively, (x ″)_l,y″_l,z″_l) Denotes the i-th discrete point, x ″)_l、y″_lAnd z ″)_lRespectively on the x, y and z axes corresponding to DTIP is more than or equal to 1, and q is more than or equal to 1;

(2b) calculating fiber of each fiber_iAnd ROI₁The minimum Euclidean distance is obtained to obtain a minimum Euclidean distance set

Calculating fiber simultaneously_iAnd ROI₂The minimum Euclidean distance is obtained to obtain a minimum Euclidean distance set

Wherein,

respectively represent fiber_iAnd ROI₁、ROI₂Minimum euclidean distance of;

(2c) setting the distance threshold between discrete points as D, D > 0, and judging

And is

If true, obtaining Fibers simultaneously with the ROI₁And ROI₂Intersecting fiber bundle extract, extract ═ fiber₁,fiber₂,...,fiber_r,...,fiber_NWherein, N is more than or equal to 1 and less than or equal to M, fiber_rDenotes the r-th strip of 3D fibre, fiber_r＝{(x₁,y₁,z₁),(x₂,y₂,z₂),...,(x_s,y_s,z_s),...,(x_t,y_t,z_t),...,(x_n,y_n,z_n)}，n≥2，s∈[1,n)，t∈(1,n]，s＜t，(x_s,y_s,z_s) Representation and ROI₁Satisfies a distance threshold D1 st discrete point, x_s、y_sAnd z_sRespectively, the x, y and z axis voxel coordinates corresponding to DTI, (x)_t,y_t,z_t) Representation and ROI₂Satisfies the last 1 discrete point, x, of the distance threshold D_t、y_tAnd z_tRespectively representing the voxel coordinates on the x, y and z axes corresponding to the DTI;

(2d) screening out fiber_rMedian ROI₁Intersecting the 1 st discrete point (x)_s,y_s,z_s) To and from ROI₂Last 1 discrete point of intersection (x)_t,y_t,z_t) At discrete points in between, a fiber bundle extract' is obtained, which is { fiber }₁′,fiber₂′,...,fiber_r′,...,fiber_N' }, wherein, fiber_r' means the r-th strip of 3D fiber, fiber_r′＝{b_r,...,c_r}，b_r＝(x_s,y_s,z_s) Is the starting point of the r-th fiber, c_r＝(x_t,y_t,z_t) Is the end point of the r-th fiber, and the Start of all fibers in a fiber bundle is obtained, Start ═ b₁,b₂,...,b_r,...,b_NAnd a final set of all fibers Last, Last ═ c₁,c₂,...,c_r,...,c_N}；

(3) Determining fiber of each fiber_r' midpoint:

calculating fiber_r' Total number Num of discrete points in, and judge if Num is odd, if yes, will the second

A discrete point as fiber_r' midpoint a_rOtherwise, it will be

A discrete point as fiber_r' midpoint a_rThen the midpoints of all the fibers form a midpoint set Middle, Middle ═ a₁,a₂,...,a_r,...,a_N}，a_r＝(x_w,y_w,z_w)，w∈(s,t)；

(4) Clustering the midpoint set Middle:

clustering the midpoint set Middle by using a Noise-Based Density spatial clustering DBSCAN (Density-Based spatial clustering of Applications with Noise) method to obtain a clustering result A of midpoint clustering, wherein A is { A ═ A { (A }₁,A₂,...,A_e,...,A_EE is more than or equal to 1 and less than or equal to N, wherein A_eIs composed of_eN corresponding to the middle point of each fiber_eGroup e fibers of sliver fibers;

(5) for each fiber class A_eAnd (3) carrying out endpoint clustering:

a is to be_eN in (1)_eThe starting points of the fibers corresponding to the strip fibers are combined into a starting point set Start_eAnd for Start_ePerforming DBSCAN clustering to obtain A_eStarting point clustering result B of_e，B_e＝{(B_e)₁,(B_e)₂,...,(B_e)_f,...,(B_e)_FAt the same time, A is added_eN in (1)_eCombining the fiber end points corresponding to the strip fibers into an end point set Last_eAnd to Last_ePerforming DBSCAN clustering to obtain A_eEnd point clustering result C of_e，C_e＝{(C_e)₁,(C_e)₂,...,(C_e)_g,...,(C_e)_GAnd (c) the step of (c) in which,

1≤F≤n_e，1≤G≤n_e，(B_e)_fis composed of_fM corresponding to each fiber starting point_fType f fibers of sliver fibers, (C)_e)_gIs composed of_gM corresponding to each fiber terminal point_gType g fibers of sliver fibers;

(6) determination of A_eEnd point clustering result of (2):

(6a) judgment B_eOr C_eWhether all of the discrete points in (a) are identified as noise points by the DBSCAN cluster,if so, A_eThe fiber in (A) is divided into n_eClass, each fiber is self-classified into 1 class, giving A_eEnd point clustering results clusters_eWherein, otherwise, step (6b) is performed;

(6b) judging whether F is equal to G or not, if so, A is_eThe fibers in (A) are classified into 1 type to obtain A_eEnd point clustering results clusters_eOtherwise, executing step (6 c);

(6c) judging whether F is 1 and G is more than 1, if so, A_eThe fibers in (A) are classified into G groups to obtain A_eEnd point clustering results clusters_e，clusters_e＝{(C_e)₁,(C_e)₂,...,(C_e)_g,...,(C_e)_GElse, executing step (6 d);

(6d) judging whether G is 1 and F is more than 1, if so, A_eThe fibers in (A) are classified into F groups to obtain A_eEnd point clustering results clusters_e，clusters_e＝{(B_e)₁,(B_e)₂,...,(B_e)_f,...,(B_e)_FElse, executing step (6 e);

(6e) each (B)_e)_fM in (1)_fCombining the fiber end points corresponding to the strip fibers into an end point set (Last)_e)_fAnd are paired (Last)_e)_fPerforming DBSCAN clustering to obtain B_eEnd point clustering result of (B)_e′，B_e′＝{(I_e)₁,(I_e)₂,...,(I_e)_f′,...,(I_e)_F′At the same time, each (C) is put in_e)_gM in (1)_gThe starting points of the fibers corresponding to the strip fibers are combined into a starting point set (Start)_e)_fAnd on (Start)_e)_fPerforming DBSCAN clustering to obtain C_eStarting point clustering result C of_e′，C_e＝{(J_e)₁,(J_e)₂,...,(J_e)_g′,...,(J_e)_G′And (c) the step of (c) in which,

1≤F′≤n_e，1≤G′≤n_e，(I_e)_f′comprises with m_f′M corresponding to each fiber terminal point_f′Sliver fiber, (J)_e)_g′Comprises with m_g′M corresponding to each fiber starting point_g′Sliver fiber;

(6f) judging whether F '> G' is true, if yes, A_eThe fibers in (A) are classified into F' groups to give A_eEnd point clustering results clusters_e，clusters_e＝{(I_e)₁,(I_e)₂,...,(I_e)_f′,...,(I_e)_F′Else, A_eThe fiber bundles in (1) are classified into G' groups to obtain A_eEnd point clustering results clusters_e，clusters_e＝{(J_e)₁,(J_e)₂,...,(J_e)_g′,...,(J_e)_G′}；

(7) Obtaining a clustering result of white matter fiber bundles of the brain:

combining the end point clustering results of all the fiber classes in the A to obtain a clustering result Cluster of the fiber bundle track', wherein the clustering result Cluster is { Cluster₁,clusters₂,...,clusters_e,...,clusters_E}, where clusters_eAnd representing the end point clustering result of the e-th fiber.

Compared with the prior art, the invention has the following advantages:

1. the method clusters the white matter fiber bundles of the brain through the fiber middle points and the fiber end points, and avoids the sensitivity reduction caused by the fact that the prior art utilizes indirect similarity measure as clustering characteristics. Compared with the prior art, the clustering accuracy of the brain fiber bundles is effectively improved.

2. The invention directly utilizes the coordinate information of the middle point and the end point of the fiber as the characteristic to carry out clustering, thereby avoiding the problem that the prior art needs to calculate the distance between the points between the fibers, obtain the similarity matrix and then extract the characteristic value as the characteristic to carry out clustering. Compared with the prior art, the method effectively improves the clustering efficiency of the white matter fiber bundles of the brain.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention;

FIG. 2 is a schematic illustration of a fiber bundle in an embodiment of the present invention;

fig. 3 is a graph comparing simulation of clustering accuracy of the present invention and the prior art.

Detailed Description

The invention will be described in further detail with reference to the following figures and specific examples, it being emphasized that the invention is not part of a method for the diagnosis and treatment of diseases:

referring to fig. 1, the present invention includes the steps of:

step 1) obtaining a whole brain fiber set:

adopting ExploreDTI software, and carrying out whole brain determination type fiber tracking through a diffusion tensor image DTI of a brain vector position with a format of NIFTI and a gradient code of DTI scanning, a bval file and a bvec file to obtain a whole brain fiber set Fibers, wherein the Fibers are { Fibers } Fibers₁,fiber₂,...,fiber_i,...,fiber₂₉₈₄₉Therein fiber_iDenotes the ith strip of 3D fibre, fiber_i＝{(x₁,y₁,z₁),(x₂,y₂,z₂),...,(x_j,y_j,z_j),...,(x_m,y_m,z_m)}，m≥1，(x_j,y_j,z_j) Denotes the jth discrete point, x_j、y_jAnd z_jRepresenting the voxel coordinates on the x, y and z axes, respectively, corresponding to the DTI.

Step 2) segmenting the whole brain fiber set Fibers:

step 2a) the whole brain fiber set comprises thousands of fibers, and in order to analyze the microstructure of the fibers, an interested fiber bundle is usually selected first and is segmented from the whole brain fiber set; in this embodiment, the method is proposed by selecting the Nap center of the automated institute of Chinese academy of sciencesRegion of interest ROI (region of interest) Atlas BN _ Atlas _246_2mm. ni i to define a 3D region of interest ROI of the start of fibres in a whole brain fibre set Fibers₁And 3D region of interest ROI of fiber end-point₂And make the atlas and whole brain fiber set in the same space, ROI₁Selected as ROI 215 in the map, named as rHipp _ L, which contains 603 discrete points inside, ROI₂Selected as ROI 216 in the map, named as rHipp _ R, which contains 537 discrete points inside₁＝{(x′₁,y′₁,z′₁),(x′₂,y′₂,z′₂),...,(x′_k,y′_k,z′_k),...,(x′₆₀₃,y′₆₀₃,z′₆₀₃)}，ROI₂＝{(x″₁,y″₁,z″₁),(x″₂,y″₂,z″₂),...,(x″_l,y″_l,z″_l),...,(x″₅₃₇,y″₅₃₇,z″₅₃₇) Of which is (x'_k,y′_k,z′_k) Denotes the k-th discrete point, x'_k、y′_kAnd z'_kDenotes the voxel coordinates on the x, y and z axes corresponding to DTI, respectively, (x ″)_l,y″_l,z″_l) Denotes the i-th discrete point, x ″)_l、y″_lAnd z ″)_lRespectively representing the voxel coordinates on the x, y and z axes corresponding to the DTI;

step 2b) calculating each fiber by using MATLAB software_iAnd ROI₁Because of fiber_iIn which there are a plurality of discrete points, ROI₁There are also a number of discrete points in it, so first the fiber is calculated_iEach discrete point (x) in_d,y_d,z_d) And ROI₁The Euclidean distances of all the discrete points in the image are calculated to obtain a distance set D, D ═ D₁,D₂,...,D_d,...,D_mThe calculation formula is as follows:

wherein (x)_d,y_d,z_d)∈fiber_i，P∈[1,432]，D_d＝{D_d-1,D_d-2,...,D_d-k′,...,D_d-432}，D_d-k′Represents (x)_d,y_d,z_d) To ROI₁The Euclidean distance of the k' th discrete point; then calculating the fiber_iAll discrete points in to ROI₁The minimum Euclidean distance of (D) to obtain a distance set D_d″，D_d″＝{D₁′,D₂′,...,D_d′,...,D_m' }, the calculation formula is as follows:

D_d′＝min(D_d),

wherein D is_d' means (x)_d,y_d,z_d) To ROI₁Minimum euclidean distance of; recalculate fiber_iAnd ROI₁Minimum euclidean distance of

The calculation formula is as follows:

finally obtaining a minimum Euclidean distance set

While calculating fiber in the same way_iAnd ROI₂The minimum Euclidean distance is obtained to obtain a minimum Euclidean distance set

Wherein,

respectively represent fiber_iAnd ROI₁、ROI₂Minimum euclidean distance of;

step 2c) when the discrete point on the fiber and the discrete point in the ROI exist in one voxel at the same time, the fiber is considered to intersect with the ROI, because the discrete point distance threshold D is influenced by the size of the voxel, which can be determined according to the actual situation, in this embodiment, the size of the voxel is 2 × 2 × 2, so the length of the diagonal line of the cube where the voxel is located is set as the threshold,

and judge

And is

If true, obtaining Fibers simultaneously with the ROI₁And ROI₂Intersecting fiber bundle extract, extract ═ fiber₁,fiber₂,...,fiber_r,...,fiber₁₆₇Therein, fiber_rDenotes the r-th strip of 3D fibre, fiber_r＝{(x₁,y₁,z₁),(x₂,y₂,z₂),...,(x_s,y_s,z_s),...,(x_t,y_t,z_t),...,(x_n,y_n,z_n)}，n≥2，s∈[1,n)，t∈(1,n]，s＜t，(x_s,y_s,z_s) Representation and ROI₁Satisfies the 1 st discrete point, x, of the distance threshold D_s、y_sAnd z_sRespectively, the x, y and z axis voxel coordinates corresponding to DTI, (x)_t,y_t,z_t) Representation and ROI₂Satisfies the last 1 discrete point, x, of the distance threshold D_t、y_tAnd z_tRespectively representing the voxel coordinates on the x, y and z axes corresponding to the DTI;

step 2d) because only ROI is concerned₁And ROI₂In between, so that in the ROI₁And ROI₂The other discrete points do not enter the subsequent steps, and the fiber is screened out_rMedian ROI₁Intersecting the 1 st discrete point (x)_s,y_s,z_s) To and from ROI₂Last 1 discrete point of intersection (x)_t,y_t,z_t) At discrete points in between, a fiber bundle extract' is obtained, which is { fiber }₁′,fiber₂′,...,fiber_r′,...,fiber₁₆₇' }, as shown in FIG. 2, in which fiber_r' means the r-th strip of 3D fiber, fiber_r′＝{b_r,...,c_r}，b_r＝(x_s,y_s,z_s) Is the starting point of the r-th fiber, c_r＝(x_t,y_t,z_t) Is the end point of the r-th fiber, and the Start of all fibers in a fiber bundle is obtained, Start ═ b₁,b₂,...,b_r,...,b₁₆₇And a final set of all fibers Last, Last ═ c₁,c₂,...,c_r,...,c₁₆₇}。

Step 3) determining fiber of each fiber_r' midpoint:

calculating fiber_r' Total number Num of discrete points in, and judge if Num is odd, if yes, will the second

A discrete point as fiber_r' midpoint a_rOtherwise, it will be

A discrete point as fiber_r' midpoint a_rThen the midpoints of all the fibers form a midpoint set Middle, Middle ═ a₁,a₂,...,a_r,...,a₁₆₇}，a_r＝(x_w,y_w,z_w)，w∈(s,t)。

Step 4) clustering the midpoint set Middle:

step 4a) because the midpoint can represent the trunk position information of the fiber bundle,therefore, clustering the midpoint set can roughly divide the fiber bundles into a plurality of fiber categories with longer distances, clustering the midpoint set Middle by using a noise-based density-based spatial clustering DBSCAN method, and finding out each midpoint a_rAnd determines a set of core objects omega, in this example 3, a for midpoint clustering_rThe number of midpoints in the neighborhood range MinPts is 3, and all the midpoints are connected with a_rIs not more than the middle point of the Euclidean distance to form a_rField of judgment a_rWhether the number of midpoints in the field is greater than or equal to MinPts, if so, a_rIs a core object, otherwise, a_rObtaining a core object set omega of Middle for a non-core object;

step 4b) randomly selecting a core object from omega as a seed, finding out all middle points which can be reached by the density of the core object, and assuming a_rIs a core object, and selects a_rAs a seed, a midpoint a in Middle_r′Is located at a_rIn the field of (1), then a_r′From a to a_rThe density is direct, and whether a midpoint sequence u exists is judged₁,u₂,...,u_vWherein u is₁＝a_r，u_v＝a_r′And u is_r+1By u_rDensity is up to a_r′From a to a_rThe density can be reached to obtain the first cluster A₁，A₁140 fibers respectively corresponding to the middle points of 140 fibers are arranged in the fiber;

step 4c) A₁Removing the core object contained in (1) from omega, wherein omega is not null, executing step 4b) to obtain a second cluster class A₂，A₂The middle of the fiber is provided with 27 fibers which respectively correspond to the middle points of the 27 fibers;

step 4d) A₂The core object included in (1) is removed from Ω, Ω is null, and all cluster classes are combined to obtain a clustering result a of Middle, where a is { a ═ a₁,A₂}。

Step 5) for fibers A₁And A₂And (3) carrying out endpoint clustering:

since in each midpoint clustering result the fibers are already compact at the stem location, onlyIt is necessary to subdivide the starting and ending points of the fibers in each fiber class, and A is₁The starting points of the fibers corresponding to 140 fibers are combined into a starting point set Start₁And for Start₁In this embodiment, in order to classify finer fibers, the end point cluster is set to 2 and MinPts is set to 3 to obtain a₁Starting point clustering result B of₁，B₁＝{(B₁)₁,(B₁)₂,(B₁)₃,(B₁)₄At the same time, A is added₁The fiber end points corresponding to 140 fibers in the fiber end point set are combined into an end point set Last₁And to Last₁Performing DBSCAN clustering to obtain A₁End point clustering result C of₁，C₁＝{(C₁)₁,(C₁)₂,(C₁)₃And (c) the step of (c) in which,

(B₁)₁contains 16 fibers corresponding to 16 fiber starting points respectively, (B)₁)₂Contains 51 fibers corresponding to 51 fiber starting points respectively, (B)₁)₃Contains 66 fibers corresponding to the 66 fiber starting points respectively, (B)₁)₄The fiber comprises 6 fibers corresponding to the 6 fiber starting points respectively, and the rest 1 fiber starting point is identified as a noise point, so that the fiber is also considered as a noise fiber and does not enter the subsequent flow; (C)₁)₁Contains 21 fibers corresponding to 21 fiber end points respectively, (C)₁)₂Contains 111 fibers corresponding to 111 fiber end points respectively, (C)₁)₃The fiber contains 7 fibers corresponding to 7 fiber end points respectively, and the rest 1 fiber starting point is identified as a noise point, so that the fiber is also considered as a noise fiber and does not enter the subsequent flow, and simultaneously, A is used₂The starting points of the fibers corresponding to the 27 fibers are combined into a starting point set Start₂And for Start₂Performing DBSCAN clustering to obtain A₂Starting point clustering result B of₂，B₂＝{(B₂)₁At the same time, A is added₂Combining the fiber end points corresponding to the 27 fibers into an end point set Last₂And to Last₂Performing DBSCAN clustering to obtain A₂End point clustering result C of₂，C₂＝{(C₂)₁And (c) the step of (c) in which,

(B₂)₁contains 27 fibers corresponding to the 27 fiber starting points respectively, (C)₂)₁The fiber contains 27 fibers corresponding to 27 fiber end points.

Step 6) determination of fibers A₁And A₂End point clustering result of (2):

step 6a) judging whether the tail ends of the fibers are orderly and consistent according to the clustering numbers of the starting point and the end point, firstly judging whether the two tail ends of the fibers are extremely dispersed, and judging whether the fibers in the fiber class are all identified as noise fibers, namely judging B₁Or C₁Whether all the discrete points in the step (6) are recognized as noise points by the DBSCAN cluster or not is judged, and if not, the step (6b) is executed;

step 6b), judging whether the two ends of the fiber are neat and consistent, namely judging whether F-G-1 is true, and if not, executing step 6 c);

step 6c) judging whether the fibers are regular at the starting point part and have bifurcations at the end point part, namely judging whether F is 1 and G is more than 1, and if not, executing step 6 d);

step 6d) judging whether the fibers are regular at the end point part and have bifurcations at the starting point part, namely judging whether G is 1 and F is more than 1, and if not, executing step 6 e);

step 6e) because the starting point position and the end point position of the fiber are both branched, the subdivision of the fiber is discussed further, the starting point position of the fiber obtained by the starting point clustering is considered to be regular, so that the end point clustering is performed on the fiber obtained by clustering each starting point again, and the step (B) is implemented₁)₁The fiber end points corresponding to the 16 fibers are combined into an end point set (Last)₁)₁And are paired (Last)₁)₁Performing DBSCAN clustering to obtain (B)₁)₂Combining the fiber end points corresponding to 51 fibers in the fiber end point set (Last)₁)₂And are paired (Last)₁)₂Performing DBSCAN clustering to obtain (B)₁)₃Combining the fiber end points corresponding to the medium 66 fibers into an end point set (Last)₁)₃And are paired (Last)₁)₃Performing DBSCAN clustering to obtain (B)₁)₄The fiber end points corresponding to the 6 fibers in the fiber bundle are combined into an end point set (Last)₁)₄And are paired (Last)₁)₄Performing DBSCAN clustering, integrating clustering results and removing fibers identified as noise to obtain B₁End point clustering result of (B)₁′，B₁′＝{(I₁)₁,(I₁)₂,(I₁)₃,(I₁)₄,(I₁)₅And (C) simultaneously considering that the end point positions of the fibers obtained by end point clustering are regular, so that the fibers obtained by clustering each end point are only required to be subjected to starting point clustering again, and₁)₁the fiber starting points corresponding to the 21 fibers in the fiber array are combined into a starting point set (Start)₁)₁And on (Start)₁)₁Performing DBSCAN clustering to obtain (C)₁)₂The starting points of the fibers corresponding to 111 fibers are combined into a starting point set (Start)₁)₂And on (Start)₁)₂Performing DBSCAN clustering to obtain (C)₁)₃The fiber starting points corresponding to 7 fibers in the fiber array are combined into a starting point set (Start)₁)₃And on (Start)₁)₃Performing DBSCAN clustering, integrating clustering results and removing fibers identified as noise to obtain C₁Starting point clustering result C of₁′，C₁′＝{(J₁)₁,(J₁)₂,(J₁)₃,(J₁)₄,(J₁)₅And (c) the step of (c) in which,

(I₁)₁containing 10 fibers corresponding to 10 fiber end points respectively, (I)₁)₂Containing 5 fibers corresponding to 5 fiber end points respectively, (I)₁)₃Containing 48 fibers corresponding to 48 fiber end points respectively, (I)₁)₄Containing 60 fibers corresponding to 60 fiber end points respectively, (I)₁)₅Containing 6 fibers corresponding to 6 fiber ends, respectively, 10 fibers were identified as noise fibers, (J)₁)₁Containing 10 fibers corresponding to 10 fiber origins, (J)₁)₂Comprising 6 fibers corresponding to 6 fiber origins, (J)₁)₃Containing 48 fibers corresponding to 48 fiber origins, (J)₁)₄Containing 60 fibers corresponding to 60 fiber origins, (J)₁)₅Contains 5 fibers corresponding to the 5 fiber starting points respectively, and 10 fibers are identified as noise fibers;

step 6F) regarding the fiber class subjected to subclass endpoint clustering, that is, considering that two tail ends are neat and consistent, in order to select a more detailed clustering result, selecting a fiber class with more clustering classes as a final clustering result, namely judging whether F '> G' is true or not, if not, F '═ G' ═ 5, and considering that a good clustering effect can be obtained no matter which classification method is selected when the class numbers are the same, wherein A is a method for determining whether the class numbers are the same or not, and A is a method for determining whether the class numbers₁In accordance with C₁' the clustering results are divided into 5 classes, resulting in A₁End point clustering results clusters₁，clusters₁＝{(J₁)₁,(J₁)₂,(J₁)₃,(J₁)₄,(J₁)₅}；

Step 6g) A₂The starting points of the fibers corresponding to the 27 fibers are combined into a starting point set Start₂And for Start₂Performing DBSCAN clustering to obtain A₂Starting point clustering result B of₂，B₂＝{(B₂)₁At the same time, A is added₂Combining the fiber end points corresponding to the 27 fibers into an end point set Last₂And to Last₂Performing DBSCAN clustering to obtain A₂End point clustering result C of₂，C₂＝{(C₂)₁And (c) the step of (c) in which,

(B₂)₁contains 27 fibers corresponding to the 27 fiber starting points respectively, (C)₂)₁Comprises 27 fibers corresponding to the 27 fiber starting points respectively;

step 6h) judging B₂Or C₂Whether all the discrete points in the step (6 i) are recognized as noise points by the DBSCAN cluster or not is judged, and if not, the step (6 i) is executed;

step 6i) determines whether F ═ G ═ 1 is true, yes, a₂The fibers in (A) are classified into 1 type to obtain A₂End point clustering results clusters₂。

Step 7) obtaining a clustering result of white matter fiber bundles of the brain:

combining the end point clustering results of all the fiber classes in the A to obtain a clustering result Cluster of the fiber bundle track', wherein the clustering result Cluster is { Cluster₁,clusters₂The total of the components of the Chinese herbal medicine is divided into 6 types, which respectively comprise 10, 6, 48, 60, 5 and 27 fibers and 11 noise fibers, and the whole process takes 0.047 s.

The technical effect of the present invention will be described below with reference to simulation experiments.

1. Simulation conditions and contents:

the fiber bundle in this example was subjected to a "extract', extract ═ fiber" using MATLAB software₁′,fiber₂′,...,fiber_r′,...,fiber₁₆₇', simulated using the white matter fiber tract clustering method involved in the paper "reproduction of superior white matter tracks using-weight imaging mapping" published by doctor Guevara, Miguel, et al, NeuroImage 2016.

Firstly, each fiber is resampled into 51 points at equal intervals, and then each fiber is calculated_i' Each point a to fiber on_j' Euclidean distance of all the points, then taking the maximum value to obtain each point a to fiber_jDistance of' and then fiber_iAll points in' to fiber_jDistance of `Separating and averaging to obtain fiber_i' to fiber_j' average farthest distance, fiber_i′,fiber_j′∈tract′，fiber_i′≠fiber_j', to obtain the average farthest distance between all the fibers in the trace', wherein fiber_i' to fiber_j' average farthest distance to fiber_j' to fiber_i' the average farthest distances are different, the minimum value among them is taken to form a distance matrix d, and then an affinity diagram W is calculated for d, and a formula is used

The width of the Gaussian kernel is calculated according to the empirical value²And (5) performing hierarchical clustering on the W to obtain a clustering result, wherein the whole process is used for 33.942 s.

2. And (3) simulation result analysis:

referring to fig. 3, fig. 3(a) shows a schematic diagram of a result of white matter fiber bundle clustering implemented by using the present invention in this embodiment, fig. 3(b) shows a schematic diagram of a result of white matter fiber bundle clustering implemented by using the prior art in this embodiment, fig. 3(c) shows a partial enlarged view in a black dashed frame in fig. 3(a), and a gray level corresponding to each fiber type is marked in a color bar in fig. 3, it can be seen from the figure that 6 types are obtained by using the clustering method of the present invention, and trace' is divided into 3 types of fibers by using the clustering method of the prior art.

Claims (3)

1. A brain white matter fiber bundle clustering method based on fiber midpoint and end points is characterized by comprising the following steps:

(1) obtaining a whole brain fiber set:

adopting image processing software, and carrying out whole brain determination type fiber tracking through a diffusion tensor image DTI of brain vector position with a format of NIFTI and a gradient code of DTI scanning, a bval file and a bvec file to obtain a whole brain determination type fiber trackingBrain Fibers collection Fibers, Fibers ═ fiber { fiber }₁,fiber₂,...,fiber_i,...,fiber_MTherein fiber_iDenotes the ith strip of 3D fibre, fiber_i＝{(x₁,y₁,z₁),(x₂,y₂,z₂),...,(x_j,y_j,z_j),...,(x_m,y_m,z_m)}，M≥1，m≥1，(x_j,y_j,z_j) Denotes the jth discrete point, x_j、y_jAnd z_jRespectively representing the voxel coordinates on the x, y and z axes corresponding to the DTI;

(2) segmenting the whole brain fiber set Fibers:

(2a) 3D region of interest ROI defining respectively the fiber start points in Whole brain fiber sets Fibers₁And 3D region of interest ROI of fiber end-point₂，ROI₁＝{(x′₁,y′₁,z′₁),(x′₂,y′₂,z′₂),...,(x′_k,y′_k,z′_k),...,(x′_p,y′_p,z′_p)}，ROI₂＝{(x″₁,y″₁,z″₁),(x″₂,y″₂,z″₂),...,(x″_l,y″_l,z″_l),...,(x″_q,y″_q,z″_q) Of which is (x'_k,y′_k,z′_k) Denotes the k-th discrete point, x'_k、y′_kAnd z'_kDenotes the voxel coordinates on the x, y and z axes corresponding to DTI, respectively, (x ″)_l,y″_l,z″_l) Denotes the i-th discrete point, x ″)_l、y″_lAnd z ″)_lRespectively representing the voxel coordinates on the x, y and z axes corresponding to the DTI, wherein p is more than or equal to 1, and q is more than or equal to 1;

(2b) calculating fiber of each fiber_iAnd ROI₁The minimum Euclidean distance is obtained to obtain a minimum Euclidean distance set

Calculating fiber simultaneously_iAnd ROI₂Minimum euclidean distance ofObtaining a minimum Euclidean distance set

Wherein,

respectively represent fiber_iAnd ROI₁、ROI₂Minimum euclidean distance of;

(2c) setting the distance threshold between discrete points as D, D > 0, and judging

And is

If true, obtaining Fibers simultaneously with the ROI₁And ROI₂Intersecting fiber bundle extract, extract ═ fiber₁,fiber₂,...,fiber_r,...,fiber_NWherein, N is more than or equal to 1 and less than or equal to M, fiber_rDenotes the r-th strip of 3D fibre, fiber_r＝{(x₁,y₁,z₁),(x₂,y₂,z₂),...,(x_s,y_s,z_s),...,(x_t,y_t,z_t),...,(x_n,y_n,z_n)}，n≥2，s∈[1,n)，t∈(1,n]，s＜t，(x_s,y_s,z_s) Representation and ROI₁Satisfies the 1 st discrete point, x, of the distance threshold D_s、y_sAnd z_sRespectively, the x, y and z axis voxel coordinates corresponding to DTI, (x)_t,y_t,z_t) Representation and ROI₂Satisfies the last 1 discrete point, x, of the distance threshold D_t、y_tAnd z_tRespectively representing the voxel coordinates on the x, y and z axes corresponding to the DTI;

(2d) screening out fiber_rMedian ROI₁Intersecting the 1 st discrete point (x)_s,y_s,z_s) To and from ROI₂Last 1 discrete point of intersection (x)_t,y_t,z_t) At discrete points in between, a fiber bundle extract' is obtained, which is { fiber }₁′,fiber₂′,...,fiber_r′,...,fiber_N' }, wherein, fiber_r' means the r-th strip of 3D fiber, fiber_r′＝{b_r,...,c_r}，b_r＝(x_s,y_s,z_s) Is the starting point of the r-th fiber, c_r＝(x_t,y_t,z_t) Is the end point of the r-th fiber, and the Start of all fibers in a fiber bundle is obtained, Start ═ b₁,b₂,...,b_r,...,b_NAnd a final set of all fibers Last, Last ═ c₁,c₂,...,c_r,...,c_N}；

(3) Determining fiber of each fiber_r' midpoint:

calculating fiber_r' Total number Num of discrete points in, and judge if Num is odd, if yes, will the second

A discrete point as fiber_r' midpoint a_rOtherwise, it will be

A discrete point as fiber_r' midpoint a_rThen the midpoints of all the fibers form a midpoint set Middle, Middle ═ a₁,a₂,...,a_r,...,a_N}，a_r＝(x_w,y_w,z_w)，w∈(s,t)；

(4) Clustering the midpoint set Middle:

clustering the midpoint set Middle by using a noise-based density-based spatial clustering DBSCAN method to obtain MiddleClustering result a of point clustering, a ═ { a ═ a₁,A₂,...,A_e,...,A_EE is more than or equal to 1 and less than or equal to N, wherein A_eIs composed of_eN corresponding to the middle point of each fiber_eGroup e fibers of sliver fibers;

(5) for each fiber class A_eAnd (3) carrying out endpoint clustering:

a is to be_eN in (1)_eThe starting points of the fibers corresponding to the strip fibers are combined into a starting point set Start_eAnd for Start_ePerforming DBSCAN clustering to obtain A_eStarting point clustering result B of_e，B_e＝{(B_e)₁,(B_e)₂,...,(B_e)_f,...,(B_e)_FAt the same time, A is added_eN in (1)_eCombining the fiber end points corresponding to the strip fibers into an end point set Last_eAnd to Last_ePerforming DBSCAN clustering to obtain A_eEnd point clustering result C of_e，C_e＝{(C_e)₁,(C_e)₂,...,(C_e)_g,...,(C_e)_GAnd (c) the step of (c) in which,

1≤F≤n_e，1≤G≤n_e，(B_e)_fis composed of_fM corresponding to each fiber starting point_fType f fibers of sliver fibers, (C)_e)_gIs composed of_gM corresponding to each fiber terminal point_gType g fibers of sliver fibers;

(6) determination of A_eEnd point clustering result of (2):

(6a) judgment B_eOr C_eWhether all discrete points in the data are identified as noise points by DBSCAN clustering or not, if so, A_eThe fiber in (A) is divided into n_eClass, each fiber is self-classified into 1 class, giving A_eEnd point clustering results clusters_eWherein, otherwise, step (6b) is performed;

(6b) judging whether F is equal to G or not, if so, A is_eThe fibers in (A) are classified into 1 type to obtain A_eEnd point clustering ofResults clusters_eOtherwise, executing step (6 c);

(6c) judging whether F is 1 and G is more than 1, if so, A_eThe fibers in (A) are classified into G groups to obtain A_eEnd point clustering results clusters_e，clusters_e＝{(C_e)₁,(C_e)₂,...,(C_e)_g,...,(C_e)_GElse, executing step (6 d);

(6d) judging whether G is 1 and F is more than 1, if so, A_eThe fibers in (A) are classified into F groups to obtain A_eEnd point clustering results clusters_e，clusters_e＝{(B_e)₁,(B_e)₂,...,(B_e)_f,...,(B_e)_FElse, executing step (6 e);

(6e) each (B)_e)_fM in (1)_fCombining the fiber end points corresponding to the strip fibers into an end point set (Last)_e)_fAnd are paired (Last)_e)_fPerforming DBSCAN clustering to obtain B_eEnd point clustering result of (B)_e′，B_e′＝{(I_e)₁,(I_e)₂,...,(I_e)_f′,...,(I_e)_F′At the same time, each (C) is put in_e)_gM in (1)_gThe starting points of the fibers corresponding to the strip fibers are combined into a starting point set (Start)_e)_fAnd on (Start)_e)_fPerforming DBSCAN clustering to obtain C_eStarting point clustering result C of_e′，C_e＝{(J_e)₁,(J_e)₂,...,(J_e)_g′,...,(J_e)_G′And (c) the step of (c) in which,

1≤F′≤n_e，1≤G′≤n_e，(I_e)_f′comprises with m_f′Each fiber end point beingCorresponding m_f′Sliver fiber, (J)_e)_g′Comprises with m_g′M corresponding to each fiber starting point_g′Sliver fiber;

(6f) judging whether F '> G' is true, if yes, A_eThe fibers in (A) are classified into F' groups to give A_eEnd point clustering results clusters_e，clusters_e＝{(I_e)₁,(I_e)₂,...,(I_e)_f′,...,(I_e)_F′Else, A_eThe fiber bundles in (1) are classified into G' groups to obtain A_eEnd point clustering results clusters_e，clusters_e＝{(J_e)₁,(J_e)₂,...,(J_e)_g′,...,(J_e)_G′}；

(7) Obtaining a clustering result of white matter fiber bundles of the brain:

combining the end point clustering results of all the fiber classes in the A to obtain a clustering result Cluster of the fiber bundle track', wherein the clustering result Cluster is { Cluster₁,clusters₂,...,clusters_e,...,clusters_E}, where clusters_eAnd representing the end point clustering result of the e-th fiber.

2. The method for clustering white matter fiber tracts based on fiber midpoints and end points according to claim 1, wherein the step (2b) of calculating each fiber_iAnd ROI₁The implementation steps of the minimum Euclidean distance of (1) are as follows:

(2b1) calculating fiber_iEach discrete point (x) in_d,y_d,z_d) And ROI₁The Euclidean distances of all the discrete points in the image are calculated to obtain a distance set D, D ═ D₁,D₂,...,D_d,...,D_mThe calculation formula is as follows:

wherein (x)_d,y_d,z_d)∈fiber_i，P∈[1,p]，D_d＝{D_d-1,D_d-2,...,D_d-k′,...,D_d-p}，D_d-k′Represents (x)_d,y_d,z_d) To ROI₁The Euclidean distance of the k' th discrete point;

(2b2) calculating fiber_iAll discrete points in to ROI₁The minimum Euclidean distance of (D) to obtain a distance set D_d″，D_d″＝{D₁′,D₂′,...,D_d′,...,D_m' }, the calculation formula is as follows:

D_d′＝min(D_d),

wherein D is_d' means (x)_d,y_d,z_d) To ROI₁Minimum euclidean distance of;

(2b3) calculating fiber_iAnd ROI₁Minimum Euclidean distance D of_d", the calculation formula is as follows:

D_d″′＝min(D_d″)。

3. the method for clustering white matter fiber tracts based on fiber center points and end points according to claim 1, wherein the step (4) of clustering the Middle point set by using a noise density-based spatial clustering DBSCAN method comprises the following steps:

(4a) given a midpoint set Middle, Middle ═ a₁,a₂,...,a_r,...,a_NFind out each midpoint a first_rAnd determines the core object set omega, all of which are compared with a_rIs not more than the middle point of the Euclidean distance to form a_rField of judgment a_rWhether the number of midpoints in the field is greater than or equal to MinPts, if so, a_rIs a core object, otherwise, a_rObtaining a core object set omega of Middle for a non-core object;

(4b) randomly selecting a core object from omega as a seed, and finding out all midpoints that can be reached by the density of the core object, assuming that a_rIs a core object, and selects a_rAs a seed, a midpoint a in Middle_r′Is located at a_rIn the field of (1), then a_r′From a to a_rThe density is direct, and whether a midpoint sequence u exists is judged₁,u₂,...,u_vWherein u is₁＝a_r，u_v＝a_r′And u is_r+1By u_rDensity is up to a_r′From a to a_rThe density can be reached to obtain a cluster A_cluster；

(4c) A is to be_clusterRemoving the core object contained in the intermediate cluster from omega, judging whether omega is empty or not, if yes, combining all cluster classes to obtain a clustering result A of middles, and if not, executing the step (4 b).

Images