Method and system for personalized microstructure difference detection and physiologically associated human brain connection group analysis

Abstract

The invention provides a method and a system for personalized microstructure difference detection and physiologically associated human brain connection group analysis. The method comprises the following steps: converting fibers and diffusion information along the fibers into cosine coefficient matrix representation; generating a FiberMap image of each sample according to the fiber and the fiber diffusion information; a FiberMap image on one group is generated from the groups; comparing the FiberMap image of the single sample with the FiberMap image of the group, and finding a difference fiber bundle; and inputting a FiberMap image of a single sample into the prediction model to obtain correlation fibers of individual features. According to the method, when the FiberMap image of the sample is generated, the position and shape information of the fiber is considered at the same time, the whole brain fiber can be fully represented, the fiber data shape which is not easy to process is converted into the image shape which is easy to process in this way, and fiber analysis becomes very convenient.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and particularly relates to a method and system for personalized microstructural difference detection and physiological-related human brain connectome analysis. Background Art

[0002] In the analysis of brain structure and function, the analysis of neural pathways is particularly crucial. It can not only find the physiological differences between healthy people and brain-damaged patients based on this, but also identify the lesion sites of nervous system diseases. Diffusion magnetic resonance imaging (dMRI) is a non-invasive imaging technique that can provide the structural information of microscopic tissues and obtain more information than traditional anatomical magnetic resonance imaging (MRI). dMRI helps to quantitatively evaluate the microstructure, integrity, and connectivity of white matter tracts, and deeply understand the performance of the brain white matter in healthy people or patients. However, such an analysis unit is the voxel in the image, and does not consider the shape and diffusion characteristics of the fiber as a whole in individual variations. However, the data shape of the fiber makes it difficult to analyze the fiber as a unit and consumes a large amount of computing resources. At present, there is no general method for analyzing the whole-brain fibers. Summary of the Invention

[0003] In order to break through the challenges of the existing technology, the present invention provides a method and system for personalized microstructural difference detection and physiological-related human brain connectome analysis, which combines the whole-brain fibers with diffusion information to generate a fiber grid image, and realizes the difference and correlation analysis based on this image.

[0004] The technical solution for achieving the object of the present invention is as follows:

[0005] A method for personalized microstructural difference detection and physiological-related human brain connectome analysis, comprising:

[0006] Step 1, obtaining the whole-brain fibers and the diffusion characteristics along the fibers through fiber tracking technology;

[0007] Step 2, converting the whole-brain fibers into a first-order cosine coefficient matrix;

[0008] Step 3, converting the diffusion characteristics of the whole-brain along the fibers into a second-order cosine coefficient matrix;

[0009] Step 4, generating the FiberMap image of each sample according to the cosine coefficient matrix of the whole-brain fibers and the second-order cosine coefficient matrix of the fiber diffusion characteristics; for the population sample, generating the FiberMap image at the population level;

[0010] Step 5, based on the FiberMap image of a single sample and the FiberMap image of the population, obtaining the differential fibers of the individual and the population through a differential analysis method;

[0011] Step 6: Query the correlated fibers on the individual through the correlation analysis method based on the FiberMap image of a single sample.

[0012] Furthermore, the dsi-studio or mrtrix fiber tracking technology is adopted.

[0013] Furthermore, the process of converting Step 2 or Step 3 into a cosine coefficient matrix is as follows:

[0014] Let the input sequence be l ∈ R n×d , which contains n points d is the dimension of each point;

[0015] First, define the function f -1 Normalize these points to the range 0 - 1 according to the length ratio in the sequence:

[0016]

[0017] For corresponding one by one to the n sampling points in , where each coordinate is approximated by a cosine expansion of order K as:

[0018]

[0019] where {ψ k} is the cosine basis, and the cosine matrix C = {c ki} ∈ R (K+1)×d , which is estimated by the least squares method.

[0020] Furthermore, the first-order cosine coefficient matrix is 2×3, the first row represents the offset of the fiber, the second row represents the shape of the fiber, and the second-order cosine coefficient matrix is 3×1.

[0021] Furthermore, generating the FiberMap images of each sample specifically includes:

[0022] For the first-order cosine coefficient matrix of the fibers, take all the offsets of the fibers as a point cloud and divide it into uniform 64×64×64 blocks, and take all the shapes of the fibers as another point cloud and divide it into uniform 10×10×10 blocks, to obtain a three-dimensional image representing the fiber offset distribution and a three-dimensional image representing the fiber shape distribution;

[0023] For each block in the obtained offset image, screen the fibers in it, screen the three shapes of fibers with the top three proportions, then take the average of each as the three main shapes of the current block, and at the same time average the diffusion characteristics of the corresponding fibers to obtain the diffusion characteristics corresponding to each of the three main shapes of the current block;

[0024] The length of the cosine coefficient matrix of the diffusion characteristics of each fiber is 3, the total length of the characteristics contained in each block is (3×(3 + 3)) = 18, and the shape of the finally obtained FiberMap image is 64×64×64×18.

[0025] Furthermore, generating the FiberMap image at the population level specifically includes:

[0026] Generate FiberMaps of m samples in the population, Define a new image α p ∈R 64×64×64×18 ;

[0027] For each block α p (i, j, k) in the image, extract the blocks at the same position of all samples in the population Each sample has 3 main shapes, so there are 3m main shapes for m samples. Select three main shapes and take the average as the main shape of the current block α p (i, j, k), and at the same time average the corresponding diffusion characteristics, and then fill the vector with a length of 18 of the result into the current block.

[0028] Furthermore, obtain the differential fibers of an individual from the population through the differential analysis method, specifically including:

[0029] Step 5.1, define the FiberMap images of the samples to be tested and the benchmark as α test ∈R 64×64×64×3×6 and αb ase ∈R 64×64×64×3×6 , and define the coordinates of one of the blocks as:

[0030] i = (a, b, c), a ∈ [0, 63], b ∈ [0, 63], c ∈ [0, 63]

[0031] Define the neighborhood as:

[0032] N(i) = [a - 1, a, a + 1]×[b - 1, b, b + 1]×[c - 1, c, c + 1]

[0033] Step 5.2, for a block i in α test , extract the corresponding neighborhood in α base :

[0034]

[0035] Step 5.3, define the shape similarity threshold θ1 = 0.5 and the diffusion feature similarity threshold θ2 = 0.5, and calculate the k-th row in α test (i) and Base flat(i) The shape similarity of the n-th row is:

[0036]

[0037] Calculate the diffusion feature similarity as:

[0038]

[0039] α test For the fiber bundle corresponding to the k-th row feature in (i), whether it is a differential fiber is calculated as follows:

[0040]

[0041] For all blocks in α test perform the same operations as in step 5.3.

[0042] Furthermore, the specific steps for querying the correlation fibers on an individual through the correlation analysis method include:

[0043] Use the FiberMap image of a single sample as the model input, train a 3D model to fit the individual features, screen out the most significant blocks in the image, and find the corresponding fibers, which are the fibers most relevant to the individual features.

[0044] Furthermore, the process of the 3D model fitting the individual features and screening out the most significant blocks in the image is as follows:

[0045] First, perform shape-invariant convolution on the image to expand the number of channels from 18 to 256, then perform global max pooling on each channel, and finally use a fully connected network to fit the individual features;

[0046] For the result after the model expands the channels, each channel is normalized, and then the sum is taken on the channel dimension to obtain the importance of each block in the image when fitting the individual features.

[0047] A system for personalized microstructural difference detection and physiological-related human connectome analysis includes:

[0048] A fiber tracking unit that obtains the whole-brain fibers and the diffusion features along the fibers through fiber tracking technology;

[0049] A conversion unit that converts the whole-brain fibers into a first-order cosine coefficient matrix and converts the diffusion features along the whole-brain fibers into a second-order cosine coefficient matrix;

[0050] A FiberMap image generation unit that generates FiberMap images of each sample according to the cosine coefficient matrix of the whole-brain fibers and the second-order cosine coefficient matrix of the fiber diffusion features; for population samples, generate FiberMap images at the population level;

[0051] A difference detection unit, based on the FiberMap image of a single sample and the FiberMap image of a population, obtains the differential fibers between an individual and the population through a differential analysis method.

[0052] An association analysis unit, based on the FiberMap image of a single sample, queries the associated fibers on an individual through a correlation analysis method.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] The method for generating a FiberMap image proposed by the present invention simultaneously considers the offset and shape information of fibers and combines the diffusion information of the whole brain, making the information representation ability of this image stronger and applicable to various tasks, such as fiber segmentation, query, and fiber analysis at the whole-brain level; the differential analysis method proposed by the present invention can be used for comparison between any individual and any population, or between individuals, or between populations to find differential fibers, and the results are stable, fast, and consume little computing resources; the correlation analysis method proposed by the present invention is applicable to a variety of different populations and data sets, with significant effects, and can play an auxiliary role in the diagnosis of nervous system diseases. Brief Description of the Drawings

[0055] Figure 1 is a flowchart of the method for analyzing individual microstructural differences and physiological correlations of brain white matter fiber bundles based on diffusion magnetic resonance and deep learning proposed by the present invention.

[0056] Figure 2 is an effect diagram of differential fibers obtained by the method according to the present invention.

[0057] Figure 3 is an effect diagram of associated fibers obtained by the method according to the present invention. Detailed Embodiments

[0058] To enable those skilled in the art to better understand the technical solutions of the present application, the following further describes in detail the overall classification process provided by the present application in conjunction with the drawings and specific embodiments.

[0059] The method for personalized microstructural difference detection and physiological association human connectome analysis proposed by the present invention is as Figure 1 shown and includes:

[0060] Step 1: Data Preparation. Use classical fiber tracking techniques (such as dsi - studio or mrtrix) to obtain whole - brain fibers and diffusion characteristics along the fibers. For a single sample, the data to be prepared includes dwi images, bval files, bvec files, and mask files. Use the command - line or visualization interface of dsi - studio to generate SRC files from whole - brain diffusion magnetic resonance images, then generate FIB files from the SRC files, and then trace the whole - brain fibers from the FIB files and export the diffusion information along the whole - brain fibers.

[0061] The process of using classical fiber tracking techniques (such as dsi - studio or mrtrix) to obtain whole - brain fibers and diffusion characteristics along the fibers is as shown in Figure 1 (a) below and is described as follows:

[0062] (1) Trace the whole - brain fibers;

[0063] (2) Obtain the diffusion characteristics along the fibers according to the whole - brain fibers and the dMRI images of the whole brain.

[0064] Step 2: Cosine Coefficient Representation. Convert the whole - brain fibers into first - order cosine coefficient representation and save them to an npy file. Convert the diffusion characteristics along the fibers into second - order cosine coefficient representation and also save them to an npy file.

[0065] The process of converting a sequence into cosine coefficient representation is as follows:

[0066] (1) Let the input sequence be \(l\in\mathbb{R}\) n×d with \(n\) points and \(d\) being the dimension of each point;

[0067] (2) First, define the function \(f\) -1 to normalize these points to the range 0 - 1 according to their length ratios in the sequence. The formula is as follows:

[0068]

[0069] (3) For corresponding one - by - one to the \(n\) sampling points in , where each coordinate can be approximated using a \(K\) - order cosine expansion as follows:

[0070]

[0071] where \(\{\psi\) k}\) is the cosine basis, and the cosine matrix \(C = \{c\) ki}\in\mathbb{R}\) (K+1)×d is estimated by the least - squares method.

[0072] Step 3: Use the method for generating FiberMap proposed by the present invention to generate a fiber grid image for a single sample. The input is the results of the two cosine coefficient representations obtained in the previous step, and the values in the image are normalized for different channels respectively.

[0073] The process of generating FiberMap for a single sample is as Figure 1 shown in (a) below and is described as follows:

[0074] (1) First, for the cosine coefficient matrix of the fibers, the shape of the cosine coefficient matrix of each fiber is 2×3, where the first row represents the offset of the fiber and the second row represents the shape of the fiber. The offsets of all fibers are divided into a point cloud and partitioned into uniform 64×64×64 blocks, and the shapes of all fibers are divided into another point cloud and partitioned into uniform 10×10×10 blocks. In this way, we obtain a three-dimensional image representing the fiber offset distribution and a three-dimensional image representing the fiber shape distribution, and each block corresponds to a part of the fibers.

[0075] (2) For each block in the offset image obtained in the previous step, screen the fibers in it, screen the fibers of three main shapes, then take the average of each as the three main shapes of the current block, and at the same time, also take the average of the diffusion characteristics of the corresponding fibers to obtain the diffusion characteristics corresponding to each of the three main shapes of the current block.

[0076] (3) In step 2, the length of the diffusion characteristic of each fiber after conversion is 3. Therefore, the total length of the characteristics contained in each block obtained is (3×(3 + 3)) = 18. Thus, the shape of the finally obtained FiberMap image is 64×64×64×18.

[0077] Step 4: Generate a FiberMap image for the population. The input is the FiberMap images of multiple samples used as the reference dataset, and the output is the FiberMap image of the population, with the data shape being the same as that of the individual images.

[0078] The process of generating the population FiberMap is as Figure 1 shown in (b) below and is described as follows:

[0079] (1) First, the FiberMaps of m samples in the population have been generated. At this time, define a new image α p ∈R 64×64×64×18 ;

[0080] (2) For each block in the image, perform the same processing: For a block α p (i,j,k) in the image, extract the blocks at the same position of all samples in the population Each sample has 3 main shapes, and there are 3m main shapes for m samples. Three main shapes are selected from them and averaged to be the current block α. p The main shape of (i, j, k), and at the same time, the corresponding diffusion features are also averaged, and then the vector with a length of 18 of the result is filled into the current block.

[0081] (3) For each block in the image, perform the same processing as in step (2).

[0082] Step 5: Obtain the differential fibers of an individual from the group through the differential analysis method proposed in the present invention. The input reference image is the group image obtained in the previous step, the input test image is the image of the sample to be tested, and the output is the differential fiber file found on the test sample. Some results are as Figure 2 shown.

[0083] The differential analysis method is used to perform differential analysis between two whole-brain FiberMaps. Taking one sample as the reference, differential fibers on the other whole brain are found. The process is as Figure 1 shown in (b) in, and includes:

[0084] (1) Define two FiberMap images α test ∈R 64×64×64×3×6 and α base ∈R 64×64×64×3×6 . Define the coordinates of one block as follows:

[0085] i = (a, b, c), a ∈ [0, 63], b ∈ [0, 63], c ∈ [0, 63]

[0086] Define the neighborhood as follows:

[0087] N(i) = [a - 1, a, a + 1] × [b - 1, b, b + 1] × [c - 1, c, c + 1]

[0088] (2) For a block i in α test , extract the corresponding neighborhood in α base :

[0089]

[0090] (3) For all blocks in α test , perform the same processing: Define the shape similarity threshold θ1 = 0.5 and the diffusion feature similarity threshold θ2 = 0.5, and calculate the shape similarity between the k-th row in α test (i) and the n-th row in Base flat (i) as follows:

[0091]

[0092] The similarity of diffusion features is as follows:

[0093]

[0094] α in the results test The calculation of whether the fiber bundle corresponding to the k-th row feature in (i) is a differential fiber is as follows:

[0095]

[0096] For all blocks in α test Perform the same operations as in step (3) for screening

[0097] Step 6: To query the correlation fibers on an individual through the correlation analysis method proposed by the present invention, it is necessary to prepare the training input, including FiberMap images of multiple samples and the labels on each sample, train the model, then use the trained model to perform inference on the samples, and use the correlation analysis method proposed by the present invention to find the fibers with the greatest correlation and save them as files as the output.

[0098] The correlation analysis process is as shown in Figure 1 (c), and some of the results are as shown in Figure 3 shown and described as follows:

[0099] (1) Use the FiberMap image obtained in step 4 as the model input to train a 3D model to fit the individual features;

[0100] (2) The model used in training first performs shape-invariant convolution on the image, expands the number of channels from 18 to 256, then performs global max pooling on each channel, and finally uses a fully connected network to fit the individual features;

[0101] (3) For the result after the model expands the channels, each channel is normalized, and then the sum is taken on the channel dimension to obtain the importance of each block in the image when fitting the individual features;

[0102] (4) According to the result output by the model, screen out the most significant blocks in the image, and find the corresponding fibers, which are the fibers most relevant to the individual features.

[0103] This embodiment also provides a system for personalized microstructural difference detection and physiological-associated human connectome analysis, including:

[0104] A fiber tracking unit that obtains whole-brain fibers and diffusion features along the fibers through fiber tracking technology;

[0105] A conversion unit that converts the whole-brain fibers into a first-order cosine coefficient matrix and converts the diffusion features along the whole-brain fibers into a second-order cosine coefficient matrix;

[0106] The FiberMap image generation unit generates FiberMap images of each sample according to the cosine coefficient matrix of whole-brain fibers and the second-order cosine coefficient matrix of fiber diffusion characteristics; for a population sample, it generates a FiberMap image at the population level.

[0107] The difference detection unit obtains the differential fibers between an individual and the population through a differential analysis method based on the FiberMap image of a single sample and the FiberMap image of the population.

[0108] The correlation analysis unit queries the correlation fibers of an individual through a correlation analysis method based on the FiberMap image of a single sample.

[0109] When generating the FiberMap image of a sample, the present invention simultaneously considers the position and shape information of the fibers, which can fully represent the whole-brain fibers. In this way, the shape of the fiber data that is not easy to process is converted into an image shape that is easy to process, making fiber analysis extremely convenient.

[0110] The above process is the entire process of the method proposed by the present invention. Those skilled in the art to which the present invention pertains can select appropriate parameters and adjust each step according to needs to adapt to the requirements, but will not deviate from the general idea of the present invention or exceed the scope defined by the appended claims. In addition, although some specific terms are used in this specification, these terms are only for convenience of description and do not constitute any limitation to the present invention.

Claims

1. A method for personalized microstructural difference detection and physiological-related human connectome analysis, characterized in that, Including: Step 1: Obtain the whole-brain fibers and the diffusion characteristics along the fibers through fiber tracking technology; Step 2: Convert the whole-brain fibers into a first-order cosine coefficient matrix; Step 3: Convert the diffusion characteristics of the whole-brain along the fibers into a second-order cosine coefficient matrix; Step 4: Generate the FiberMap images of each sample based on the cosine coefficient matrix of the whole-brain fibers and the second-order cosine coefficient matrix of the fiber diffusion characteristics; For the population samples, generate the FiberMap image at the population level; Step 5: Based on the FiberMap image of an individual sample and the FiberMap image of the population, obtain the differential fibers between the individual and the population through a differential analysis method; Step 6: Based on the FiberMap image of an individual sample, query the correlated fibers of the individual through a correlation analysis method.

2. The method for personalized microstructural difference detection and physiological-related human connectome analysis according to claim 1, characterized in that, Adopt the dsi-studio or mrtrix fiber tracking technology.

3. A method for personalized microstructural difference detection and physiological-related human connectome analysis according to claim 1, characterized in that, The process of converting to the cosine coefficient matrix in Step 2 or Step 3 is as follows: Let the input sequence be \(l\in\mathbb{R}\) n×d , which contains \(n\) points where \(d\) is the dimension of each point; First, define the function f -1 Normalize these points to the range 0 to 1 according to their length ratios in the sequence: For corresponding one by one to the n sampling points in where each coordinate is approximated by a cosine expansion of order K as follows: Among them, {ψ k} is a cosine basis, and the cosine matrix C = {c ki} ∈ R (K+1)×d , which is estimated by the least squares method.

4. A method for personalized microstructural difference detection and physiological-related human connectome analysis according to claim 1, characterized in that The first-order cosine coefficient matrix is 2×3. The first row represents the offset of the fiber, and the second row represents the shape of the fiber. The second-order cosine coefficient matrix is 3×1.

5. A method for personalized microstructural difference detection and physiological-related human connectome analysis according to claim 4, characterized in that Specifically, generating the FiberMap images of each sample includes: For the first-order cosine coefficient matrix of the fibers, take the offsets of all fibers as a point cloud and divide it into evenly sized 64×64×64 blocks, and take the shapes of all fibers as another point cloud and divide it into evenly sized 10×10×10 blocks, to obtain a three-dimensional image representing the fiber offset distribution and a three-dimensional image representing the fiber shape distribution; For each block in the obtained offset image, screen the fibers in it, select the top three fiber shapes in terms of proportion, then take the average of each as the three main shapes of the current block, and at the same time, also take the average of the corresponding diffusion characteristics of the fibers, to obtain the diffusion characteristics corresponding to the three main shapes of the current block respectively; The length of the cosine coefficient matrix of the diffusion characteristics of each fiber is 3, the total length of the characteristics contained in each block is (3×(3 + 3)) = 18, and the final shape of the obtained FiberMap image is 64×64×64×18.

6. A method for personalized microstructural difference detection and physiological-related human connectome analysis according to claim 5, characterized in that, Specifically, generating the FiberMap image at the population level includes: Generate the FiberMap of m samples in the population, Define a new image α p ∈R 64 ×64×64×18 ; For each block α in the image p (i, j, k), extract the blocks at the same position of all samples in the population Each sample has 3 main shapes. For m samples, there are 3m main shapes. Select three main shapes from them and take the average as the main shape of the current block α p (i, j, k), and also average the corresponding diffusion features. Then fill the vector of length 18 of the result into the current block.

7. A method for personalized microstructure difference detection and physiological-related human connectome analysis according to claim 6, characterized in that, Obtaining the differential fibers between the individual and the population through a differential analysis method specifically includes: Step 5.1, define the FiberMap images of the samples to be tested and the reference as α test ∈R 64×64×64×3×6 and α base ∈R 64 ×64×64×3×6 , define the coordinates of one of the blocks as: i = (a, b, c, a ∈ [0, 63], b ∈ [0, 63], c ∈ [0, 63] Define the neighborhood as: N(i) = [a - 1, a, a + 1] × [b - 1, b, b + 1] × [c - 1, c, c + 1] Step 5.2, for a block i in α test , extract the corresponding neighborhood in α base : Step 5.3, define the shape similarity threshold θ1 = 0.5 and the diffusion feature similarity threshold θ2 = 0.5, and calculate α test the k-th row in (i) and Base flat the shape similarity between the n-th row in (i) and Base is: Calculate the diffusion feature similarity as: α test Whether the fiber bundle corresponding to the feature in the \(k\)-th row of (i) is a differential fiber is calculated as follows: For α test Perform the same operations as in step 5.3 for all blocks in 8. A method for personalized microstructural difference detection and physiological-related human connectome analysis according to claim 6, characterized in that, Specifically, querying the correlated fibers of the individual through a correlation analysis method includes: Use the FiberMap image of an individual sample as the model input, train a 3D model to fit the individual characteristics, screen out the most significant blocks in the image, and find the corresponding fibers, which are the fibers most relevant to the individual characteristics.

9. A method for personalized microstructural difference detection and physiological-related human connectome analysis according to claim 5, characterized in that The process of the 3D model fitting the individual characteristics and screening out the most significant blocks in the image is: First, perform convolution on the image with invariant shape to expand the number of channels from 18 to 256. Then, perform global max pooling on each channel. Finally, use a fully connected network to fit individual features; For the result after expanding the channels of the model, standardize each channel and then sum in the channel dimension to obtain the importance of each block in the image when fitting individual features.

10. A system for personalized microstructural difference detection and physiological-related human connectome analysis for implementing the method according to any one of claims 1-9, characterized in that, Including: A fiber tracking unit that obtains whole-brain fibers and diffusion features along the fibers through fiber tracking technology; A conversion unit that converts whole-brain fibers into a first-order cosine coefficient matrix and converts the diffusion features of the whole brain along the fibers into a second-order cosine coefficient matrix; A FiberMap image generation unit that generates FiberMap images of each sample based on the cosine coefficient matrix of the whole-brain fibers and the second-order cosine coefficient matrix of the fiber diffusion features; For population samples, generate FiberMap images at the population level; A difference detection unit that obtains the differential fibers between an individual and the population through a differential analysis method based on the FiberMap image of an individual sample and the FiberMap image of the population; A correlation analysis unit that queries the correlation fibers of an individual through a correlation analysis method based on the FiberMap image of an individual sample.

Images