A method for morphological characterization and classification of ab amyloid plaque based on multi-dimensional image features

Abstract

The application discloses an A beta plaque morphology characterization and classification method based on multi-dimensional image features, and comprises the following steps: S1, constructing an A beta plaque data set; S2, for each A beta plaque in the A beta plaque data set, extracting low-order morphological features and high-order morphological features, and constructing a feature space; S3, screening and verifying all features in the feature space for multiple times, and classifying the screened features into diffuse plaque features, core plaque features and CAA plaque features; each type of feature is divided into key features, main features and secondary features according to the importance; S4, training a supervised classifier to obtain a plaque morphology classification model; and S5, classifying A beta plaques. The scheme realizes multi-angle and high-precision quantitative description of plaque morphology, can more comprehensively and finely depict the complex morphology of A beta plaques, and reduces individual differences in artificial judgment.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and in particular to a method for morphological characterization and classification of Aβ plaques based on multidimensional image features. Background Technology

[0002] Aβ (amyloid-β) plaques are one of the core pathological features of Alzheimer's disease (AD), widely distributed in the brains of AD patients. The morphology of Aβ plaques is not uniform, but rather exhibits a diverse appearance, ranging from diffuse to highly dense. Different plaque morphologies have different biological significance and are closely related to disease progression and pathological changes. Currently, plaque morphology analysis relies heavily on qualitative judgment by pathologists, lacking quantitative references and unified diagnostic criteria. To date, a universal and systematic analytical method for resolving plaque morphological heterogeneity remains lacking. Accurate and quantitative characterization of Aβ plaque morphology is of great significance for research on the mechanisms of AD, drug screening, and pathological staging.

[0003] Existing Aβ plaque morphology analysis mainly relies on the following methods: Firstly, there is the method of subjective sensory judgment: Pathologists rely on their visual experience to identify typical plaque morphologies by integrating information such as texture, color, and spatial relationships against a complex stained background. They classify these plaques as diffuse plaques, nuclear plaques, or cerebral amyloid angiopathy (CAA) plaques and make descriptive judgments on atypical morphologies. However, this method is inherently subjective and qualitative. Different pathologists may have different criteria for classifying plaque morphology. For transitional and special morphological plaques that lack quantitative references, it is even more difficult to achieve a unified judgment. In addition, this method is inefficient and cannot meet the needs of large-scale, standardized quantitative analysis.

[0004] Secondly, basic morphological measurement: In most basic research, the differentiation of plaque morphology is often overlooked. Current mesoscale analysis of the whole brain of mice is still mainly based on quantitative statistics, and detailed analysis of plaque morphology is relatively lacking. Most existing studies only extract rough indicators such as plaque count and size changes within the whole brain or specific brain regions. For example, the β-amyloid plaque identification and measurement method based on image processing disclosed in the authorization announcement number CN107561264B identifies plaques in microscope images through fluorescence microscopy and performs automatic quantity statistics and feature calculation. A few studies use image processing software to extract basic statistical features such as plaque volume, radius, and gray-level histogram, but these features can only describe the macroscopic morphology of the plaque and cannot capture its complex internal texture, gray-level changes, and structural heterogeneity. The ability to describe the morphologically variable and diffusely edged Aβ plaques is seriously insufficient. Summary of the Invention

[0005] Therefore, to solve the above problems, this invention provides a method for Aβ patch morphology characterization and classification based on multidimensional image features.

[0006] This invention is achieved through the following technical solution: A method for Aβ patch morphology representation and classification based on multidimensional image features includes the following steps: S1: Construct an Aβ patch dataset with accurate labels, including a balanced number of diffuse patches, nuclear patches, and CAA patches, covering multiple mouse ages and brain regions. S2: For each Aβ patch in the Aβ patch dataset, extract low-order morphological features that reflect its macroscopic morphology and intuitive grayscale distribution, and high-order morphological features that characterize its internal texture heterogeneity, spatial complexity patterns and sub-visual radiomics attributes, and construct a feature space that integrates the low-order and high-order morphological features of the Aβ patch. S3: All features in the feature space are screened and verified multiple times. The screened features are classified into diffuse patch features, nuclear patch features and CAA patch features. Each type of feature is divided into key features, main features and secondary features according to their importance. S4: Select a supervised classifier type, assign weights to the selected features based on the importance level of each feature, and train the selected supervised classifier using the Aβ patch dataset containing the selected patch features to obtain a patch morphology classification model. S5: Classify Aβ patches using the patch morphology classification model.

[0007] Preferably, step S1 includes the following steps: S11: Acquire three-dimensional image data of mouse brains at multiple age stages; S12: Preprocess the mouse brain 3D image data to suppress the background region outside the brain outline, and crop the preprocessed mouse brain 3D image data into several data blocks of the same size. S13: Select multiple sets of mouse data from each age group, and use variable proportion sampling to extract plaque samples from the selected mouse data to obtain a plaque sample set with a balanced number of samples from each age group. S14: A classifier is used to make a preliminary classification on the coronal section of the patch, and the classification is manually checked and corrected by combining the three-dimensional morphology and the two-dimensional section in three orthogonal directions to achieve the classification labeling of patch samples in the patch sample set.

[0008] Preferably, in step S2, the extraction of low-order morphological features for each Aβ patch includes the following steps: calculating the radius, volume, mean gray value, standard deviation, and skewness of the Aβ patch; Calculate the ratio of the longest axis to the shortest axis of the three-dimensional minimum circumscribed cuboid of patch Aβ to obtain the ratio of the longest axis to the shortest axis. Polynomial fitting was performed on the grayscale curve passing through the center point of the Aβ patch to calculate the grayscale difference between the central peak and the two side valleys and secondary peaks, and the larger value was taken as the grayscale peak-valley drop. Within a radius of 1 / 2 from the center of patch Aβ, the maximum gray value is selected as the peak value of the patch signal.

[0009] Preferably, the extraction step of the higher-order morphological features includes: Obtain the original image of the patch and its foreground signal mask; Based on the original image and its foreground signal mask, high-order morphological features are extracted from five dimensions: shape features, gray-level co-occurrence matrix, gray-level size region matrix, adjacent gray-level tone difference matrix, and gray-level dependence matrix.

[0010] Preferably, step S3 includes the following steps: S31: Perform the Kruskal-Wallis test on each feature in the feature space in turn to determine whether there are significant differences among the three types of plaques: diffuse, nuclear, and CAA. S32: Use bulldozer distance to assess the degree of separation of feature distributions among different morphological categories; S33: Validate the ability of features to distinguish patch types using k-means clustering; S34: Based on the evaluation results of steps S31-S33, the importance of each feature is quantified using an integral system, and the features are divided into key level, primary level and secondary level. S35: Decompose the patch image into 8 frequency domain sub-bands using wavelet transform, and perform the Kruskal-Wallies test again on each sub-band. Retain features that still show significant inter-class differences on at least 4 sub-bands. S36: Principal component analysis is used to perform dimensionality reduction evaluation on each feature set, determine the minimum number of features required to describe each type of patch, and verify whether the number of features in the feature set is sufficient to describe the morphological attributes of the patch to which it belongs.

[0011] Preferably, in step S4, the original features are used for classification, and a small number of randomly selected patches are used to pre-train commonly used supervised classifiers. The supervised classifier with the highest test accuracy is selected as the best supervised classifier.

[0012] Preferably, in step S4, commonly used supervised classifiers include XGBoost, Random Forest, Extreme Random Tree, Gaussian Naive Bayes, k-Nearest Neighbors, Logistic Regression, Decision Tree, and Support Vector Machine.

[0013] Preferably, step S4 includes the following steps: S41: Select XGBoost classifier as the best supervised classifier; S42: Obtain the Aβ patch dataset containing the filtered patch features, and divide the Aβ patch dataset into a training set and a test set according to a preset ratio; S43: Perform step-by-step debugging of key parameters of XGBoost to determine the optimal parameter combination; S44: Set the weights for key features, primary features, and secondary features respectively, and assign an initial weight to each feature selected from the feature space; S45: Using the above optimal parameter combination and initial weight settings, train the XGBoost classifier using the training set to obtain the patch morphology classifier; S46: Use the test reagent to evaluate the performance of the patch morphology classifier.

[0014] Preferably, after the patch morphology classifier outputs the classification result, step S6 is further included: introducing low-order features, setting judgment thresholds based on the statistical distribution of low-order features among CAA patches, nuclear patches and diffuse patches respectively, and performing secondary correction on the classification result output by the patch morphology classifier.

[0015] The patch morphology classification model was obtained using the Aβ patch morphology representation and classification method based on multidimensional image features as described above.

[0016] The beneficial effects of the technical solution of this invention are mainly reflected in: 1. Based on an Aβ patch dataset containing accurate labels of diffuse patches, nuclear patches, and CAA patches with balanced numbers covering multiple mouse ages and brain regions, a feature space for Aβ patches was constructed from three key dimensions: shape, grayscale, and texture. This feature space integrates intuitive low-order features with high-order features describing deep attributes, enabling multi-angle and high-precision quantitative description of patch morphology, extracting deep image information that is difficult for the human eye to capture directly, and more comprehensively and finely depicting the complex morphology of Aβ patches.

[0017] 2. By repeatedly screening and validating all features in the feature space, on the one hand, the importance hierarchy of features is divided and invalid features are eliminated. Based on this importance hierarchy, differentiated initial weights are assigned to each feature during the model training phase, thereby guiding the classifier to prioritize experimentally validated strong discriminative features during iterative splitting, thus accelerating model convergence and improving computational efficiency. On the other hand, it can objectively and efficiently achieve accurate differentiation of the three classic morphologies: diffuse patches, nuclear patches, and CAA patches, significantly reducing individual differences in human judgment and meeting the urgent need for standardization and high precision in large-scale patch data analysis.

[0018] 3. Based on the trained patch morphology classifier, low-order features are introduced to post-correct the output of the patch morphology classifier, so that the classification decision has a clear biological and image basis and improves the credibility of the classification results. Attached Figure Description

[0019] Figure 1 This is a flowchart of the Aβ patch morphology characterization and classification method based on multidimensional image features of the present invention; Figure 2 This is a flowchart of steps S1-S4 in the Aβ patch morphology characterization and classification method based on multidimensional image features in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the low-order morphological feature extraction process in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of steps S3 and S4 in Embodiment 1 of the present invention. Detailed Implementation

[0020] To make the objectives, advantages, and features of the present invention clearer and more detailed, the following non-limiting description of preferred embodiments will be illustrated and explained. These embodiments are merely typical examples of applying the technical solutions of the present invention; all technical solutions formed by equivalent substitutions or equivalent transformations fall within the scope of protection claimed by the present invention.

[0021] It is also stated that, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] Example 1: This embodiment discloses a method for Aβ patch morphological characterization and classification based on multidimensional image features, such as... Figure 1 , Figure 2 As shown, it includes the following steps: S1: Construct an Aβ patch dataset with accurate labels, including a balanced number of diffuse patches, nuclear patches, and CAA patches, covering multiple mouse ages and brain regions. Step S1 includes the following steps: S11: Acquire three-dimensional image data of mouse brains at multiple age stages; In this study, the three-dimensional brain images of mice were obtained from the Aβ plaques in the brains of 5xFAD transgenic mice using High-Definition fluorescence Micro-Optical Sectioning Tomography (HD-fMOST). Experimental groups were set up at five time points: 6 weeks, 2 months, 4 months, 6 months, and 10 months of age. Three-dimensional brain images of 5xFAD transgenic mice labeled with the DANIR-8c fluorescent probe were obtained, with an image resolution of 0.325 × 0.325 × 1 μm³.

[0023] S12: Preprocess the mouse brain 3D image data to suppress the background region outside the brain outline, and crop the preprocessed mouse brain 3D image data into several data blocks of the same size. In this embodiment, the preprocessing step involves downsampling the data to 10×10×10 μm³, drawing the brain outline, and setting the external background to zero. Subsequently, the preprocessed mouse brain 3D image data is downsampled to isotropic resolution (1×1×1 μm³) and cropped into fixed-size data blocks (128×128×128 voxels).

[0024] S13: Select multiple sets of mouse data from each age group, and use variable proportion sampling to extract plaque samples from the selected mouse data to obtain a plaque sample set with a balanced number of samples from each age group. Three sets of mouse data were selected from the five age groups corresponding to the above five age stages. In each set of data, the sampling ratio was adjusted according to the difference in the number of plaques in different age groups. Since there were fewer plaques in younger mice, the sampling ratio was increased, while the sampling ratio was decreased in older mice with more plaques. In this embodiment, the sampling ratios for the five age groups of 6 weeks, 2 months, 4 months, 6 months and 10 months were 10%, 7%, 5%, 2% and 1% respectively, in order to balance the differences in the number of mice in different age groups. Finally, about 84,000 plaques were extracted.

[0025] S14: In this embodiment, the pathologically interpretable classification method for Alzheimer's disease based on a convolutional neural network process, published by Tang et al. (2019) in Nature Communications, is first adopted. [1] Preliminary classification was performed on the coronal section of the plaques. Then, the three-dimensional morphology and two-dimensional sections in three orthogonal directions were combined for manual verification and correction to obtain accurate labels for diffuse plaques, nuclear plaques, and CAA plaques, with 31,329, 26,840, and 25,670 labels, respectively. Based on the above accurate labels for diffuse plaques, nuclear plaques, and CAA plaques, an Aβ plaque dataset was constructed. S2: For each Aβ patch in the Aβ patch dataset, extract low-order morphological features that reflect its macroscopic morphology and intuitive grayscale distribution, and high-order morphological features that characterize its internal texture heterogeneity, spatial complexity patterns and sub-visual radiomics attributes, and construct a feature space that integrates the low-order and high-order morphological features of the Aβ patch. like Figure 2 , Figure 3 As shown, in this embodiment, the extraction of low-order morphological features for each Aβ patch includes the following steps: Calculate the radius, volume, mean gray value, standard deviation, and skewness of patch Aβ; Obtain the minimum 3D bounding cube of the patch and determine its longest axis. With the shortest axis Calculate the ratio of major to minor axis: ; Polynomial fitting is performed on the grayscale curve passing through the center point of the Aβ patch. Starting from the central peak, the search proceeds to both sides to locate the deepest valley and its adjacent highest peak. The grayscale difference between the valley and the adjacent peak is calculated, and the larger value between the two sides is taken as the grayscale peak-valley drop (Gray_drop). Within a radius of 1 / 2 from the center of the Aβ patch, the maximum gray value is selected as the peak value of the patch signal (Max_gray) to avoid mutual interference of diffuse halos from neighboring patches.

[0026] In this embodiment, the extraction step of the higher-order morphological features includes: Obtain the original image of the patch and its foreground signal mask; The process involves using three-dimensional maximum entropy threshold segmentation to extract the initial foreground signal, thus preserving the weak signals in the diffuse region. Then, connected component analysis is performed to remove small noise points smaller than 10% of the maximum connected component. Next, the Canny edge detection operator is used to extract the edges of the connected components, and the complete boundary of the patch is obtained through the maximum circumscribed contour algorithm. Finally, a binary mask of the patch is obtained.

[0027] Based on the original image and its foreground signal mask, high-order morphological features are extracted from five dimensions: shape features, gray-level co-occurrence matrix, gray-level size region matrix, adjacent gray-level tone difference matrix, and gray-level dependence matrix. In this embodiment, the PyRadiomics toolkit of Python is used for feature extraction, with the grayscale level set to 0–255 and the voxel spacing set to [1, 1, ...]. [1], and enabled five sets of features: Shape, Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Size Region Matrix (GLSZM), Neighboring Gray-Level Tone Difference Matrix (NGTDM), and Gray-Level Dependency Matrix (GLDM). The image and its mask were input into PyRadiomics, and the feature extraction method was called to calculate each patch, resulting in 74 high-order feature values. Among them: 17 shape features, including mesh volume, voxel volume, surface area, surface area to volume ratio, sphericity, compactness 1, compactness 2, sphericity misalignment, maximum 3D diameter, maximum 2D diameter (slice plane), maximum 2D diameter (column plane), maximum 2D diameter (row plane), principal axis length, secondary axis length, shortest axis length, elongation, and flatness; 22 gray-level co-occurrence matrix features, including joint mean, cluster prominence, cluster shadowness, cluster trend, contrast, correlation, difference mean, difference entropy, difference variance, joint energy, joint entropy, inverse moment, normalized inverse moment, and inverse... The system includes 16 features for gray-level region matrices, such as normalized inverse variance, inverse variance, maximum probability, entropy, sum of squares, autocorrelation, information relevance measure 1, and information relevance measure 2. These features include gray-level non-uniformity, normalized gray-level non-uniformity, gray-level variance, low gray-level region emphasis, high gray-level region emphasis, small region low gray-level emphasis, small region high gray-level emphasis, large region low gray-level emphasis, large region high gray-level emphasis, size region non-uniformity, normalized size region non-uniformity, region variance, and small region strength. Features include: tone, large area emphasis, area percentage, and area entropy; five features in the adjacent gray-level tone difference matrix, including busyness, contrast, complexity, intensity, and roughness; and fourteen features in the gray-level dependency matrix, including gray-level non-uniformity, dependency non-uniformity, normalized dependency non-uniformity, gray-level variance, dependency variance, dependency entropy, low gray-level emphasis, high gray-level emphasis, small dependency low gray-level emphasis, small dependency high gray-level emphasis, large dependency low gray-level emphasis, large dependency high gray-level emphasis, small dependency emphasis, and large dependency emphasis.

[0028] In this embodiment, for all Aβ patches in the Aβ patch dataset, a process pool is created using the Python standard library multiprocessing. The size of the process pool can be set according to the hardware conditions. The patch list is divided into batches and put into the process pool for parallel processing. Each process independently completes the extraction steps of the low-order morphological features and high-order morphological features. The main process collects all results and summarizes them to obtain the feature space of all Aβ patches in the Aβ patch dataset.

[0029] S3: All features in the feature space are screened and verified multiple times. The screened features are classified into diffuse patch features, nuclear patch features and CAA patch features. Each type of feature is divided into key features, main features and secondary features according to their importance. like Figure 2 , Figure 4 As shown, in this embodiment, step S3 includes the following steps: S31: Perform the Kruskal-Wallis test (KW test) on each feature in the feature space in turn to determine whether there are significant differences among the three types of plaques: diffuse, nuclear, and CAA. Based on the significance level (p<0.05, p<0.01, p<0.001), the features are initially classified as secondary, primary, and critical features.

[0030] S32: Normalize the feature values ​​and use the bulldozer distance (EMD) to calculate the distribution interval (EMD_a) between the current class and the other classes, as well as the minimum interval (EMD_b) between the current class and all classes. Use the mean of all features as the threshold. If the EMD_a and EMD_b of a certain feature are both greater than the corresponding threshold, it is included in the candidate feature. This is used to evaluate the degree of separation of feature distributions among different morphological categories.

[0031] S33: Use k-means clustering to verify the ability of features to distinguish patch types; wherein, binary clustering (merging the current class with the rest of the classes) and ternary clustering of features are compared respectively, and the feature level is determined according to the accuracy of cluster classification: >50% is secondary, >70% is primary, and >90% is critical.

[0032] S34: Based on the evaluation results of steps S31-S33, the importance of each feature is quantified using an integral system, and the features are divided into key level, primary level and secondary level. The KW test results are used as the preliminary evaluation results. Based on this, combined with the evaluation results of steps S31-S33: 1 point is awarded for each secondary or alternative feature selected, 2 points for primary features, and 3 points for key features; a score >5 is a key feature, >3 is a primary feature, and >1 is a secondary feature. The features with insufficient scores in the preliminary evaluation results are adjusted according to the above rules.

[0033] S35: Wavelet Domain Mapping Secondary Screening: A discrete wavelet transform is performed on the 3D image of each Aβ patch, decomposing the patch image into 8 frequency domain sub-bands, corresponding to combinations of high frequency (H) and low frequency (L) in three dimensions: HHH, HHL, HLH, HLL, LHH, LHL, LLH, LLL. The KW test is performed again on each sub-band, retaining features that still show significant inter-class differences in at least 4 sub-bands to ensure that the features have stable discriminative performance in different frequency domains; if a feature is significant in all 8 sub-bands, it is upgraded to a key feature; after this screening, 60 diffuse patch features, 61 nuclear patch features, and 68 CAA patch features are finally identified, and the importance level of each feature is determined.

[0034] S36: After feature selection, principal component analysis (PCA) was used to perform dimensionality reduction evaluation on diffuse plaque features, nuclear plaque features, and CAA plaque features to determine the minimum number of features required to describe each type of plaque and to verify whether the number of features in the feature set was sufficient to describe the morphological attributes of the plaques. Specifically, all feature matrices of each morphological class were input into the PCA model, the explained variance ratio of each principal component was calculated, and the cumulative explained variance curve was plotted. When the first 14 principal components were selected, the cumulative explained variance reached 90%. When the cumulative explained variance reached 99%, the number of principal components required for diffuse plaques, nuclear plaques, and CAA plaques were 38, 37, and 38, respectively. These values ​​were less than the actual number of features selected for each morphological class (60, 61, and 68), proving that the number of selected features was sufficient to completely characterize the morphological features of various plaques.

[0035] S4: Select a supervised classifier type, assign weights to the selected features based on the importance level of each feature, and train the selected supervised classifier using the Aβ patch dataset containing the selected patch features to obtain a patch morphology classification model.

[0036] like Figure 2 , Figure 4 As shown, in this embodiment, step S4 includes the following steps: S41: Select XGBoost classifier as the best supervised classifier; The optimal supervised classifier was selected from eight commonly used supervised classifiers: XGBoost, Random Forest, Extremely Random Tree, Gaussian Naive Bayes, k-Nearest Neighbors, Logistic Regression, Decision Tree, and Support Vector Machine. The performance of each of these eight supervised classifiers was tested, including tests with and without pre-training conditions. The tests without pre-training conditions used the original features for classification, while the tests with pre-training conditions used 3000 randomly selected patches. In the subsequent testing, the supervised classifier with the highest accuracy was selected as the best supervised classifier. In this embodiment, the accuracy of XGBoost without pre-training was 0.686, which was higher than that of Random Forest (0.685) and Extreme Random Tree (0.683), which ranked second and third respectively. After small-sample pre-training, the accuracy of XGBoost increased to 0.8227, which was still better than that of Random Forest (0.8152) and Extreme Random Tree (0.8045), which ranked second and third respectively. Therefore, XGBoost was selected as the patch morphology classifier in this embodiment.

[0037] S42: Obtain the Aβ patch dataset containing the filtered patch features, and divide the Aβ patch dataset into a training set and a test set according to a preset ratio. In this embodiment, the preset ratio of the training set to the test set is 5:1. The training set contains approximately 69,866 patches, and the test set contains approximately 13,973 patches.

[0038] S43: Perform step-by-step debugging of key parameters of XGBoost to determine the optimal parameter combination; the key parameters finally confirmed include: tree structure control parameters (max_depth=12, min_child_weight=1), node splitting loss threshold (gamma=2), sampling parameters (subsample=0.9, colsample_bytree=0.2), regularization parameters (reg_alpha=0.1, reg_lambda=1), number of iterations (n_estimators=140), and learning rate (learning_rate=0.1).

[0039] S44: Set the weights for key features, primary features, and secondary features respectively, and assign an initial weight to each feature selected from the feature space; the weight ratio of key features, primary features, and secondary features was determined to be 1.4 : 1.2 : 0.4 after testing.

[0040] S45: Using the above optimal parameter combination and initial weight settings, train the XGBoost classifier using the training set to obtain the patch morphology classifier.

[0041] S46: The performance of the patch morphology classifier is evaluated using the test reagent. In this embodiment, the overall accuracy of the classification results output by the patch morphology classifier reaches over 96%, and the precision, recall, and F1 score of the three types of patches are all over 94%.

[0042] S5: Apply the patch morphology classification model to classify Aβ patches and output the classification results.

[0043] Example 2: This embodiment discloses a patch morphology classification model, which is obtained by using the Aβ patch morphology representation and classification method based on multidimensional image features as described in Embodiment 1.

[0044] Example 3: This embodiment discloses a method for characterizing and classifying Aβ patch morphology based on multidimensional image features. To further enhance the biological rationality of the classification, in addition to the steps in Embodiment 1, this embodiment further includes step S6 after the patch morphology classifier outputs the classification results: introducing low-order features, setting judgment thresholds based on the statistical distribution of low-order features among CAA patches, nuclear patches and diffuse patches, and performing secondary correction on the classification results output by the patch morphology classifier.

[0045] The thresholds are set based on the statistical distribution of the corresponding features in the dataset: the lower limit threshold r2 = 2.40 and the upper limit threshold r1 = 2.60 for the ratio of major to minor axes, which are taken from the 95th percentile (2.39) and the 99th percentile (2.58) of non-CAA patches, respectively, rounded up; the signal peak threshold is 225 (for 8-bit images), which is taken from the larger value of the 95th percentile of non-nuclear patches and the 5th percentile of nuclear patches; the grayscale peak-valley difference correction is based on waveform detection, and is triggered when there is a trough-peak transition.

[0046] The correction rules are as follows: When Axial_ratio > 2.40, if the classification is CAA, the confidence level is increased by 0.1; otherwise, it is decreased by 0.1. If Axial_ratio > 2.60 and the current classification is not CAA, the result is reset to CAA. If a grayscale peak-valley difference is detected, and the classification is a kernel patch, the confidence level is increased by 0.1; otherwise, it is decreased by 0.1. When Max_gray > 225, the kernel patch confidence level is also adjusted. In addition, a logistic regression classifier is used for auxiliary verification. If the result is consistent with XGBoost, the confidence level is increased by 0.1; otherwise, it is decreased by 0.1. The initial confidence level of all patches is set to 0.7, and the above rules are applied sequentially, with a maximum confidence level of 1.0. For patches with a corrected confidence level lower than 0.5, a secondary classification is initiated: random forest is used for reclassification. If the new result is consistent with the original result and the new confidence level is > 0.5, the original result is retained. If they are inconsistent and the new confidence level is > 0.7, the new result is adopted; otherwise, the original result is retained and manual review can be selected.

[0047] References: [1] Z. Tang, KV Chuang, C. DeCarli, LW Jin, L. Beckett, MJ Keiser, et al. Interpretable classification of Alzheimer's diseasepathologies with a convolutional neural network pipeline. Nature Communications., 2019, 10: 2173. This invention has many other embodiments, and all technical solutions formed by equivalent transformations or equivalent modifications fall within the protection scope of this invention.

Claims

1. A method for Aβ patch morphological characterization and classification based on multidimensional image features, characterized in that: Includes the following steps: S1: Construct an Aβ patch dataset with accurate labels, including a balanced number of diffuse patches, nuclear patches, and CAA patches, covering multiple mouse ages and brain regions. S2: For each Aβ patch in the Aβ patch dataset, extract low-order morphological features that reflect its macroscopic morphology and intuitive grayscale distribution, and high-order morphological features that characterize its internal texture heterogeneity, spatial complexity patterns and sub-visual radiomics attributes, and construct a feature space that integrates the low-order and high-order morphological features of the Aβ patch. S3: All features in the feature space are screened and verified multiple times. The screened features are classified into diffuse patch features, nuclear patch features and CAA patch features. Each type of feature is divided into key features, main features and secondary features according to their importance. S4: Select a supervised classifier type, assign weights to the selected features based on the importance level of each feature, and train the selected supervised classifier using the Aβ patch dataset containing the selected patch features to obtain a patch morphology classification model. S5: Classify Aβ patches using the patch morphology classification model.

2. The method for Aβ patch morphology characterization and classification based on multidimensional image features according to claim 1, characterized in that: Step S1 includes the following steps: S11: Acquire three-dimensional image data of mouse brains at multiple age stages; S12: Preprocess the mouse brain 3D image data to suppress the background region outside the brain outline, and crop the preprocessed mouse brain 3D image data into several data blocks of the same size. S13: Select multiple sets of mouse data from each age group, and use variable proportion sampling to extract plaque samples from the selected mouse data to obtain a plaque sample set with a balanced number of samples from each age group. S14: A classifier is used to make a preliminary classification on the coronal section of the patch, and the classification is manually checked and corrected by combining the three-dimensional morphology and the two-dimensional section in three orthogonal directions to achieve the classification labeling of patch samples in the patch sample set.

3. The method for Aβ patch morphology characterization and classification based on multidimensional image features according to claim 1, characterized in that: In step S2, the extraction of low-order morphological features for each Aβ patch includes the following steps: Calculate the radius, volume, mean gray value, standard deviation, and skewness of patch Aβ; Calculate the ratio of the longest axis to the shortest axis of the three-dimensional minimum circumscribed cuboid of patch Aβ to obtain the ratio of the longest axis to the shortest axis. Polynomial fitting was performed on the grayscale curve passing through the center point of the Aβ patch to calculate the grayscale difference between the central peak and the two side valleys and secondary peaks, and the larger value was taken as the grayscale peak-valley drop. Within a radius of 1 / 2 from the center of patch Aβ, the maximum gray value is selected as the peak value of the patch signal.

4. The method for Aβ patch morphology characterization and classification based on multidimensional image features according to claim 1, characterized in that: The steps for extracting the higher-order morphological features include: Obtain the original image of the patch and its foreground signal mask; Based on the original image and its foreground signal mask, high-order morphological features are extracted from five dimensions: shape features, gray-level co-occurrence matrix, gray-level size region matrix, adjacent gray-level tone difference matrix, and gray-level dependence matrix.

5. The method for Aβ patch morphology characterization and classification based on multidimensional image features according to claim 1, characterized in that: Step S3 includes the following steps: S31: Perform the Kruskal-Wallis test on each feature in the feature space in turn to determine whether there are significant differences among the three types of plaques: diffuse, nuclear, and CAA. S32: Use bulldozer distance to assess the degree of separation of feature distributions among different morphological categories; S33: Validate the ability of features to distinguish patch types using k-means clustering; S34: Based on the evaluation results of steps S31-S33, the importance of each feature is quantified using an integral system, and the features are divided into key level, primary level and secondary level. S35: Decompose the patch image into 8 frequency domain sub-bands using wavelet transform, and perform the Kruskal-Wallies test again on each sub-band. Retain features that still show significant inter-class differences on at least 4 sub-bands. S36: Principal component analysis is used to perform dimensionality reduction evaluation on each feature set, determine the minimum number of features required to describe each type of patch, and verify whether the number of features in the feature set is sufficient to describe the morphological attributes of the patch to which it belongs.

6. The method for Aβ patch morphology characterization and classification based on multidimensional image features according to claim 1, characterized in that: In step S4, the original features are used for classification, and small sample patches are randomly selected to pre-train commonly used supervised classifiers. The supervised classifier with the highest test accuracy is selected as the best supervised classifier.

7. The method for Aβ patch morphology characterization and classification based on multidimensional image features according to claim 6, characterized in that: In step S4, commonly used supervised classifiers include XGBoost, Random Forest, Extreme Random Tree, Gaussian Naive Bayes, k-Nearest Neighbors, Logistic Regression, Decision Tree, and Support Vector Machine.

8. The method for Aβ patch morphology characterization and classification based on multidimensional image features according to claim 7, characterized in that: Step S4 includes the following steps: S41: Select XGBoost classifier as the best supervised classifier; S42: Obtain the Aβ patch dataset containing the filtered patch features, and divide the Aβ patch dataset into a training set and a test set according to a preset ratio; S43: Perform step-by-step debugging of key parameters of XGBoost to determine the optimal parameter combination; S44: Set the weights for key features, primary features, and secondary features respectively, and assign an initial weight to each feature selected from the feature space; S45: Using the above optimal parameter combination and initial weight settings, train the XGBoost classifier using the training set to obtain the patch morphology classifier; S46: Use the test reagent to evaluate the performance of the patch morphology classifier.

9. The method for Aβ patch morphology characterization and classification based on multidimensional image features according to claim 8, characterized in that: After the patch morphology classifier outputs the classification results, step S6 is also included: introducing low-order features, setting judgment thresholds based on the statistical distribution of low-order features among CAA patches, nuclear patches and diffuse patches, and performing secondary correction on the classification results output by the patch morphology classifier.

Images