Intelligent auxiliary diagnosis method for Alzheimer disease
Through rotary invariant spherical harmony feature decomposition and whole-brain fiber bundle tracking technology, combined with white matter subregion clustering and composite nuclear support vector machine, the problems of insufficient sensitivity of subtle lesions and multicenter diagnostic deviation in the early diagnosis of Alzheimer's disease are solved, and high-precision Alzheimer's disease diagnosis is achieved.
Patent Information
- Application Number
- CN202510453994.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-11
- Publication Date
- 2025-08-12
AI Technical Summary
The prior art lacks sensitivity to subtle white matter lesions in the early diagnosis of Alzheimer's disease, and the diagnosis deviation caused by multicenter data heterogeneity, traditional methods are difficult to effectively eliminate the risk of misdiagnosis caused by device parameter differences.
A nonlinear mapping model was established using rotary invariant spherical harmony feature decomposition technology, combining whole-brain fiber bundle tracking and white matter subregion clustering, and a cross-center high-precision classification model was constructed through multimodal registration and edge beam analysis, and a composite nuclear support vector machine was used for diagnosis.
Accurate positioning and high-precision diagnosis of early Alzheimer's disease lesions are achieved, reducing the risk of misdiagnosis caused by parameter differences in multi-center equipment, and improving the robustness and accuracy of the diagnosis.
Smart Images
Figure CN120473115A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical image processing technology, and in particular to a multi-center diffusion magnetic resonance data coordination and white matter microstructure analysis method, which is suitable for a computer-aided diagnosis method for neurodegenerative diseases such as Alzheimer's disease. Background Art
[0002] Alzheimer's disease (AD), a degenerative disease of the central nervous system characterized by insidious cognitive decline, has yet to be fully elucidated. Early, non-invasive diagnosis faces significant challenges. Traditional studies have suggested that AD pathological changes are primarily associated with gray matter atrophy. However, recent studies have shown that microstructural abnormalities in white matter fiber networks may be a core feature of the early pathological progression of AD, providing new avenues for the development of disease biomarkers. Diffusion Magnetic Resonance Imaging (dMRI) can non-invasively assess the integrity of white matter fiber microstructure by detecting the anisotropic properties of water molecule diffusion. Indicators such as fractional anisotropy (FA) and mean diffusivity (MD) derived from diffusion tensor imaging (DTI) have been widely used to quantify white matter damage in AD patients. Clinical evidence shows that AD patients have significantly reduced FA values and increased MD values in key white matter pathways such as the corpus callosum, cingulate gyrus, and uncinate fasciculus, suggesting myelin degeneration and loss of axonal integrity. However, existing technologies are mostly limited to global analysis of average indicators of the entire bundle, which makes it difficult to reveal the spatially heterogeneous damage patterns within the fiber bundles. In addition, mainstream automated fiber tracking technologies (such as automated fiber quantification, AFQ) can only identify 20-30 main fiber bundles, resulting in insufficient sensitivity to early subtle lesions. In addition, in multicenter clinical applications, the systematic deviation of FA values (typical deviation 0.05-0.08) caused by differences in scanning equipment parameters (such as the number of gradient directions and b-value settings) in different medical institutions may introduce a misdiagnosis risk of up to 30%. Traditional data reconciliation methods (such as linear normalization) can eliminate site technical variation, but they tend to over-smooth the true biological differences between individuals, weakening the generalization ability of the diagnostic model. Summary of the Invention
[0003] To overcome the insufficient sensitivity of traditional methods to subtle white matter lesions in the early stages of Alzheimer's disease (AD) and the diagnostic bias caused by multi-center data heterogeneity, the present invention provides an intelligent auxiliary diagnosis method for Alzheimer's disease, forming a full-chain solution from data standardization to intelligent diagnosis, providing a reliable technical tool for early and precise intervention of AD.
[0004] The technical solution of the present invention is:
[0005] An intelligent auxiliary diagnosis method for Alzheimer's disease, comprising the following steps:
[0006] 101. Data Acquisition: Multicenter diffusion MRI data were collected, covering scanning sites from different medical institutions. Multicenter data coordination was achieved by constructing a difference model between reference sites and target sites.
[0007] 102. Data preprocessing: De-noising, distortion correction, and calculation of diffusion tensor imaging indicators such as fractional anisotropy and mean diffusivity were performed using FMRIB software tools;
[0008] 103. Multi-center Data Harmonization: Based on a standardized algorithm for diffusion MRI data, a nonlinear mapping model is established from the target site dataset to the reference site dataset using rotationally invariant spherical harmonic eigendecomposition. The trained mapping model is applied to all subjects at the target site to reconstruct the harmonized dMRI signals, eliminating cross-site device parameter differences while retaining 95% of individual biological variation.
[0009] 104. Whole-brain fiber tract tracing: Using whole-brain fiber tract tracing technology and unscented Kalman filter double tensor model to perform whole-brain fiber tract tracing and generate streamlines;
[0010] 105. White matter tract substructure identification: anatomical clustering of 800 white matter subregions was achieved through the automatic annotation fiber clustering framework for white matter analysis;
[0011] 106. Along-bundle feature calculation: Perform along-bundle analysis on each fiber cluster, divide the streamline into 100 equal segments and remove 15% of the segments at both ends that are easily affected by partial volume effects, and extract the segmental FA / MD mean features;
[0012] 107. Statistical analysis: The t-test with Bonferroni correction was used to screen the regions with significant differences.
[0013] 108. Classification model construction and validation: Combine minimum redundancy maximum relevance feature selection with composite kernel support vector machine classifier to achieve high-precision classification across centers, and use ROC / AUC to evaluate performance.
[0014] Furthermore, in 103, the process of multi-center data coordination is as follows:
[0015] 201. RISH feature extraction: Perform spherical harmonic decomposition on the dMRI data of the matched subjects and calculate the rotationally invariant spherical harmonic feature RISH, whose mathematical expression is:
[0016]
[0017] Where l is the order of the spherical harmonic function, which determines the directional resolution of the feature. In dMRI, l is an even number, l = 0, 2, 4, 6, 8. The symmetry of the diffusion signal makes the odd-order coefficients approach zero. m is the azimuthal quantum number, which is used to describe the azimuthal distribution of the spherical harmonic function on the sphere. a lm Spherical harmonic coefficients: coefficients obtained from dMRI data by spherical harmonic decomposition, representing the intensity component of the diffusion signal in a specific direction;
[0018] 202. Nonlinear Mapping Learning: A random forest regression algorithm is used to establish a cross-site feature mapping model from the target site dataset to the reference site dataset. The mean square error (MSE) is controlled to < 0.05 through 5-fold cross-validation to ensure the accuracy of nonlinear relationship fitting.
[0019] 203. Individual Data Harmonization: Apply the learned mapping model to the target site data, reconstruct the harmonized dMRI signal, and simultaneously generate the corrected bval / bvec files to eliminate the differences in gradient direction and diffusion sensitivity factor b-value. The calculation formula is:
[0020]
[0021] in, is the coordinated dMRI signal, is the coordinated rotationally invariant spherical harmonic feature, is the spherical harmonic basis function: the superscript 0 indicates the case of magnetic quantum number m = 0, the subscript l indicates the order of the spherical harmonic function; θ indicates the zenith angle, ranging from 0 to π; represents the azimuth angle, ranging from 0 to 2π; θ and Used to determine direction in space;
[0022] 204. Quality Verification: The coordination effect is verified by the following three indicators: Jensen-Shannon divergence, i.e. JSD ≤ 0.15, quantifying the distribution consistency between sites; biological variation index BVI ≥ 3.0, ensuring that the individual difference retention rate is > 95%; FA offset threshold ΔFA < 0.03, controlling the indicator deviation introduced by technical variation.
[0023] Furthermore, in steps 104 to 106, the white matter analysis process is as follows:
[0024] 301. Fiber Tracking: Using whole-brain fiber tracking technology and the UKF dual-tensor model, whole-brain fiber tract reconstruction was performed, generating approximately 350,000 streamlines covering key white matter pathways such as the corpus callosum and uncinate fasciculus;
[0025] 302. Subregion Identification: White matter anatomical subregion segmentation is achieved through a standardized processing pipeline, and whole-brain fiber tracts are divided into 800 anatomically defined subregions through multimodal registration;
[0026] Along-beam analysis: A single streamline was evenly divided into 100 anatomical nodes, and a kd-tree spatial index was established to match the DTI voxel coordinates. 15% of the segments at each end that were susceptible to partial volume effects were removed, and the middle 70% of valid nodes (k = 16-84) were retained. The mean segmental FA / MD ratio was calculated.
[0027] In the above 108, the classification model combines mRMR feature selection with a composite kernel SVM classifier (SVM classifier) to achieve cross-center high-precision classification, and the SVM classifier includes a radial basis and a linear kernel.
[0028] This invention achieves breakthroughs through the following core technologies: First, a nonlinear mapping algorithm based on Rotationally Invariant Spherical Harmonic (RISH) is used to decompose the diffusion signal into direction-dependent (fiber orientation) and direction-independent (microstructural integrity) components, and establish an accurate correction model for cross-site device parameter differences (FA value offset ≤ 0.03, Jensen-Shannon divergence ≤ 0.15), while eliminating technical variation while retaining more than 95% of individual biological differences (biological variation index ≥ 3.0). Secondly, combining a whole-brain white matter cluster atlas (WMA) with tractometric analysis, multimodal registration was used to partition whole-brain fiber tracts into 800 anatomically defined subregions. A parametric streamline resampling method was then used to segment individual fiber tracts into 100 equally spaced segments. After removing 15% of segments susceptible to partial volume effects at both ends, the mean values of the segmental fractional anisotropy (FA) and mean diffusivity (MD) were calculated. This enabled millimeter-level localization of localized microstructural lesions in key pathways (such as the anterior uncinate fasciculus and posterior cingulate gyrus). The resulting multimodal fusion diagnostic model, based on a modified minimum redundancy maximum relevance (mRMR) feature selection algorithm and a composite kernel support vector machine (SVM), achieved high-accuracy classification across centers. Performance was evaluated using receiver operating characteristic (ROC) and area under the curve (AUC). This comprehensive solution, from multicenter data collection to RISH coordination to subregional clustering to tractometric analysis to intelligent diagnosis, provides a reliable technical tool for early and precise intervention in AD.
[0029] The beneficial effects of the present invention are: forming a full-chain solution from data standardization to intelligent diagnosis, and providing reliable technical tools for early and precise intervention of AD. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1This is a specific flow chart of an intelligent auxiliary diagnosis method for Alzheimer's disease according to an embodiment of the present invention.
[0031] Figure 2 yes Figure 1 Flowchart of an embodiment of a multi-center data coordination system in an embodiment.
[0032] Figure 3 yes Figure 1 A flowchart of an embodiment of a white matter analysis process in an embodiment of the present invention. DETAILED DESCRIPTION
[0033] The present invention will be further described below with reference to the accompanying drawings.
[0034] Figure 1 This is a specific flow chart of an intelligent auxiliary diagnosis method for Alzheimer's disease according to an embodiment of the present invention.
[0035] like Figure 1 As shown, an intelligent auxiliary diagnosis method for Alzheimer's disease based on multi-center diffusion magnetic resonance data coordination and white matter tract microstructure analysis according to an embodiment of the present invention includes:
[0036] 101. Data Collection: Multicenter Diffusion Magnetic Resonance Imaging (dMRI) data were collected, covering scanning sites in different medical institutions. Multicenter data coordination was achieved by constructing a difference model between the reference site (as a cross-site benchmark) and the target site. The specific process required that the number of subjects at both the reference site and the target site be no less than 30. The matching criteria included an age difference of less than 1.5 years and a chi-square test value of χ2 for gender distribution. 2 <3.84 (p>0.05), and the two-sample t-test verified that there was no significant difference between the groups.
[0037] 102. Data preprocessing: De-noising, distortion correction and DTI index (FA / MD) calculation were performed using FMRIB Software Library (FSL).
[0038] 103. Multi-center data harmonization: Based on the dMRI harmonization algorithm and utilizing the rotationally invariant spherical harmonic features (RISH) decomposition technique, a nonlinear mapping model is established from the target site dataset to the reference site dataset. The trained mapping model is applied to all subjects at the target site to reconstruct the harmonized dMRI signals, thereby eliminating cross-site device parameter differences (Jensen-Shannon divergence ≤ 0.15) while retaining 95% of individual biological variation.
[0039] like Figure 2As shown, the process of multi-center data coordination according to the embodiment shown in the figure is:
[0040] 201. RISH feature extraction: Perform spherical harmonic decomposition on the dMRI data of the matched subjects and calculate the rotationally invariant spherical harmonic feature (RISH), whose mathematical expression is:
[0041]
[0042] Where l is the order of the spherical harmonic function, which determines the directional resolution of the feature. In dMRI, l is an even number, l = 0, 2, 4, 6, 8. The symmetry of the diffusion signal makes the odd-order coefficients approach zero. m is the azimuthal quantum number, which is used to describe the azimuthal distribution of the spherical harmonic function on the sphere. a lm Spherical harmonic coefficients: The coefficients obtained from dMRI data by spherical harmonic decomposition represent the intensity components of the diffusion signal in a specific direction.
[0043] 202. Nonlinear Mapping Learning: A random forest regression algorithm is used to establish a cross-site feature mapping model from the target site dataset to the reference site dataset. The mean squared error (MSE<0.05) is controlled through 5-fold cross-validation to ensure the accuracy of nonlinear relationship fitting.
[0044] 203. Individual Data Harmonization: Apply the learned mapping model to the target site data, reconstruct the harmonized dMRI signal, and simultaneously generate the corrected bval / bvec files to eliminate the differences in gradient direction and diffusion sensitivity factor b-value. The calculation formula is:
[0045]
[0046] in, is the coordinated dMRI signal, is the coordinated rotationally invariant spherical harmonic feature, is the spherical harmonic basis function: the superscript 0 indicates the case of magnetic quantum number m = 0, the subscript l indicates the order of the spherical harmonic function; θ indicates the zenith angle, ranging from 0 to π; represents the azimuth angle, ranging from 0 to 2π; θ and Used to determine orientation in space.
[0047] 204. Quality Verification: The coordination effect was verified by the following three indicators: Jensen-Shannon divergence (JSD ≤ 0.15), which quantifies the consistency of distribution between sites; biological variation index (BVI ≥ 3.0), which ensures that the individual difference retention rate is > 95%; FA shift threshold (ΔFA < 0.03), which controls the indicator deviation introduced by technical variation.
[0048] 104. Whole-brain fiber tract tracing: Whole-brain fiber tract tracing is performed using whole-brain fiber tract tracing techniques, such as the Unscented Kalman Filter (UKF) dual tensor model, generating 350,000 streamlines.
[0049] 105. White matter tract substructure identification: anatomical clustering of 800 white matter subregions was achieved through the automatic annotation fiber clustering framework of white matter analysis (WMA);
[0050] 106. Along-bundle feature calculation: Perform along-bundle analysis on each fiber cluster, divide the streamline into 100 equal segments and remove 15% of the segments at both ends that are easily affected by partial volume effects, and extract the segmental FA / MD mean features;
[0051] like Figure 3 As shown, according to the white matter analysis process (104-106) of the embodiment shown in the figure, the white matter analysis process is as follows:
[0052] 301. Fiber Tracking: Whole-brain fiber tracking techniques, such as the UKF dual-tensor model, are used to reconstruct whole-brain fiber bundles, generating approximately 350,000 streamlines covering key white matter pathways such as the corpus callosum and uncinate fasciculus.
[0053] 302. Subregion Identification: White matter anatomical subregion segmentation is achieved through a standardized processing pipeline, and whole-brain fiber tracts are divided into 800 anatomically defined subregions through multimodal registration.
[0054] Along-beam analysis: A single streamline was evenly divided into 100 anatomical nodes, and a kd-tree spatial index was established to match the DTI voxel coordinates. 15% of the segments at each end that were susceptible to partial volume effects were removed, and the middle 70% of valid nodes (k = 16-84) were retained. The mean segmental FA / MD ratio was calculated.
[0055] 107. Statistical analysis: The t-test with Bonferroni correction was used to screen the regions with significant differences.
[0056] 108. Classification Model Construction and Validation: Combining mRMR feature selection with a composite kernel SVM classifier (radial basis + linear kernel) to achieve high-accuracy classification across centers, using ROC / AUC performance evaluation. This provides a comprehensive solution from data standardization to intelligent diagnosis, providing a highly robust technology platform for precision medicine in AD.
[0057] The embodiments of this specification are merely examples of implementations of the invention and are provided for illustrative purposes only. The scope of protection of the present invention should not be considered limited to the specific embodiments described in these embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by a person of ordinary skill in the art based on the invention.
Claims
1. An intelligent auxiliary diagnosis method for Alzheimer's disease, characterized in that: The method comprises the following steps:
101. Data Acquisition: Multicenter diffusion MRI data were collected, covering scanning sites from different medical institutions. Multicenter data coordination was achieved by constructing a difference model between reference sites and target sites.
102. Data preprocessing: De-noising, distortion correction, and calculation of diffusion tensor imaging indicators such as fractional anisotropy and mean diffusivity were performed using FMRIB software tools; 103. Multi-center Data Harmonization: Based on a standardized algorithm for diffusion MRI data, a nonlinear mapping model is established from the target site dataset to the reference site dataset using rotationally invariant spherical harmonic eigendecomposition. The trained mapping model is applied to all subjects at the target site to reconstruct the harmonized dMRI signals, eliminating cross-site device parameter differences while retaining 95% of individual biological variation.
104. Whole-brain fiber tract tracing: Using whole-brain fiber tract tracing technology and unscented Kalman filter double tensor model to perform whole-brain fiber tract tracing and generate streamlines; 105. White matter tract substructure identification: anatomical clustering of 800 white matter subregions was achieved through the automatic annotation fiber clustering framework for white matter analysis; 106. Along-bundle feature calculation: Perform along-bundle analysis on each fiber cluster, divide the streamline into 100 equal segments and remove 15% of the segments at both ends that are easily affected by partial volume effects, and extract the segmental FA / MD mean features; 107. Statistical analysis: The t-test with Bonferroni correction was used to screen the regions with significant differences.
108. Classification model construction and validation: Combine minimum redundancy maximum relevance feature selection with composite kernel support vector machine classifier to achieve high-precision classification across centers, and use ROC / AUC to evaluate performance.
2. The intelligent auxiliary diagnosis method for Alzheimer's disease according to claim 1, characterized in that: In 103, the process of multi-center data coordination is as follows:
201. RISH feature extraction: Perform spherical harmonic decomposition on the dMRI data of the matched subjects and calculate the rotationally invariant spherical harmonic feature RISH, whose mathematical expression is: Where l is the order of the spherical harmonic function, which determines the directional resolution of the feature. In dMRI, l is an even number, l = 0, 2, 4, 6, 8. The symmetry of the diffusion signal makes the odd-order coefficients approach zero. m is the azimuthal quantum number, which is used to describe the azimuthal distribution of the spherical harmonic function on the sphere. a lm Spherical harmonic coefficients: coefficients obtained from dMRI data by spherical harmonic decomposition, representing the intensity component of the diffusion signal in a specific direction; 202. Nonlinear Mapping Learning: A random forest regression algorithm is used to establish a cross-site feature mapping model from the target site dataset to the reference site dataset. The mean square error (MSE) is controlled to < 0.05 through 5-fold cross-validation to ensure the accuracy of nonlinear relationship fitting.
203. Individual Data Harmonization: Apply the learned mapping model to the target site data, reconstruct the harmonized dMRI signal, and simultaneously generate the corrected bval / bvec files to eliminate the differences in gradient direction and diffusion sensitivity factor b-value. The calculation formula is: in, is the coordinated dMRI signal, is the coordinated rotationally invariant spherical harmonic feature, is the spherical harmonic basis function: the superscript 0 indicates the case of magnetic quantum number m = 0, the subscript l indicates the order of the spherical harmonic function; θ indicates the zenith angle, ranging from 0 to π; represents the azimuth angle, ranging from 0 to 2π; θ and Used to determine direction in space; 204. Quality Verification: The coordination effect is verified by the following three indicators: Jensen-Shannon divergence, i.e. JSD ≤ 0.15, quantifying the distribution consistency between sites; biological variation index BVI ≥ 3.0, ensuring that the individual difference retention rate is > 95%; FA offset threshold ΔFA < 0.03, controlling the indicator deviation introduced by technical variation.
3. The intelligent auxiliary diagnosis method for Alzheimer's disease according to claim 1 or 2, characterized in that: In 104-106, the white matter analysis process is as follows:
301. Fiber Tracking: Using whole-brain fiber tracking technology and the UKF dual-tensor model, whole-brain fiber tract reconstruction was performed, generating approximately 350,000 streamlines covering key white matter pathways such as the corpus callosum and uncinate fasciculus; 302. Subregion Identification: White matter anatomical subregion segmentation is achieved through a standardized processing pipeline, and whole-brain fiber tracts are divided into 800 anatomically defined subregions through multimodal registration; Along-beam analysis: A single streamline was evenly divided into 100 anatomical nodes, and a kd-tree spatial index was established to match the DTI voxel coordinates. 15% of the segments at each end that were susceptible to partial volume effects were removed, and the middle 70% of valid nodes (k = 16-84) were retained. The mean segmental FA / MD ratio was calculated.
4. The intelligent auxiliary diagnosis method for Alzheimer's disease according to claim 1 or 2, characterized in that: In the above 108, the classification model combines mRMR feature selection with a composite kernel SVM classifier (SVM classifier) to achieve cross-center high-precision classification, and the SVM classifier includes a radial basis and a linear kernel.