An early image detection method for Alzheimer's disease based on single-mode MRI

Abstract

This invention discloses an early image detection method for Alzheimer's disease based on monomodal MRI, belonging to the field of image detection technology. The method includes: setting intensity correction parameters based on quality control marker information; performing intensity correction, non-brain tissue stripping, and gray matter separation on uniform structural MRI images to generate gray matter probability images and gray matter uncertainty maps; performing standard spatial alignment on the gray matter probability images and gray matter uncertainty maps, and segmenting them according to anatomical brain region rules to generate a set of brain region gray matter distribution sub-images and brain region uncertainty constraint information; calculating brain region abnormality scores using the set of brain region gray matter distribution sub-images, and using the brain region uncertainty constraint information to suppress false detections, generating a brain region abnormality score map and a set of suspected structural abnormality regions; using reversible reverse verification to screen out false abnormalities and perform credibility convergence, outputting credible abnormality region images and early detection conclusions. This invention enhances the interpretability and consistency of image-level evidence.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, and in particular to a method for early image detection of Alzheimer's disease based on single-modal MRI. Background Technology

[0002] In the early auxiliary diagnosis of Alzheimer's disease, image detection of monomodal structural MRI often involves using a unified imaging sequence followed by format parsing and spatial normalization, then intensity correction and non-brain tissue dissection, and gray matter separation. Subsequently, the gray matter probabilistic images are aligned to a standard space and brain regions are divided according to anatomical brain region templates. Abnormal responses are constructed at the brain region scale and potential structural abnormality areas are presented with heat distribution, thus forming image evidence and risk indications for clinical interpretation and follow-up assessment.

[0003] The standard procedure faces two challenges in the preclinical and mild cognitive impairment stages: First, gray matter separation and registration are sensitive to noise and individual differences in the boundary region, and abnormal responses are easily mixed with false abnormalities caused by uncertainty, which affects the control of false detections; Second, the spatial morphology of abnormal regions is greatly affected by local score fluctuations, and without reverse verification and iterative stability constraints for abnormal candidates, the stability and verifiability of the detection conclusions are insufficient under repeated processing or parameter changes. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an early image detection method for Alzheimer's disease based on single-modal MRI to solve the problem of false anomalies caused by uncertainties in gray matter separation and registration in the early stage, resulting in insufficient stability and verifiability of detection conclusions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for early image detection of Alzheimer's disease based on single-modal MRI, which includes acquiring raw structural MRI images of a single sequence of the subject, performing normalization processing, and generating uniform structural MRI images and quality control marker information. Intensity correction parameters are set based on quality control marker information, and intensity correction, non-brain tissue stripping, and gray matter separation are performed on uniform structure MRI images to generate gray matter probability images and gray matter uncertainty maps. The gray matter probability image and gray matter uncertainty map are aligned in standard space and segmented according to the rules of anatomical brain regions to generate a set of sub-images of gray matter distribution in brain regions and uncertainty constraint information of brain regions. Brain region abnormality scores are calculated using a set of brain region gray matter distribution sub-images, and false detections are suppressed using brain region uncertainty constraint information, generating brain region abnormality score maps and sets of suspected structural abnormality regions; Reversible reverse verification is used to screen out false anomalies and perform credibility convergence, outputting images of credible anomaly regions and early detection conclusions.

[0007] As a preferred embodiment of the Alzheimer's disease early image detection method based on single-modal MRI described in this invention, the steps for generating uniform structured MRI images are as follows: The original structural MRI images of a single sequence of the subject are acquired and the acquisition time and sequence identification information are written in. At the same time, sequence consistency is verified and a structural MRI baseline image is generated. The structural MRI baseline images are processed for format parsing, coordinate orientation unification, and standardization to generate uniform structural MRI images.

[0008] As a preferred embodiment of the early image detection method for Alzheimer's disease based on monomodal MRI described in this invention, the quality control label information is obtained by quality assessment of uniform structure MRI images.

[0009] As a preferred embodiment of the early image detection method for Alzheimer's disease based on monomodal MRI described in this invention, the steps of performing intensity correction, non-brain tissue stripping, and gray matter separation on uniform structural MRI images to generate gray matter probability images are as follows. Intensity correction parameters are set based on quality control marker information, and intensity correction is performed on uniform structural MRI images to generate grayscale reference structural MRI images. Non-brain tissue dissection was performed on grayscale baseline MRI images to obtain brain parenchyma range constraint information, and brain parenchyma regions were extracted based on the brain parenchyma range constraint information to generate brain parenchyma structure MRI images. Gray matter tissue separation was performed on MRI images of brain parenchyma structures to generate gray matter probability images.

[0010] As a preferred embodiment of the early image detection method for Alzheimer's disease based on monomodal MRI described in this invention, the gray matter uncertainty map is obtained by using a perturbation reseparation and comparison strategy to evaluate the gray matter consistency of the gray matter probability image and statistically analyze the gray matter probability fluctuation.

[0011] As a preferred embodiment of the early image detection method for Alzheimer's disease based on monomodal MRI described in this invention, the steps of segmenting the brain according to anatomical brain region rules to generate a set of sub-images of gray matter distribution in the brain region and brain region uncertainty constraint information are as follows: The gray-quality probability image and the gray-quality uncertainty map are aligned in a standard space to generate a standard gray-quality probability image and a standard gray-quality uncertainty map. Based on the rules of anatomical brain regions, standard gray matter probability images are segmented into brain regions to generate a set of sub-images of gray matter distribution in brain regions. The standard gray matter uncertainty map is mapped and aggregated according to the brain region mask corresponding to the set of brain region gray matter distribution sub-images to generate brain region uncertainty constraint information.

[0012] As a preferred embodiment of the early image detection method for Alzheimer's disease based on monomodal MRI described in this invention, the anatomical brain region rule refers to numbering the brain tissue in the standard space according to the partition boundaries of the preset brain region template and generating corresponding brain region masks, and then performing brain region division on the standard gray matter probability image.

[0013] As a preferred embodiment of the early image detection method for Alzheimer's disease based on monomodal MRI described in this invention, the steps for suppressing false detections and generating a brain region abnormality scoring map using brain region uncertainty constraint information are as follows: The abnormal response intensity of each brain region is calculated based on the set of sub-images of gray matter distribution in the brain region. The spatial heat distribution is obtained through voxel-level mapping and normalized and calibrated to generate an initial abnormality score map of the brain region. Based on the uncertainty constraint information of brain regions, the credibility suppression and sensitivity calibration of the initial abnormality scoring map of the brain region are performed to obtain the abnormality scoring map of the brain region.

[0014] As a preferred embodiment of the early image detection method for Alzheimer's disease based on monomodal MRI described in this invention, the set of suspected structural abnormality regions is obtained by locating abnormal peak regions according to the brain region abnormality scoring map and applying spatial coherence constraints.

[0015] As a preferred embodiment of the Alzheimer's disease early image detection method based on single-modal MRI described in this invention, the steps of using reversible reverse verification to screen out false anomalies and perform confidence convergence to output credible anomaly region images and early detection conclusions are as follows. Based on the brain region abnormality scoring map, local structural restoration and reconstruction are performed on the candidate regions corresponding to the suspected structural abnormality regions to generate structural restoration images. Structural consistency difference assessment is performed on the structural reconstruction image and the gray matter probability image, and pseudo-anomaly regions in the suspected structural anomaly region set are screened out to generate an anomaly candidate set; The anomaly candidate set is corrected by neighborhood consistency propagation, the credibility of the anomaly candidate set is updated round by round, an anomaly credibility map is generated and credibility convergence is performed, and a credible anomaly region image and early detection conclusion are generated.

[0016] The beneficial effects of this invention are as follows: a perturbation-reseparation-comparison strategy is used to generate a gray matter uncertainty map, which transforms gray matter probability fluctuations into quantifiable uncertainty constraints and performs credibility suppression and sensitivity calibration during the formation of the brain region abnormality scoring map, making the candidate generation of the suspected structural abnormality region set more focused on stable abnormal responses; then, local structural restoration and reconstruction are performed on the candidate regions corresponding to the suspected structural abnormality region set to obtain structural restoration images, and false abnormal regions are screened out by combining structural consistency difference evaluation, and then the convergence of the abnormality credibility map is driven by neighborhood consistency propagation correction, outputting credible abnormality region images and early detection conclusions, thereby enhancing the interpretability and verification consistency of image-level evidence. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of early image detection for Alzheimer's disease based on monomodal MRI.

[0019] Figure 2 This is a flowchart for standard spatial alignment and brain region segmentation.

[0020] Figure 3 A flowchart for calculating brain region abnormality scores and identifying suspected structural abnormality areas.

[0021] Figure 4 This is a brain region abnormality scoring map. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for early image detection of Alzheimer's disease based on single-modal MRI, including the following steps: S1. Acquire raw structural MRI images of a single sequence from the subject and perform standardized processing to generate uniform structural MRI images and quality control marker information. The original structural MRI images of a single sequence of the subject are acquired and the acquisition time and sequence identification information are written in. At the same time, sequence consistency is verified and a structural MRI baseline image is generated. Furthermore, after acquiring raw structural MRI images of a single sequence from the subject, the acquisition time and sequence identifier are extracted from the image header information and written into the corresponding image record. The consistency of the imaging sequence type, slice thickness, and number of slices are verified based on the sequence identifier and acquisition parameters, and the verification conclusions are associated with the raw structural MRI images. The raw structural MRI images containing the acquisition time, sequence identifier, and verification conclusions are packaged into structural MRI reference images. The structural MRI reference images then proceed to subsequent format parsing, coordinate orientation unification, and standardization processing steps.

[0026] The structural MRI baseline images are processed for format parsing, coordinate orientation unification, and standardization to generate uniform structural MRI images.

[0027] Furthermore, when performing format parsing on the structural MRI reference image, the image pixel matrix and spatial resolution information are read and the sequence identifier consistency is verified. The image storage format conversion is completed and a format parsing record is formed. When performing coordinate orientation unification on the structural MRI reference image, the axial direction is rearranged according to the image head orientation information and the position of the coordinate origin is corrected to obtain a coordinate orientation unified structural MRI reference image. When performing normalization processing on the coordinate orientation unified structural MRI reference image, the intensity range is standardized and the target voxel size is resampled while maintaining the continuity of the anatomical structure to generate a unified structural MRI image. A structural reference with consistent spatial coordinates and grayscale scale is obtained, so that subsequent quality assessment and intensity correction can be performed under a unified standard.

[0028] Quality assessment of uniform structural MRI images is performed to generate quality control marker information.

[0029] Furthermore, by calculating the ratio of the standard deviation of the background region to the mean of the brain tissue region and taking its reciprocal, a signal-to-noise ratio (SNR) quantitative assessment is performed on the uniform structure MRI image to generate a SNR numerical label. Based on the SNR numerical label and the uniform structure MRI image, motion artifact detection is performed to generate a motion artifact level label. Combined with the motion artifact level label, the inter-slice intensity consistency test is performed on the uniform structure MRI image to generate an inter-slice consistency label. Based on the inter-slice consistency label, the gray-white matter boundary contrast is calculated on the uniform structure MRI image to generate a contrast label. Referring to the contrast label, the integrity of brain tissue coverage is analyzed on the uniform structure MRI image to generate a coverage integrity label. The single label with the lowest quality level among the SNR numerical label, motion artifact level label, inter-slice consistency label, contrast label, and coverage integrity label is used as the overall quality level label of the uniform structure MRI image.

[0030] It should be noted that the motion artifact detection and generation of motion artifact level labels based on signal-to-noise ratio (SNR) numerical labels and uniform structure MRI images are carried out as follows: The mutual information entropy value between adjacent slices in the uniform structure MRI image is calculated; the lower the mutual information entropy value, the worse the similarity between slices, and the higher the probability of motion artifacts. A fast Fourier transform is performed on the uniform structure MRI image to analyze the distribution density and spatial location of abnormal high-frequency components in the frequency domain; areas of abnormal concentration of high-frequency components indicate the presence of motion artifacts. The frequency domain analysis results are weighted and corrected according to the SNR numerical labels. When the signal-to-noise ratio (SNR) numerical label is low, the sensitivity to high-frequency components is enhanced; when the SNR numerical label is high, the sensitivity is appropriately reduced to avoid false detections. The mutual information entropy value, high-frequency component density, and SNR weighted result are fused to obtain a comprehensive motion artifact score. The degree of motion artifacts is divided into four levels by a grading threshold set based on the statistical characteristics of motion artifact distribution in images of healthy elderly people. For example, a comprehensive motion artifact score less than 0.2 is classified as level 0, 0.2 to 0.4 as level 1, 0.4 to 0.7 as level 2, and greater than or equal to 0.7 as level 3.

[0031] S2. Set intensity correction parameters based on quality control marker information, and perform intensity correction, non-brain tissue stripping, and gray matter separation on uniform structure MRI images to generate gray matter probability images and gray matter uncertainty maps. Intensity correction parameters are set based on quality control marker information, and intensity correction is performed on uniform structural MRI images to generate grayscale reference structural MRI images. Furthermore, based on the quality control marker information, signal-to-noise ratio (SNR) numerical markers are extracted, and based on the SNR numerical markers, an inverse mapping method is used to obtain the intensity correction smoothing coefficient. Based on the quality control marker information, motion artifact level markers are extracted, and based on the motion artifact level markers, the intensity correction offset order is determined. The intensity correction smoothing coefficient and the intensity correction offset order are combined to form intensity correction parameters. The intensity correction parameters are used to perform N4 offset correction on the uniform structure MRI image to generate an offset-corrected structure MRI image. The offset-corrected structure MRI image is then matched with a standard grayscale template using grayscale histogram matching to generate a grayscale matched structure MRI image. Finally, the grayscale matched structure MRI image is subjected to global grayscale linear stretching to generate a grayscale reference structure MRI image.

[0032] It should be noted that the inverse mapping method refers to directly calculating the intensity correction smoothing coefficient by multiplying the reciprocal of the signal-to-noise ratio (SNR) value by a preset scaling factor. The lower the SNR value, the higher the intensity correction smoothing coefficient, and vice versa, forming a monotonically decreasing linear correspondence.

[0033] Non-brain tissue dissection was performed on grayscale baseline MRI images to obtain brain parenchyma range constraint information, and brain parenchyma regions were extracted based on the brain parenchyma range constraint information to generate brain parenchyma structure MRI images. Furthermore, the BET brain tissue extraction algorithm is used to perform skull dissection calculations on the grayscale reference structure MRI image to obtain an initial brain tissue mask. Three-dimensional morphological closure operation is performed on the initial brain tissue mask to fill the internal holes and form a complete brain tissue mask. The complete brain tissue mask is defined as the brain parenchyma range constraint information. The brain parenchyma range constraint information and the grayscale reference structure MRI image are subjected to a voxel-by-voxel logical AND operation to retain the voxel gray values ​​marked as 1 and set the gray values ​​of the remaining voxels to zero. The output is a brain parenchyma structure MRI image containing only the grayscale distribution of the brain parenchyma region. The brain parenchyma structure MRI image eliminates interference from non-brain tissues such as the skull and scalp, providing a pure brain tissue input basis for subsequent gray matter separation.

[0034] It should be noted that the BET brain tissue extraction algorithm refers to a skull dissection method based on a deformable model. This method calculates the initial spherical contour of the brain tissue surface from grayscale reference MRI images, iteratively shrinks the contour in the opposite direction of the grayscale gradient until it reaches the intensity boundary at the junction of cerebrospinal fluid and skull, retains the voxels inside the contour and removes the voxels outside the contour to generate the initial brain tissue mask.

[0035] Three-dimensional morphological closing operation refers to first performing a three-dimensional spherical structural element expansion operation on the initial brain tissue mask to fill the small holes and broken areas inside the mask, and then performing a three-dimensional spherical structural element erosion operation of the same radius on the expansion result to restore the original boundary scale of the mask, eliminating the discontinuity inside the mask while maintaining the overall morphology of the brain tissue and generating a complete brain tissue mask.

[0036] Brain parenchyma range constraint information refers to the three-dimensional binary voxel-labeled image obtained by converting a complete brain tissue mask. In this image, voxels with a value of 1 accurately mark the spatial location range of the brain parenchyma region within the MRI image of the brain parenchyma structure, while voxels with a value of 0 mark non-brain parenchyma regions. The brain parenchyma range constraint information provides a spatial range limit for the subsequent gray matter separation step.

[0037] Gray matter tissue separation is performed on MRI images of brain parenchyma structures to generate gray matter probability images; Furthermore, the MRI images of brain parenchymal structures are registered to the MNI standard space to obtain spatially normalized MRI images of brain parenchymal structures. The spatially normalized MRI images of brain parenchymal structures are then fused with the gray matter prior probability map in the MNI standard space using voxel-level Bayesian fusion calculation to obtain an initial estimate of the gray matter posterior probability. The initial estimate of the gray matter posterior probability is then applied to the spatial smoothing constraint of the Hidden Markov Random Field to optimize neighborhood consistency and generate an optimized gray matter posterior probability image. The optimized gray matter posterior probability image is then inversely transformed back to the original space of the MRI images of brain parenchymal structures to generate a gray matter probability image.

[0038] A perturbation-based reseparation and comparison strategy is adopted to assess gray quality consistency in gray quality probability images and statistically analyze gray quality probability fluctuations, thereby generating a gray quality uncertainty map.

[0039] Furthermore, a perturbation re-separation and contrast strategy was adopted to add Gaussian noise perturbation to the MRI images of brain parenchymal structures to generate perturbed brain parenchymal structure MRI images. The same gray matter tissue separation process as the gray matter probability image was performed on the perturbed brain parenchymal structure MRI images to generate perturbed gray matter probability images. The process of adding Gaussian noise perturbation and gray matter tissue separation was repeated ten times to generate ten perturbed gray matter probability images. The gray matter probability images and the ten perturbed gray matter probability images were successively subtracted at the voxel level at the same spatial location to obtain ten sets of gray matter probability fluctuation sequences. The standard deviation of the ten sets of gray matter probability fluctuation sequences was calculated at each voxel position to quantify the gray matter probability fluctuation amplitude and generate a gray matter probability fluctuation standard deviation image. The gray matter probability fluctuation standard deviation image was Gaussian smoothed and filtered to eliminate isolated noise points and generate a gray matter uncertainty map.

[0040] It should be noted that the perturbation-reseparation-comparison strategy is an uncertainty assessment method for quantifying the stability of gray matter segmentation results. Its core idea is to indirectly measure the confidence of the segmentation boundary by injecting small, controllable perturbations into the input image and repeatedly executing the same segmentation process, observing the fluctuation of the output results. "Perturbation" refers to superimposing Gaussian noise of a level consistent with natural noise onto the grayscale of the brain parenchyma MRI image. "Reseparation" refers to repeatedly performing gray matter tissue separation calculations on the perturbated image that are completely consistent with the original gray matter probability image. "Comparison" refers to statistically analyzing the difference in voxel-level probability values ​​between the original gray matter probability image and the perturbation gray matter probability image generated by multiple perturbation-reseparation at the same anatomical location. This strategy ensures that the perturbation intensity is within the range of common noise in clinical images by setting the standard deviation of Gaussian noise to 5% of the global grayscale standard deviation of the brain parenchyma MRI image, while setting the number of repetitions to 10 to achieve a balance between statistical stability and computational efficiency. Finally, a gray matter uncertainty map is generated by the probability fluctuation standard deviation to reflect the segmentation confidence of the local gray matter boundary.

[0041] S3. Perform standard spatial alignment on the gray matter probability image and the gray matter uncertainty map, and segment according to the rules of anatomical brain regions to generate a set of sub-images of gray matter distribution in brain regions and uncertainty constraint information of brain regions. The gray-quality probability image and the gray-quality uncertainty map are aligned in a standard space to generate a standard gray-quality probability image and a standard gray-quality uncertainty map. Furthermore, affine registration calculations are performed between the grayscale probability image and the MNI152 standard spatial template to generate an affine transformation matrix. The affine transformation matrix is ​​then applied to the grayscale probability image to generate an affine-aligned grayscale probability image. Nonlinear registration calculations are then performed between the affine-aligned grayscale probability image and the MNI152 standard spatial template to generate a nonlinear deformation field. Based on the nonlinear deformation field, a third-order spline interpolation spatial transformation is performed on the affine-aligned grayscale probability image to generate a standard grayscale probability image. Simultaneously, the affine transformation matrix and the nonlinear deformation field are sequentially applied to the grayscale uncertainty map to generate an affine-aligned grayscale uncertainty map, which is then further transformed to generate a standard grayscale uncertainty map.

[0042] Based on the rules of anatomical brain regions, standard gray matter probability images are segmented into brain regions to generate a set of sub-images of gray matter distribution in brain regions. Furthermore, based on the anatomical brain region rules, the MNI152 standard spatial preset brain region template is loaded. Spatial position verification is performed on the standard gray matter probability image and the MNI152 standard spatial preset brain region template to confirm that their coordinate systems are consistent. Each brain region label numbered 1 to M in the MNI152 standard spatial preset brain region template is traversed. For each brain region label, the corresponding brain region voxel position is extracted in the MNI152 standard spatial preset brain region template to generate a single brain region mask. The single brain region mask is multiplied with the standard gray matter probability image at the voxel level to extract the gray matter probability value within the brain region and generate a single brain region gray matter distribution sub-image. The M single brain region gray matter distribution sub-images are arranged in the order of brain region number to generate a set of brain region gray matter distribution sub-images.

[0043] The standard gray matter uncertainty map is mapped and aggregated according to the brain region mask corresponding to the set of brain region gray matter distribution sub-images to generate brain region uncertainty constraint information.

[0044] Furthermore, the process iterates through each brain region gray matter distribution sub-image numbered 1 to M in the set of brain region gray matter distribution sub-images. For each sub-image, binarization thresholding is performed to generate a corresponding brain region mask. The brain region mask is then multiplied voxel-by-voxel by the standard gray matter uncertainty map at the same spatial location. Uncertainty values ​​at mask values ​​of 1 are retained, and the remaining values ​​are set to zero to generate the uncertainty value distribution within that brain region. The voxel-level mean of the uncertainty value distribution within each brain region is calculated to generate the mean uncertainty value for that brain region. The M mean uncertainty values ​​are arranged in order of brain region number to generate a mean uncertainty vector. This mean uncertainty vector and the M brain region masks are then encapsulated together to generate brain region uncertainty constraint information. This constraint information provides a quantitative basis for the credibility of gray matter segmentation in each brain region in subsequent brain region anomaly scoring steps, thus suppressing false detection responses in high-uncertainty brain regions.

[0045] Anatomical brain region rules refer to numbering brain tissue in standard space according to the partition boundaries of a preset brain region template and generating corresponding brain region masks, and then performing brain region division on the standard image. For example, using the AAL2 (Automatic Anatomical Marking Version 2) preset brain region template, the brain tissue in the MNI152 standard space is divided into 120 anatomical brain regions. Each anatomical brain region is assigned a unique integer number from 1 to 120. Based on number 1, the spatial range of the left hippocampus is extracted to generate brain region mask number 1. Based on number 2, the spatial range of the right hippocampus is extracted to generate brain region mask number 2, and so on until number 120 is generated to generate all 120 brain region masks. All 120 brain region masks are then subjected to voxel-level multiplication operations with the standard gray matter probability image to extract the gray matter probability distribution of each brain region and generate 120 brain region gray matter distribution sub-images.

[0046] The brain region template is based on the anatomical division system and standard spatial anatomical positioning relationship of the standard brain atlas. It is set by projecting the boundary markings of each brain region onto the standard space and writing the corresponding brain region number label into each voxel.

[0047] High uncertainty brain regions refer to brain regions with relatively large mean uncertainty in the uncertainty constraint information. The mean uncertainty of such brain regions is obtained by calculating the arithmetic mean of the uncertainty values ​​of all voxels in the corresponding brain region mask coverage space of the standard gray matter uncertainty map. It reflects that the gray matter segmentation results of such brain regions are affected by noise, registration error or blurred anatomical boundaries, resulting in large fluctuations in gray matter probability and low reliability of abnormal scoring.

[0048] S4. Calculate the brain region abnormality score using the set of brain region gray matter distribution sub-images, and use the brain region uncertainty constraint information to suppress false detections, generating a brain region abnormality score map and a set of suspected structural abnormality regions. The abnormal response intensity of each brain region is calculated based on the set of sub-images of gray matter distribution in the brain region. The spatial heat distribution is obtained through voxel-level mapping and normalized and calibrated to generate an initial abnormality score map of the brain region. The expression for calculating the intensity of abnormal responses in brain regions is: ; in, It is the first The intensity of abnormal response in a brain region indicates the standard deviation of the degree of gray matter atrophy in that brain region from the reference value of healthy individuals; the value is non-negative. It is the first in the mean gray matter probability database of healthy elderly populations. The reference mean of gray matter probability in each brain region reflects the typical gray matter probability level of that brain region in a healthy elderly population. It is the first The spatial extent of each brain region; Spatial range A single voxel index within; It is a voxel The gray probability value in a standard gray probability image ranges from [0, 1]. It is the first in the gray matter probability mean database of healthy elderly populations. The reference standard deviation of the gray matter probability of each brain region reflects the degree of natural variation in the gray matter probability of that brain region in healthy individuals. Furthermore, the abnormal response intensity of each brain region is assigned to all voxel positions covered by the brain region mask obtained from the binarization of the corresponding brain region gray matter distribution sub-image. The results of all brain region mask assignments are accumulated to generate a spatial heat distribution. Min-max normalization is performed on the spatial heat distribution, and the voxel values ​​are linearly mapped to the closed interval [0,1] so that the minimum voxel value of the spatial heat distribution is mapped to 0 and the maximum voxel value is mapped to 1, thereby generating a normalized calibration spatial heat distribution. The normalized calibration spatial heat distribution is defined as the initial abnormality scoring map of the brain region.

[0049] It should be noted that spatial thermal distribution refers to a three-dimensional voxel-level image formed by uniformly mapping the abnormal response intensity of each brain region to the corresponding anatomical space of the brain region. The value of each voxel in the image is equal to the abnormal response intensity of its respective brain region. The higher the value, the more significantly the degree of gray matter atrophy in that brain region deviates from the healthy reference value. Spatial thermal distribution intuitively presents the spatial distribution pattern of the degree of abnormality at each anatomical location within the whole brain.

[0050] Based on the uncertainty constraint information of brain regions, the initial abnormality scoring map of brain regions is subjected to credibility suppression and sensitivity calibration to obtain the abnormality scoring map of brain regions. Furthermore, the uncertainty mean of each brain region in the brain region uncertainty constraint information sequence is traversed, and an exponential decay function is performed on the uncertainty mean of each brain region to calculate the corresponding brain region confidence weight. The confidence weight of each brain region is filled into the spatial range marked by the corresponding brain region mask in the brain region uncertainty constraint information to generate a spatialized confidence weight map. The initial abnormal score map of the brain region and the spatialized confidence weight map are multiplied on a voxel-by-voxel basis at the same spatial position, so that the abnormal scores of high uncertainty brain regions are suppressed by the exponential decay weight while the abnormal scores of low uncertainty brain regions are retained, generating a confidence-suppressed initial abnormal score map of the brain region. The sigmoid function is applied to the confidence-suppressed initial abnormal score map of the brain region for nonlinear compression calibration of the abnormal response dynamic range, generating a brain region abnormal score map.

[0051] The exponential decay weight is a confidence suppression coefficient obtained by mapping the mean uncertainty of a brain region through an exponential function. The larger the mean uncertainty of a brain region, the smaller the weight. This weight is set by nonlinearly transforming the mean uncertainty of a brain region through an exponential decay function.

[0052] Figure 4A two-dimensional slice projection of the brain region anomaly scoring map in standard space is presented. The horizontal axis represents the standard space x-axis, the vertical axis represents the standard space y-axis, and the color bars represent the brain region anomaly scoring values. Warm-colored areas correspond to the spatial range of brain regions with higher anomaly response intensity. The brain region anomaly scoring map is obtained by multiplying the initial brain region anomaly scoring map voxel by voxel and then calibrating it using the Sigmoid function. This suppresses the voxel values ​​of the brain region anomaly scoring map corresponding to the spatial location of high-uncertainty brain regions with exponential decay weights, thereby providing a more stable basis for the localization of anomaly peaks for subsequent local maximum detection and region growing, and reducing false detections.

[0053] The abnormal peak regions are located based on the brain region abnormality scoring map, and spatial coherence constraints are applied to generate a set of suspected structural abnormality regions.

[0054] Furthermore, local maxima detection is performed on the brain region abnormality scoring map to locate the set of voxel-level abnormal peak points. With each abnormal peak point as the center, a region growing algorithm is executed to gradually incorporate voxels whose brain region abnormality scoring map values ​​are not lower than a preset proportion of the corresponding abnormal peak point value into the growing region to generate an initial abnormal region set. The spatial volume of each initial abnormal region in the initial abnormal region set is calculated, and initial abnormal regions with spatial volumes less than a preset minimum volume threshold are removed to generate a volume-filtered abnormal region set. Three-dimensional morphological closing operations are performed on each volume-filtered abnormal region in the volume-filtered abnormal region set to eliminate internal voids and boundary jaggedness to generate a spatially coherent abnormal region set. The spatially coherent abnormal region set is defined as the suspected structural abnormal region set.

[0055] It should be noted that the preset ratio of abnormal peak point values ​​refers to the relative threshold for determining whether adjacent voxels belong to the same abnormal region during the region growth process. This ratio is defined as the lower limit of the ratio of the abnormal score value of adjacent voxels to the abnormal score value of abnormal peak points. It is set by analyzing the spatial gradient distribution statistical characteristics of abnormal score maps of brain regions in healthy elderly people and subjects with mild cognitive impairment. The minimum volume threshold is based on the pathological characteristics of Alzheimer's disease-related gray matter atrophy lesions, which present as continuous diffusion rather than isolated small point distribution in anatomical space (e.g., progressive atrophy of the hippocampus and entorhinal cortex usually extends in a sheet-like or clump-like manner along the nerve fiber projection pathway; brain regions affected by frontotemporal degeneration show continuous cortical thinning across gyri, rather than randomly scattered isolated small lesions). By statistically analyzing the volume distribution of false positive connected regions in the abnormal brain region scoring map of healthy elderly people and taking its 95th percentile as the upper limit reference, the minimum volume threshold is set to effectively filter out segmentation noise and small isolated responses caused by normal anatomical variations. An exemplary value range is 100 to 500 cubic millimeters.

[0056] The region growth algorithm refers to using an abnormal peak point as a seed point, and sequentially examining the abnormal score map values ​​of adjacent voxels in the six-neighbor or twenty-six-neighbor areas of the seed point. If the value of an adjacent voxel is not lower than the product of the seed point value and the abnormal peak point value in a preset ratio, the adjacent voxel is included in the growth region and used as a new seed point to continue expansion. The process of neighbor examination and region expansion is repeated until no new voxels meet the growth conditions, and finally a connected abnormal region is formed with the abnormal peak point as the center and the abnormal score values ​​of the voxels inside the region showing a continuous spatial decay.

[0057] S5. Reversible reverse verification is used to screen out false anomalies and perform credibility convergence, outputting images of credible anomaly areas and early detection conclusions.

[0058] Based on the brain region abnormality scoring map, local structural restoration and reconstruction are performed on the candidate regions corresponding to the suspected structural abnormality regions to generate structural restoration images. Furthermore, a three-dimensional spatial mask is extracted for each suspected structural abnormality region in the set of suspected structural abnormality regions. A voxel-level multiplication operation is performed between the three-dimensional spatial mask of each suspected structural abnormality region and the brain region abnormality scoring map to extract the abnormality score distribution of each suspected structural abnormality region. Threshold segmentation is performed on the abnormality score distribution of each suspected structural abnormality region to generate a core abnormality mask for each suspected structural abnormality region. A logical NOT operation is performed between the core abnormality mask of each suspected structural abnormality region and the standard gray matter probability image to generate a surrounding normal gray matter mask for each suspected structural abnormality region. Three-dimensional Gaussian interpolation is performed on the voxel gray values ​​covered by the surrounding normal gray matter mask of each suspected structural abnormality region to generate a structural restoration gray value distribution for each suspected structural abnormality region. The structural restoration gray value distribution of each suspected structural abnormality region is filled into the spatial location of the corresponding suspected structural abnormality region in the standard gray matter probability image to generate a structural restoration image.

[0059] It should be noted that the voxel difference between structural restoration images and standard gray matter probability images can directly quantify the degree of local gray matter volume loss, avoiding interference from individual head circumference differences introduced by global brain volume normalization in traditional methods; at the same time, the structural restoration process is regulated by the uncertainty constraint information of brain regions, and reliable reconstruction is performed only in brain regions with low uncertainty, effectively suppressing false restoration in high-noise regions and improving the anatomical specificity and clinical reliability of atrophy quantitative assessment.

[0060] Structural consistency difference assessment is performed on the structural reconstruction image and the gray matter probability image, and pseudo-anomaly regions in the suspected structural anomaly region set are screened out to generate an anomaly candidate set; Furthermore, a voxel-level absolute value difference operation is performed on the structural reconstruction image and the gray matter probability image to generate a structural consistency difference image. Each suspected structural anomaly region in the suspected structural anomaly region set is traversed and the spatial mask corresponding to each suspected structural anomaly region is extracted. A voxel-level multiplication operation is performed on the spatial mask corresponding to each suspected structural anomaly region and the structural consistency difference image to obtain the difference voxel set of each suspected structural anomaly region. The arithmetic mean of the voxel values ​​of the difference voxel set of each suspected structural anomaly region is calculated to generate the region consistency difference of each suspected structural anomaly region. Suspected structural anomaly regions with region consistency differences less than the consistency difference threshold are identified as pseudo-anomaly regions. All pseudo-anomaly regions are removed from the suspected structural anomaly region set. The remaining suspected structural anomaly regions are arranged in order of spatial coordinates and voxel index to generate an anomaly candidate set.

[0061] It should be noted that the consistency difference threshold is defined based on the statistical distribution difference in the gray matter probability value amplitude between irreversible changes in local structure caused by gray matter atrophy and reversible artifacts caused by segmentation noise or registration error. By analyzing the distribution of structural consistency differences caused by segmentation algorithm fluctuations and spatial normalization residuals in images of healthy elderly people and taking its 99th percentile as the noise upper limit, and combining it with the lower limit of structural consistency differences in the actual atrophic areas confirmed by clinical diagnosis in subjects with mild cognitive impairment, the consistency difference threshold is set as a critical value that can effectively distinguish between pathological gray matter volume loss and technical artifacts. An exemplary value range is 0.08 to 0.15.

[0062] The anomaly candidate set is corrected by neighborhood consistency propagation, the credibility of the anomaly candidate set is updated round by round, an anomaly credibility map is generated and credibility convergence is performed, and a credible anomaly region image and early detection conclusion are generated.

[0063] Furthermore, an initial confidence value is assigned to each abnormal candidate region in the abnormal candidate set based on its abnormal response intensity in the brain region abnormality scoring map. All pairs of abnormal candidate regions in the abnormal candidate set are traversed, and the presence of voxel-level boundary contact in three-dimensional space is detected for each pair of regions. Adjacency connections are established for regions with boundary contact to form a spatial adjacency relationship map, and the six neighboring regions of each abnormal candidate region are determined. The confidence value of each abnormal candidate region is iteratively updated by neighborhood weighted average based on the spatial adjacency relationship map. The iteration is repeated until the change in confidence value is less than the confidence convergence threshold for two consecutive rounds, generating a converged confidence value set. The confidence value of each abnormal candidate region in the converged confidence value set is filled into all voxel positions covered by the spatial mask of the corresponding abnormal candidate region to generate an abnormal confidence map. Threshold segmentation is performed on the abnormal confidence map to extract voxel regions with confidence values ​​greater than the confidence judgment threshold to generate a credible abnormal region image. An early detection conclusion is generated based on the spatial distribution of abnormal regions in the credible abnormal region image and the anatomical matching relationship with the template of typical atrophied brain regions in early Alzheimer's disease.

[0064] Early detection results include the determination of the early risk of Alzheimer's disease, the name of the anatomical brain region corresponding to the abnormal region in the credible abnormal region image, and the three-dimensional spatial coordinate range of the abnormal region in standard space.

[0065] It should be noted that the confidence threshold is defined based on the need for a balance between sensitivity and specificity in early detection of Alzheimer's disease. For example, by analyzing the convergence confidence value distribution of abnormal regions in clinically diagnosed mild cognitive impairment subjects and the confidence value distribution of false positive regions in healthy elderly people, the intersection of the two distributions is taken as the optimal decision boundary to maximize the Youden index. At the same time, considering the priority requirement of sensitivity in early screening, the confidence threshold is appropriately shifted towards lower specificity. An exemplary value range is 0.65 to 0.80.

[0066] In summary, this invention generates a gray matter uncertainty map by employing a perturbation-reseparation-comparison strategy, transforming gray matter probability fluctuations into quantifiable uncertainty constraints and performing credibility suppression and sensitivity calibration during the formation of the brain region abnormality scoring map. This makes the candidate generation of the suspected structural abnormality region set more focused on stable abnormal responses. Subsequently, local structural restoration and reconstruction are performed on the candidate regions corresponding to the suspected structural abnormality region set to obtain structural restoration images. False abnormal regions are screened out by combining structural consistency difference evaluation. Then, neighborhood consistency propagation correction drives the convergence of the abnormality credibility map, outputting credible abnormality region images and early detection conclusions, thereby enhancing the interpretability and verification consistency of image-level evidence.

[0067] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for early image detection of Alzheimer's disease based on single-modal MRI, characterized in that: include, Acquire raw structural MRI images of a single sequence from the subject and perform standardized processing to generate uniform structural MRI images and quality control marker information; Intensity correction parameters are set based on quality control marker information, and intensity correction, non-brain tissue stripping, and gray matter separation are performed on uniform structure MRI images to generate gray matter probability images and gray matter uncertainty maps. The gray matter probability image and gray matter uncertainty map are aligned in standard space and segmented according to the rules of anatomical brain regions to generate a set of sub-images of gray matter distribution in brain regions and uncertainty constraint information of brain regions. Brain region abnormality scores are calculated using a set of brain region gray matter distribution sub-images, and false detections are suppressed using brain region uncertainty constraint information, generating brain region abnormality score maps and sets of suspected structural abnormality regions; Reversible reverse verification is used to screen out false anomalies and perform credibility convergence, outputting images of credible anomaly regions and early detection conclusions.

2. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 1, characterized in that: The steps for generating uniform structural MRI images are as follows. The original structural MRI images of a single sequence of the subject are acquired and the acquisition time and sequence identification information are written in. At the same time, sequence consistency is verified and a structural MRI baseline image is generated. The structural MRI baseline images are processed for format parsing, coordinate orientation unification, and standardization to generate uniform structural MRI images.

3. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 2, characterized in that: The quality control marker information is obtained by performing quality assessment on uniform structure MRI images.

4. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 1, characterized in that: The steps for performing intensity correction, non-brain tissue stripping, and gray matter separation on uniform structural MRI images to generate gray matter probability images are as follows. Intensity correction parameters are set based on quality control marker information, and intensity correction is performed on uniform structural MRI images to generate grayscale reference structural MRI images. Non-brain tissue dissection was performed on grayscale baseline MRI images to obtain brain parenchyma range constraint information, and brain parenchyma regions were extracted based on the brain parenchyma range constraint information to generate brain parenchyma structure MRI images. Gray matter tissue separation was performed on MRI images of brain parenchyma structures to generate gray matter probability images.

5. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 4, characterized in that: The gray-quality uncertainty map is obtained by using a perturbation reseparation and comparison strategy to evaluate the gray-quality consistency of the gray-quality probability image and statistically analyze the gray-quality probability fluctuation.

6. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 1, characterized in that: The steps for segmenting brain regions according to anatomical brain region rules to generate sub-image sets of gray matter distribution and uncertainty constraint information for brain regions are as follows: The gray-quality probability image and the gray-quality uncertainty map are aligned in a standard space to generate a standard gray-quality probability image and a standard gray-quality uncertainty map. Based on the rules of anatomical brain regions, standard gray matter probability images are segmented into brain regions to generate a set of sub-images of gray matter distribution in brain regions. The standard gray matter uncertainty map is mapped and aggregated according to the brain region mask corresponding to the set of brain region gray matter distribution sub-images to generate brain region uncertainty constraint information.

7. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 6, characterized in that: The anatomical brain region rule refers to numbering the brain tissue in the standard space according to the partition boundaries of the preset brain region template and generating corresponding brain region masks, and then performing brain region division on the standard gray matter probability image.

8. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 1, characterized in that: The steps for using brain region uncertainty constraint information to suppress false detections and generate a brain region abnormality scoring map are as follows. The abnormal response intensity of each brain region is calculated based on the set of sub-images of gray matter distribution in the brain region. The spatial heat distribution is obtained through voxel-level mapping and normalized and calibrated to generate an initial abnormality score map of the brain region. Based on the uncertainty constraint information of brain regions, the credibility suppression and sensitivity calibration of the initial abnormality scoring map of the brain region are performed to obtain the abnormality scoring map of the brain region.

9. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 8, characterized in that: The set of suspected structurally abnormal regions is obtained by locating abnormal peak regions based on brain region abnormality scoring maps and applying spatial coherence constraints.

10. The method for early image detection of Alzheimer's disease based on single-modal MRI as described in claim 1, characterized in that: The steps for using reversible reverse verification to screen out false anomalies and perform confidence convergence, outputting a reliable anomaly region image and early detection conclusions, are as follows: Based on the brain region abnormality scoring map, local structural restoration and reconstruction are performed on the candidate regions corresponding to the suspected structural abnormality regions to generate structural restoration images. Structural consistency difference assessment is performed on the structural reconstruction image and the gray matter probability image, and pseudo-anomaly regions in the suspected structural anomaly region set are screened out to generate an anomaly candidate set; The anomaly candidate set is corrected by neighborhood consistency propagation, the credibility of the anomaly candidate set is updated round by round, an anomaly credibility map is generated and credibility convergence is performed, and a credible anomaly region image and early detection conclusion are generated.

Images