White matter fiber multi-dimensional geometrical morphology quantification method and system based on magnetic resonance imaging

Abstract

The invention provides a white matter fiber multi-dimensional geometrical morphology quantification method and system based on magnetic resonance imaging, and the method comprises the steps: reconstructing diffusion magnetic resonance imaging data of a whole-brain white matter fiber bundle, generating a three-dimensional fiber track, setting the sampling spacing of the fiber track according to a resolution ratio requirement, and carrying out the reconstruction of the diffusion magnetic resonance imaging data of the whole-brain white matter fiber bundle; sampling is carried out through a sub-voxel-level fiber trajectory interpolation algorithm; carrying out hierarchical iterative integral calculation point by point along a fiber track by adopting a sliding window through a path signature method, and generating a local path signature feature in each window; the path signature features of the whole-brain white matter fibers are emitted to the original magnetic resonance image, and a three-dimensional geometric feature map meeting the resolution requirement is generated. According to the method, through a multi-order tensor analysis and super-resolution track feature mapping technology of a path signature (PS), the problems that the spatial resolution of traditional diffusion indexes (such as an anisotropic fraction (FA) and an average diffusivity) is insufficient and a high-order geometric quantization tool is lacked are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fiber image processing, and in particular relates to a method and system for quantifying the multi-dimensional geometric morphology of white matter fibers based on magnetic resonance imaging. Background Art

[0002] The geometric morphological analysis of white matter fibers is an important means to reveal the structural mechanism of the brain, especially in the field of diagnosis and treatment of brain diseases. However, its methodology still has the following two significant limitations: (1) The lack of high-order geometric quantification tools. The existing geometric analysis methods of white matter fibers lack the ability to systematically quantify high-order features, resulting in insufficient comprehensive analysis of the multi-scale geometric characteristics of white matter fibers, limiting the depth of research on the structural mechanism of white matter. (2) The spatial resolution bottleneck of diffusion indicators in traditional diffusion magnetic resonance imaging (dMRI). The current mainstream dMRI technology relies on scalar indicators such as fractional anisotropy (FA) and mean diffusivity for white matter analysis. Limited by the resolution limitations of clinical equipment (typical voxel volume 2-3mm 3 ), these indices are not sensitive enough to characterize white matter fiber tracts. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging. By using multi-order tensor analysis of path signatures (PS) and super-resolution trajectory feature mapping methods, this method breaks through the spatial resolution limitations of traditional diffusion indices, achieves collaborative quantification from the perspectives of multidimensional geometric morphology quantification and super-resolution, and improves the sensitivity of quantitative measurement of brain white matter fiber structure.

[0004] The technical solutions for achieving the purpose of the present invention are:

[0005] A method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging, comprising:

[0006] Step 1: Reconstruct the diffusion magnetic resonance imaging data of the whole-brain white matter fiber bundles to generate three-dimensional fiber trajectories. Set the fiber trajectory sampling interval according to the resolution requirement and perform subvoxel-level fiber trajectory interpolation sampling.

[0007] Step 2: Using the path signature method, a sliding window is used to perform hierarchical iterative integral calculation point by point along the fiber trajectory to generate the local path signature feature within each window;

[0008] Step 3: Project the path signature features of the whole-brain white matter fibers onto the original magnetic resonance image to generate a three-dimensional geometric feature map that meets the resolution requirements.

[0009] Furthermore, the path signature feature is a multi-order geometric tensor including displacement, curvature, torsion and high-order nonlinear dynamics.

[0010] Furthermore, the path signature feature is a 3rd-order geometric tensor.

[0011] Furthermore, the first-order geometric tensor of the path signature feature is:

[0012]

[0013] in, represents the sampling path at time t, i=1,...,d, d=3, represents the d-dimensional space, express The cumulative linear displacement in [a,t].

[0014] Furthermore, the second-order geometric tensor of the path signature feature is:

[0015]

[0016] Where i, j = 1, ..., d, Reflects the path self-interaction, i≠j, Reflects the curvature tensor.

[0017] Furthermore, the third-order geometric tensor of the path signature feature is:

[0018]

[0019] Where i, j, k = 1, ..., d, Characterize high-order geometric deformations.

[0020] Furthermore, the path signature features of the whole-brain white matter fibers are projected into the original magnetic resonance image, including: distinguishing the positive and negative values ​​of the same type of path signature feature values ​​within the same voxel, and writing them into the original image space on an average.

[0021] Furthermore, the fiber trajectory sampling interval is z times the voxel size, 0 <z<1。

[0022] Furthermore, the sliding window length is 3 sampling points and the step length is 1 sampling point.

[0023] A system for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging, comprising:

[0024] The 3D fiber trajectory generation and sampling unit reconstructs the diffusion MRI data of the whole-brain white matter fiber bundles to generate 3D fiber trajectories. The fiber trajectory sampling interval is set according to the resolution requirements, and sampling is performed using a subvoxel-level fiber trajectory interpolation algorithm.

[0025] The path signature unit uses a path signature method to perform hierarchical iterative integral calculation point by point using a sliding window along the fiber trajectory to generate local path signature features within each window;

[0026] The three-dimensional geometric feature map generation unit projects the path signature features of the whole-brain white matter fibers to the original magnetic resonance image to generate a three-dimensional geometric feature map that meets the resolution requirements.

[0027] Compared with the existing technology, the beneficial effects of the present invention are as follows: (1) The PS method is used to perform hierarchical iterative integral calculation on the reconstructed whole-brain white matter fiber trajectory to generate a multi-order geometric tensor containing displacement, curvature, torsion and high-order nonlinear dynamics, where the k-th order tensor represents the k-level geometric evolution characteristics of the trajectory, realizing the multi-dimensional morphological feature quantification of white matter fibers. (2) Based on the continuous differentiable characteristics of the fiber trajectory in space, the original dMRI data is super-reconstructed by the subvoxel trajectory interpolation algorithm, and the geometric features are mapped to ≤1mm by combining the affine invariance of the path signature. 3 High-resolution images, breaking through the inherent voxels (2-3mm) of clinical dMRI equipment 3 ) resolution limitation; (3) The present invention solves the problem of insufficient spatial resolution (2-3mm) of traditional diffusion indicators (such as fractional anisotropy (FA) and average diffusivity) by using multi-order tensor analysis of path signature (PS) and super-resolution trajectory feature mapping technology. 3 ) and the lack of high-order geometric quantization tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of the method for quantifying the multidimensional geometric morphology of white matter fibers of the present invention.

[0029] Figure 2 This figure shows the results of analyzing the differences in white matter structure between males and females in a large sample data set.

[0030] Figure 3 Cross-sectional comparison of feature maps constructed for original resolution and super-resolution.

[0031] Figure 4 This is the distribution map of white matter fiber bundles that undergo significant changes before and after SSRI treatment.

[0032] Figure 5It is a correlation result graph of the change rate of the average PS value in the white matter region before and after SSRI treatment and the reduction rate of the Hamilton Depression Rating Scale (HAMD-17). Detailed implementation manners

[0033] Combined with Figure 1 , a white matter fiber multi-dimensional geometric shape quantification method based on magnetic resonance imaging provided by the present invention includes the following steps:

[0034] Step 1: Reconstruct the diffusion magnetic resonance imaging data of the whole-brain white matter fiber bundle to generate three-dimensional fiber trajectories, set the spacing of fiber trajectory sampling according to the resolution requirement, and sample the fiber trajectories through a sub-voxel-level fiber trajectory interpolation algorithm;

[0035] Sample the whole-brain white matter fibers, and the sampling spacing parameter is selected as z times the voxel size (0 < z < 1, the smaller the value, the higher the resolution of the PS feature image in the subsequent steps. The specific parameter setting is based on the specific experimental purpose. It is set to 0.5 in this step).

[0036] Step 2: Through the path signature method, perform hierarchical iterative integration calculations point by point along the fiber trajectories using a sliding window to generate local path signature features within each window;

[0037] The window length = 3 sampling points, and the step size = 1.

[0038] For the discrete sampling path of the white matter fiber trajectory in three-dimensional space Its path signature S(X) a,t ​​​​​​​​​​​​​​​​​​​​​​​​​​

[0047] Capturing the helical twist and asymmetric interactions of three-dimensional fibers (such as S ijk ≠S ikj ), characterizing high-order geometric deformation.

[0048] N-order recursive formula:

[0049]

[0050] Finally, the path signature is an infinite sequence S(X) a,b =(1,{S i},{S ij},{S ijk}, ...), the present invention adopts a signature truncated to the third order (39 items) to balance computational efficiency and feature discrimination.

[0051] In the present invention, the N parameter is set to 1, 2, and 3 respectively.

[0052] For the path segment p1→p2→p3 within the window, calculate the third-order truncation signature and obtain the local PS value set of the window:

[0053]

[0054] A total of 39 PS values ​​were generated, covering linear displacement (items 1-3), curvature (items 4-12), and torsion (items 13-39).

[0055] Step 3: Project the path signature features of the whole-brain white matter fibers onto the original magnetic resonance image to generate a three-dimensional geometric feature map that meets the resolution requirements.

[0056] Map the PS values ​​of the whole brain fibers to the original image space. To obtain images of different resolutions, different resolution sampling can be performed in the original image space. The same type of PS values ​​within the same voxel are distinguished as positive and negative, and the average is written into the original image space. Finally, multiple PS feature maps corresponding to different signature items are generated:

[0057]

[0058] Among them, v represents a voxel, p represents the number of line sampling points contained in v, The number of punctual points contained in v, The number of punctual points contained in v, S n (p) is the nth value in the PS value set. Finally, 78 feature maps (PS_1_pos to PS_39_neg, Note: 1-39 correspond to the numerical items in step 3; pos and neg refer to positive and negative images respectively) are generated, supporting 2×2×3mm 3 , 1mm 3and higher resolution image output.

[0059] The present invention also provides a system for quantifying the multidimensional geometric morphology of white matter fibers for implementing the method, comprising:

[0060] The 3D fiber trajectory generation and sampling unit reconstructs the diffusion MRI data of the whole-brain white matter fiber bundles to generate 3D fiber trajectories. The fiber trajectory sampling interval is set according to the resolution requirements, and sampling is performed using a subvoxel-level fiber trajectory interpolation algorithm.

[0061] The path signature unit uses a path signature method to perform hierarchical iterative integral calculation point by point using a sliding window along the fiber trajectory to generate local path signature features within each window;

[0062] The three-dimensional geometric feature map generation unit projects the path signature features of the whole-brain white matter fibers to the original magnetic resonance image to generate a three-dimensional geometric feature map that meets the resolution requirements.

[0063] Experiment 1

[0064] Based on the above method, this example conducted a large dataset analysis of the differences in white matter structure between men and women, including:

[0065] 1. Data collection and preprocessing

[0066] This implementation uses two public human brain imaging datasets. The discovery dataset is the S1200 dataset from the Human Connectome Project (HCP), which includes 3T dMRI and T1-weighted (T1w) images from 1,065 subjects (490 males and 575 females). The external testing dataset is the Chinese Human Connectome Project (CHCP).

[0067] The dMRI and T1 data scanning parameters in the HCP dataset are as follows: dMRI scan parameters are TE = 89.5 ms, TR = 5520 ms, FOV = 210 mm × 180 mm, spatial resolution 1.25 mm × 1.25 mm × 1.25 mm, and image size 145 × 174 × 145 mm. Each dMRI scan sequence includes 18 b = 0 images and 270 diffusion-weighted images acquired at b = 1000, 2000, and 3000 s / mm². T1w data scan parameters are TE = 2.14 ms, TR = 2400 ms, FOV = 224 mm × 224 mm, and spatial resolution 0.7 mm × 0.7 mm × 0.7 mm. The dMRI data used in this patent download were processed using the designed HCP preprocessing protocol. The specific preprocessing process includes non-brain tissue removal, motion correction, eddy current correction, EPI distortion correction, and registration of the dMRI with the corresponding T1w images.

[0068] After excluding subjects with missing 3T dMRI, T1w, and T2w images in the CHCP dataset, a total of 326 subjects (153 males and 173 females, age: 34.37 ± 18.73 years) were included in the study. Male and female subjects were age-matched (p > 0.05). CHCP dMRI scan parameters used were TE = 86 ms, TR = 3500 ms, FOV = 210 mm × 210 mm, spatial resolution 1.5 mm × 1.5 mm × 1.5 mm, and image size 140 × 140 × 100. Each subject received two dMRI images with opposite phase encoding directions (i.e., AP and PA). T1w data scan parameters were TE = 2.22 ms, TR = 2400 ms, FOV = 256 mm × 240 mm, and spatial resolution 0.8 mm × 0.8 mm × 0.8 mm. Scan parameters for the T2w data were TE = 563 ms, TR = 3200 ms, FOV = 256 mm × 240 mm, and a spatial resolution of 0.8 mm × 0.8 mm × 0.8 mm. The downloaded raw image data were preprocessed using Mrtrix3 and fsl (version 6.0.5). Both T1w and T2w images were preprocessed to remove non-brain tissue. dMRI preprocessing steps included denoising, removal of Gibbs ring artifacts, merging of AP and PA data, EPI correction using T2w images, bias field correction, gradient table correction, and alignment of the dMRI with the T1w images.

[0069] 2. Whole-brain fiber reconstruction and sampling

[0070] The SD_STREAM algorithm, implemented using Mrtrix3 software, provided directional information for fiber tracking using the intravoxel fiber orientation distribution (FOD) model estimated from the dMRI signal. A fourth-order Runge-Kutta integration algorithm was used to smooth the resulting streamlines. The minimum length of the tracked fibers was 20 mm and the maximum was 250 mm. The tracking step size was set to 0.3125 mm, and the FOD amplitude threshold for stopping tracking was set to 0.06. After reconstructing whole-brain white matter fiber tracts, spherical-deconvolution-informed filtering of tractograms (SIFT) was used to filter out false-positive connections. Two SIFT operations were performed on the whole-brain fiber tracts reconstructed by SD_STREAM. The first SIFT operation filtered the whole-brain fiber tracts to 1 million fibers, and the second SIFT operation filtered the fibers to 500,000 fibers. Finally, all fiber streamlines were resampled with a step size of 0.625 mm (half the voxel size).

[0071] 3. PS calculation and super-resolution mapping

[0072] Along each fiber streamline, a sliding window (window length 3 sampling points, step length 1 point) is used to extract local path segments and geometric features are calculated in stages. The first-order features quantify the linear displacement (Δx, Δy, Δz) of the streamline segments, reflecting the stretching or contraction of the fiber; the second-order features analyze the path curvature and covariance to capture the bending or bifurcation morphology of the fiber; the third-order features analyze the nonlinear torsional effects (such as spiral reorganization) to characterize high-order geometric deformation. The 39 PS features (including positive and negative values) of each sampling point are mapped to the 3D image space to generate a 1mm 3 The PS feature map of the resolution is obtained, and the positive and negative eigenvalues ​​in the same voxel are averaged to preserve the direction information.

[0073] 4. Analysis of differences in white matter structure between men and women

[0074] Each brain image was divided into 216 brain regions based on the JHU-ICBM 10 and AAL3v111 atlases. The average value of each brain region was used as a classification feature. A gender classification model was constructed based on the average values ​​of PS maps and FA for 216 brain regions. The accuracy of four different classifiers (support vector classifier (SVC), Gaussian process classifier (GPC), random forest (RF), and logistic regressor (LR)) was compared using first-, second-, and third-order PS features, as well as FA features. The 1065-case HCP data set was used as the discovery set, and the 326-case CHCP data set was used as an independent external test set. This patent performed inner-outer nested 10-fold cross-validation on the discovery set data, with the outer 9-fold data set serving as the training set and the inner 1-fold data set serving as the test set. In each round of the outer 10-fold cross-validation, lasso feature selection and classification model parameter adjustment were performed using the training set from that round. After the optimal model for that round was determined, the classification accuracy of the model was reported on the internal test set. The discovery set data is randomly shuffled for 10 rounds and the above operation is repeated, that is, each classification model can obtain the classification results of the 100 best models trained on the discovery set.

[0075] A key step before external validation of a model is to minimize nonbiological variability in data from different sites due to factors such as differences in scanners and acquisition protocols. In this experiment, the ComBat method was used to coordinate the features calculated from the CHCP to the feature space of the HCP. The 100 best model hyperparameters for each of the four classifiers were averaged as the final classification model parameters. High-frequency features that were selected more than 50 times in 100 rounds of lasso selection were used as model inputs, and the final model was retrained on the discovery set. Finally, the accuracy of the final model is reported on the external test set.

[0076] 5. Results

[0077] Figure 2 As shown, on both the discovery set and the external test set, the classification accuracy of the model based on path signatures was significantly higher than that of FA features. Furthermore, the classification accuracy of second- and third-order path signatures was generally higher than that of first-order signatures. This was consistent across all four classifiers. Compared to the FA metric, at their respective optimal results, this method achieved a 28% improvement in correct identification accuracy, demonstrating superior sensitivity.

[0078] Experiment 2

[0079] Based on the above method, this example conducted a longitudinal analysis of white matter fiber images in patients with depression before and after 8 weeks of antidepressant SSRI drug treatment, including:

[0080] 1. Data collection and preprocessing

[0081] Patients with major depressive disorder (MDD) who met the DSM-IV diagnostic criteria were recruited. The condition was confirmed by two independent physicians, and the severity of symptoms was quantified using the Hamilton Depression Rating Scale (HAMD-17). Patients were required to meet clinical remission criteria (HAMD-17 score <7) after completing 12 weeks of SSRI treatment to be included in the analysis. Subsequently, a 3T magnetic resonance imaging device (Siemens Verio) was used for image acquisition, including T1-weighted structural imaging and diffusion magnetic resonance imaging (dMRI). The T1 image parameters were set to TR = 1900ms, TE = 2.48ms, and a spatial resolution of 1mm×0.9766mm×0.9766mm; the dMRI parameters included TR = 6600ms, TE = 93ms, and a voxel resolution of 1.875×1.875×3mm 3 , collect 1 b=0 image and 270 b=1000s / mm 2 Diffusion-weighted image of .

[0082] The acquired dMRI data were standardized, including denoising based on random matrix theory, eddy current correction implemented by the FSL toolkit, and Gibbs artifact removal to ensure that the image quality met the analysis requirements. A convolutional neural network was then used to generate a mask to accurately segment the brain region and remove non-brain tissue signals. The T1 image was aligned with the dMRI data through rigid registration, preserving the original diffusion space coordinate system.

[0083] 2. Whole-brain fiber reconstruction and sampling

[0084] The probabilistic algorithm iFOD2 was used for whole-brain fiber tracking. Three million seed points, the Runge-Kutta fourth-order integration method, and angle thresholds of 15 and 30 degrees were set to generate white matter fiber bundles. The bundles were optimized to 1 million streamlines through SIFT filtering and resampled to 0.6 mm resolution to improve trajectory accuracy.

[0085] 3. PS calculation and super-resolution mapping

[0086] Along each fiber streamline, a sliding window (window length 3 sampling points, step size 1 point) is used to extract local path segments, and geometric features are calculated in stages. The first-order features quantify the linear displacement (Δx, Δy, Δz) of the streamline segments, reflecting the stretching or contraction of the fiber; the second-order features analyze the path curvature and covariance to capture the bending or bifurcation morphology of the fiber; the third-order features analyze the nonlinear torsional effects (such as spiral reorganization) to characterize high-order geometric deformation. The 39 PS features (including positive and negative values) of each sampling point are mapped to the 3D image space to generate the original resolution (1.875×1.875×3mm 3 ) and super resolution (1mm 3 )’s PS feature map (the PS feature results are as follows Figure 3 shown).

[0087] 4. Longitudinal analysis of patients with depression before and after 8 weeks of SSRI treatment

[0088] Based on the JHU-ICBM white matter atlas, 50 white matter regions were divided and the PS (1mm) of each region before and after treatment was calculated. 3 The mean of the PS feature (PS) was calculated. Paired-sample t-tests were used to analyze differences in PS feature values ​​before and after treatment, with false discovery rate (FDR) correction for multiple comparisons. The significance threshold was set at p < 0.05. The Spearman rank correlation coefficient between the rate of change in PS feature value (percentage change in mean after treatment) and the rate of decrease in HAMD-17 score was further calculated to identify white matter regions significantly associated with symptom improvement.

[0089] 5. Results

[0090] Figure 4 The areas showing significant longitudinal changes in patients with depression before and after 8 weeks of antidepressant SSRI drug treatment were mainly distributed in the transverse pontine tract (TPT), the left anterior limb of the internal capsule (ALIC), and the left corticospinal tract (CST). Figure 5 The results show a significant correlation between the PS value change rate of these main areas and the HAMD-17 reduction rate.

[0091] In a longitudinal analysis of patients with depression before and after 8 weeks of SSRI antidepressant drug treatment, this method showed excellent sensitivity in characterizing changes in brain white matter structure, and these changes were significantly correlated with data that can characterize clinical treatment effects (maximum r = 0.41, p = 0.0025).

[0092] This invention converts the biological characteristics of white matter fibers (continuous and differentiable) into an algorithm through mathematical modeling. Without upgrading the hardware, it can extract more precise geometric information from existing clinical dMRI data, providing a new technical path for neurological disease research and precision medicine.

[0093] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the claims.

Claims

1. A method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging, characterized in that: include: Step 1: Reconstruct diffusion MRI data of whole-brain white matter fiber bundles to generate three-dimensional fiber trajectories. Set the fiber trajectory sampling interval according to the resolution requirement and perform sampling using a subvoxel-level fiber trajectory interpolation algorithm. Step 2: Using the path signature method, a sliding window is used to perform hierarchical iterative integral calculation point by point along the fiber trajectory to generate the local path signature feature within each window; Step 3: Project the path signature features of the whole-brain white matter fibers onto the original magnetic resonance image to generate a three-dimensional geometric feature map that meets the resolution requirements.

2. The method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging according to claim 1, characterized in that: The path signature features are multi-order geometric tensors that include displacement, curvature, torsion, and high-order nonlinear dynamics.

3. The method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging according to claim 2, characterized in that: The path signature feature is a 3rd-order geometric tensor.

4. The method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging according to claim 3, characterized in that: The first-order geometric tensor of the path signature feature is: in, represents the sampling path at time t, i=1,...,d, d=3, represents the d-dimensional space, express The cumulative linear displacement in [a,t].

5. The method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging according to claim 4, characterized in that: The second-order geometric tensor of the path signature feature is: in, Reflecting the path self-interaction, Reflects the curvature tensor.

6. The method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging according to claim 5, characterized in that: The third-order geometric tensor of the path signature feature is: in, Characterize high-order geometric deformations.

7. The method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging according to claim 1, characterized in that: The path signature features of the whole brain white matter fibers are projected into the original magnetic resonance image, including: distinguishing the positive and negative values ​​of the same type of path signature feature values ​​in the same voxel, and writing them into the original image space on an average basis.

8. The method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging according to claim 1, characterized in that: The fiber trajectory sampling interval is z times the voxel size, 0 <z<1。 9. The method for quantifying the multidimensional geometric morphology of white matter fibers based on magnetic resonance imaging according to claim 8, characterized in that: The sliding window length is 3 sampling points, and the step length is 1 sampling point.

10. A system for quantifying the multidimensional geometric morphology of white matter fibers for implementing the method according to any one of claims 1 to 9, characterized in that: include: The 3D fiber trajectory generation and sampling unit reconstructs the diffusion MRI data of the whole-brain white matter fiber bundles to generate 3D fiber trajectories. The fiber trajectory sampling interval is set according to the resolution requirements, and sampling is performed using a subvoxel-level fiber trajectory interpolation algorithm. The path signature unit uses a path signature method to perform hierarchical iterative integral calculation point by point using a sliding window along the fiber trajectory to generate local path signature features within each window; The three-dimensional geometric feature map generation unit projects the path signature features of the whole-brain white matter fibers to the original magnetic resonance image to generate a three-dimensional geometric feature map that meets the resolution requirements.

Images