A graph convolution-assisted learning method based on hippocampal subfield surfaces

Abstract

The invention discloses a graph convolution-assisted learning method based on the surface of a hippocampal sub-region, and belongs to the field of artificial intelligence. Specifically, the method comprises the following steps: firstly, extracting hippocampal voxel data from MRI images of an ADNI database, and then acquiring hippocampal surface triangular mesh data in combination with a spherical harmonic mapping algorithm, and segmenting the hippocampus and hippocampal sub-regions; then, integrating the sub-regions divided on the hippocampal surface; adding a pooling module on the basis of an existing graph convolutional neural network model to form an improved graph convolutional pooling model, and training the hippocampal surface triangular mesh data after the sub-regions are integrated; finally, after graph convolution and pooling, mapping the trained point features and surface features to the same latitude for splicing, completing the classification task by an MLP classifier, and obtaining a final prediction result; the invention adopts a graph convolutional pooling model that improves convolution kernels and graph pooling, and achieves a higher classification accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and relates to an artificial neural network model, specifically a graph convolution-assisted learning method based on the surface of hippocampal subregions. Background Art

[0002] Alzheimer's disease (AD) is a common and devastating neurodegenerative disease affecting the elderly. As the global population ages, the number of AD patients continues to rise, placing a heavy burden on patients' families and society.

[0003] The hippocampus experiences significant volume shrinkage during the progression of Alzheimer's disease. The hippocampus is divided into several subregions, each with distinct cell types and functional characteristics. Therefore, studying hippocampal subregions is crucial for understanding the brain's memory mechanisms and exploring the causes of Alzheimer's disease.

[0004] Since the development of AD is highly correlated with hippocampal atrophy, many studies have extracted 2D slice images or 3D voxel data from MRI images and classified AD based on whether the hippocampus has atrophied and the degree of atrophy. For example, Lin Weiming et al. [1] preprocessed MRI images, extracted multiple 2.5D images in the hippocampus region, and then used CNN to train and identify these images. Yu Lu et al. [2] used gray-level co-occurrence matrix and run-length matrix to extract the three-dimensional texture features of the hippocampus of each subject, and then used principal component analysis (PCA), linear discriminant analysis (LDA), and nonlinear discriminant analysis (NDA) to preprocess the data. Finally, a back propagation neural network (BP) was used to establish a classification model.

[0005] The raw data of MRI images are three-dimensional voxels, so many studies have extracted three-dimensional voxels of the hippocampus and used neural network models for 3D data for classification. For example, the hippocampus was segmented from MRI images, and the 3D hippocampal voxel data was classified using a convolutional neural network (DenseNet) extended to 3D [3-5]. Katabathula et al. [6] added a Laplace-Beltrami operator to calculate the global shape features of the hippocampus based on DenseCNN, slightly improving the classification accuracy. Sreevani [7] proposed a lightweight 3D deep convolutional network model to process hippocampal MRI to predict AD.

[0006] While 2D slice images or 3D voxel data from MRI images are the primary focus of research on hippocampal deformation and AD classification, surface data best reflects hippocampal atrophy. Due to the difficulties in generating and processing this data and the limited availability of corresponding algorithms and models, this type of research is rarely conducted. While classification accuracy using 2D brain images is reasonable, it fails to capture the primary areas of hippocampal deformation. Using 3D voxels results in excessive computational space and time complexity, and valid voxel points within a voxel are often limited to the outermost layer, resulting in a waste of computational resources.

[0007] The data that can best reflect the degree of hippocampal atrophy is the hippocampal surface composed of several grids. However, there is currently little research on the shape of the hippocampal surface, and there is still a large gap in the relevant research on high-precision hippocampal surfaces.

[0008] References:

[0009] [1]Babcock KR, Page JS, Fallon JR, et al. Adult hippocampalneurogenesis in aging and Alzheimer's disease[J].2021,16(4):681-693.

[0010] [2]Wilson RS, Segawa E, Boyle PA, et al. The natural history of cognitive decline in Alzheimer's disease[J].2012,27(4):1008.

[0011] [3]World-Alzheimer-Report-2021[J].2021:

[0012] [4]Scheltens P, De Strooper B, Kivipelto M, et al. Alzheimer's disease[J].2021,397(10284):1577-1590.

[0013] [5]Alzheimer's As AJ,dementia.2019Alzheimer's disease facts and figures[J].2019,15(3):321-387.

[0014] [6]Breijyeh Z, Karaman RJ M. Comprehensive review on Alzheimer's disease: causes and treatment[J].2020,25(24):5789.

[0015] [7] Liu Yuanyuan, Xiao Shifu. Reliability and validity of commonly used neuropsychological tests and scales for Alzheimer's disease[J]. China Medical Herald, 2011, 8(09):11-14. Summary of the Invention

[0016] To address these issues, the present invention combines the distribution of hippocampal subregions on the hippocampal surface with graph convolutional neural networks and assisted learning to construct a graph convolution-assisted learning method based on the hippocampal subregion surface. This method studies the correlation between hippocampal subregion surface deformation and AD, and verifies the feasibility of using hippocampal surface deformation features for early AD prediction. The method achieves better results in classification tasks than other three-dimensional classification models, and also demonstrates the performance of each hippocampal subregion in the classification task.

[0017] The graph convolution-assisted learning method based on the hippocampal subfield surface is divided into the following steps:

[0018] Step 1: Extract hippocampal voxel data from MRI images in the ADNI database, then use the spherical harmonic mapping algorithm to obtain the hippocampal surface triangulated mesh data and segment the hippocampus and hippocampal subregions;

[0019] The specific segmentation process includes six steps: MRI preprocessing, skull dissection, volume annotation, intensity normalization, white matter segmentation and hippocampal sub-region segmentation.

[0020] The triangular mesh data consists of vertex information (vertex 3D coordinates and vertex hippocampal subregion labels) and face information (the indices of the three vertices that constitute the face);

[0021] The surface of the hippocampus is divided into 12 subregions: parasubiculum, anterior subiculum, subiculum, CA1, CA2 / 3, CA4, dentate gyrus granule cell layer, hippocampal molecular cell layer, hippocampal fimbria, hippocampal tail, and hippocampal fissure.

[0022] Step 2: Integrate the sub-regions divided on the hippocampal surface;

[0023] The labels of the vertices of the sub-regions of the hippocampal surface were integrated, and it was found that the three regions of the parasubiculum, presubiculum and subiculum were closely related and integrated into one region; CA1, CA2 / 3 and CA4 were integrated into one region; the granule cell layer of the dentate gyrus, the molecular cell layer of the hippocampus, the fimbria and the hippocampal fissure were integrated into one region; and the hippocampal tail was a separate region.

[0024] Step 3: Add a pooling module to the existing graph convolutional neural network model to form an improved graph convolutional pooling model, and train it on the hippocampal surface triangular mesh data after integrating the sub-regions;

[0025] The specific process is:

[0026] First, all the hippocampal surface triangular mesh data were randomly divided into training sets and test sets within different categories;

[0027] A standard template was constructed using the NC (Normal Control) sample. For the point data in the triangular mesh data, the distance from each point to the corresponding point on the template was calculated as a feature to reflect the atrophy of the hippocampal surface at that point. The template distance combined with the three-dimensional coordinates of the vertex constituted a four-dimensional vertex feature.

[0028] For the face data in the triangular mesh data, the coordinate information of the three vertices that constitute the face is used to calculate the area, inner product of the face normal and the angles of the three inner angles of each face, and these features are combined into a 5-dimensional face feature.

[0029] Then, two different graph convolutional network models, GCN-V and GCN-F, are used to train the point data and surface data respectively. The corresponding pooling rules are adopted, and the vertex and surface features undergo three rounds of pooling. Each specification of data is first processed by the graph convolution layer and then pooled.

[0030] Step 4: After graph convolution and pooling, the trained point features and surface features are mapped to the same latitude for splicing. The MLP (Multi-Layer Perception) classifier completes the classification task and obtains the final prediction result.

[0031] MLP consists of three fully connected layers: input layer, hidden layer and output layer; the number of hidden layers and neurons is selected according to the specific situation.

[0032] After the graph convolution operation, the features of the vertices of each subregion or integrated region are extracted according to the labels of the vertices of the hippocampal subregion. Then, an MLP classifier is constructed for each subregion, and the prediction results of each subregion are obtained. The loss functions of each subregion are integrated through the automatic weighted loss function to obtain the final loss and prediction results:

[0033]

[0034] Where c i is a trainable parameter, 2·c i 2As the weighted coefficient of each auxiliary task Loss, the weight of Loss is automatically optimized through network training.

[0035] The prediction result is: based on the atrophy of the hippocampus, it is judged to be CN (normal cognition), MCI (mild cognitive impairment), or AD (Alzheimer's disease).

[0036] The advantages of the present invention are:

[0037] 1) A graph convolution-assisted learning method based on the surface of hippocampal subregions, which simultaneously utilizes both vertex and face features, can fully utilize the characteristics of hippocampal surface data; at the same time, prior knowledge of hippocampal subregions is added to the network model to assist the model in completing the classification task.

[0038] 2) A graph convolution-assisted learning method based on the surface of hippocampal subregions, which adopts a graph convolution pooling model with improved convolution kernels and graph pooling, as well as a graph convolution-assisted learning model that incorporates prior knowledge of hippocampal subregions, both of which achieve higher classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flowchart of a graph convolution-assisted learning method based on the surface of hippocampal subregions of the present invention;

[0040] Figure 2 The present invention uses the spherical harmonic mapping algorithm to segment the hippocampus surface and obtain a schematic diagram of triangular mesh data;

[0041] Figure 3 Schematic diagram of the graph convolution-assisted learning model used in this invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the accompanying drawings and examples.

[0043] The present invention provides a graph convolution-assisted learning method based on the surface of hippocampal subregions. Aiming at the correlation between the hippocampal surface and Alzheimer's disease, the graph convolution-assisted learning model is improved by combining the distribution information of hippocampal subregions on the hippocampal surface.

[0044] like Figure 1 As shown, the following steps are taken:

[0045] Step 1: Extract hippocampal voxel data from MRI images in the ADNI database, then use the spherical harmonic mapping algorithm to obtain the hippocampal surface triangulated mesh data and segment the hippocampus and hippocampal subregions;

[0046] like Figure 2 As shown in the figure, the specific segmentation process includes six steps: MRI preprocessing, skull dissection, volume annotation, intensity normalization, white matter segmentation, and hippocampal subregion segmentation:

[0047] Preprocessing refers to: removing noise and bias fields from the original MRI images, improving image contrast, and then registering the images to a standard space for group analysis or atlas comparison;

[0048] Skull stripping refers to removing non-brain tissue from the image and using a neural network to predict the signed distance of each voxel to the boundary between white matter and gray matter, as well as its coordinates in the spherical atlas space, to generate an implicit isosurface as the result of skull stripping.

[0049] Volume annotation refers to: aligning an individual's T1 image with a template image and performing intensity normalization; then, roughly segmenting the aligned T1 image to obtain tissue categories such as white matter, gray matter, and cerebrospinal fluid; classifying each voxel on the segmented T1 image according to the probability atlas; using the Bayesian algorithm, the probability of each structure appearing in space in the atlas and the relationship between each structure and each voxel measurement (such as intensity, direction) are used to calculate the probability that each voxel belongs to each structure, and ultimately determine which anatomical structure each voxel belongs to.

[0050] Intensity normalization refers to the homogenization of the intensity of the original T1 image in order to better distinguish tissue types such as white matter and gray matter and make the segmentation process easier. An algorithm based on mutual information is used to adjust the position and intensity of the image by utilizing the statistical correlation between images.

[0051] White matter segmentation refers to classifying voxels into white matter or non-white matter based on image intensity and spatial information, and using the level set method algorithm to find the optimal segmentation interface by utilizing the intensity difference and continuity between white matter and gray matter.

[0052] Hippocampal subregion segmentation involves resegmenting the image after white matter segmentation and warping it to the input data space using the FS60 statistical atlas (which contains the probability distribution and shape variation of hippocampal subregions) to maximize the similarity between the atlas and the data. Based on the warped atlas and the input data, the posterior probability of each voxel belonging to a different hippocampal subregion is calculated, and discrete labels are assigned according to the maximum a posteriori probability principle.

[0053] The triangular mesh data is composed of vertex information (vertex three-dimensional coordinates and vertex hippocampal subregion labels) and face information (indexes of the three vertices constituting the face);

[0054] The steps of the spherical harmonic mapping algorithm to process the voxel-format hippocampal data into a triangular mesh surface are as follows:

[0055] (1) Binarization and data format conversion: Each voxel in the hippocampal voxel data has a segmented hippocampal sub-region label (the non-hippocampal region label is 0). In order to facilitate subsequent topological repair and spherical parameterization operations, the voxel label needs to be binarized and all hippocampal region labels are set to 1.

[0056] (2) Topological patching, which processes binary voxel data to eliminate noise in 3D binary voxel data and fill surface holes so that the voxel surface has a spherical topology.

[0057] (3) Spherical parameterization, creating a mapping from the voxel surface to the unit sphere.

[0058] (4) Spherical harmonic expansion, which expands the voxel surface into a complete set of spherical harmonic basis functions and uses Fourier coefficients to reconstruct the object surface.

[0059] (5) Spherical harmonic alignment, using FOE to establish surface correspondence and align objects so that each FOE has a standard orientation in object space and parameter space.

[0060] (6) Obtaining surface vertex labels. Since the relevant information of the hippocampal subregions only exists in the segmented voxel data and is eliminated in the entire spherical harmonic mapping process, additional processing is required to transfer the hippocampal subregion labels from the voxel data to the triangular mesh surface.

[0061] First, the patched voxel data is labeled, and the topologically patched binary voxel data (denoted as data A) is compared with the original data. The label of the padded part is the label with the highest frequency within a certain range, and the label of the deleted part is set to 0 (the processed data is denoted as data B); secondly, the voxel label is transferred to the point label. According to the principle of SPHARM, the surface of the voxel in data A is found to obtain the three-dimensional coordinates of all boundary points. Then, based on this coordinate, the hippocampal subfield labels of all surfaces in data B are obtained.

[0062] (7) Generate a triangular mesh surface. For the data after surface registration, map the spherical points onto the template to generate triangular mesh data. There are 6 template specifications, as shown in Table 1. Considering the fineness requirements of the triangular mesh surface of the hippocampus and the parameter limitations of the subsequent artificial neural network model, this example uses the lv4 template as the standard for a single hippocampal surface. The hippocampal voxels of all samples will be mapped onto this template.

[0063] Table 1

[0064]

[0065]

[0066] The labels of the vertices of the sub-regions of the hippocampal surface were integrated, and the hippocampal surface was divided into 12 sub-regions, namely: parasubiculum, anterior subiculum, subiculum, CA1, CA2 / 3, CA4, dentate gyrus granule cell layer, hippocampal molecular cell layer, hippocampal fimbria, hippocampal tail, and hippocampal fissure.

[0067] Step 2: Integrate the sub-regions divided on the hippocampal surface;

[0068] The labels of the vertices of the sub-regions of the hippocampal surface were integrated, and it was found that the three regions of the parasubiculum, presubiculum and subiculum were closely related and integrated into one region; CA1, CA2 / 3 and CA4 were integrated into one region; the granule cell layer of the dentate gyrus, the molecular cell layer of the hippocampus, the fimbria and the hippocampal fissure were integrated into one region; and the hippocampal tail was a separate region.

[0069] Step 3: Add a pooling module to the existing graph convolutional neural network model to form an improved graph convolutional pooling model, and train it on the hippocampal surface triangular mesh data after integrating the sub-regions;

[0070] The specific process is:

[0071] First, all the hippocampal surface triangulated mesh data were randomly divided into training sets and test sets within different categories, with a ratio of 4:1;

[0072] Each node feature on the hippocampal surface includes: vertex three-dimensional coordinates, vertex discrete curvature and vertex template distance;

[0073] The point data in the triangular mesh data reflects the spatial characteristics of the hippocampal morphology. A standard template is constructed using the NC (Normal Control) sample, and the distance from each point to the corresponding point on the template is calculated as a feature to reflect the atrophy status of the hippocampal surface at that point. The template distance is combined with the three-dimensional coordinates of the vertex to form a four-dimensional vertex feature.

[0074] The surface data in the triangular mesh data reflects the structural characteristics of the hippocampus surface. The coordinate information of the three vertices that constitute the surface is used to calculate the area, inner product of the surface normal and the angles of the three internal angles of each surface, and these features are combined into a 5-dimensional surface feature.

[0075] Next, two different graph convolutional network models, GCN-V and GCN-F, were trained on the point data and face data, respectively. Using corresponding pooling rules, vertex and face features underwent three rounds of pooling, concentrating the original data from level 4 to level 1. In the actual network model, these three rounds of pooling are not performed all at once; instead, data of each size is first processed through a graph convolutional layer before pooling. Because the number of data points and faces varies for each size of hippocampal surface, a separate graph convolutional network model was constructed for each size of hippocampal surface.

[0076] Step 4: After graph convolution and pooling, the multi-layer perceptron model (MLP) maps the trained point features and surface features to the same latitude for splicing. The MLP classifier completes the classification task and obtains the final prediction result.

[0077] The MLP consists of three fully connected layers: an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is equal to the number of extracted embedding features. The number of hidden layers is selected based on the specific situation.

[0078] Based on the 11 sub-areas of the hippocampal surface and the 4 sub-areas after integration, the auxiliary task scheme adopted is as follows:

[0079] (1) The first half of the model uses a graph convolutional neural network, which is different from the improved graph convolution pooling model. Since the labels of the hippocampal subregions only exist in the hippocampal surface data of lv4, the pooling operation is discarded in the auxiliary learning model and only one layer of graph convolution is used.

[0080] (2) After the graph convolution operation, the features of the vertices of each subregion or integrated region are extracted according to the labels of the vertices of the hippocampal subregions. Then, an MLP classifier is constructed for each subregion, and the prediction results of each subregion are obtained;

[0081] (3) Through the automatic weighted loss function (Automatic Weighted Loss), the loss functions of each sub-region are integrated to obtain the final loss and prediction results:

[0082]

[0083] Where c i is a trainable parameter, 2·c i 2 As the weighted coefficient of each auxiliary task Loss, the weight of Loss is automatically optimized through network training.

[0084] The prediction result is: based on the atrophy of the hippocampus, it is judged to be CN (normal cognition), MCI (mild cognitive impairment), or AD (Alzheimer's disease).

[0085] Example:

[0086] Step 1: Acquisition of hippocampal surface data.

[0087] MRI images obtained from the ADNI database were segmented into the hippocampus and its subregions. Spherical harmonics were then used to extract hippocampal surface data from the hippocampal voxel data. Each hippocampal surface is composed of 5120 triangles with 2562 vertices. The hippocampal surface is divided into 12 subregions: the parasubiculum, the anterior subiculum, the subiculum, CA1, CA2 / 3, CA4, the granule cell layer of the dentate gyrus, the molecular cell layer of the hippocampus, the fimbria, the hippocampal tail, and the hippocampal fissure.

[0088] Because the surface areas of some hippocampal subregions are too small, the hippocampal surface subregions were integrated: the parasubiculum, presubiculum, and subiculum are closely related and are integrated into one region; CA1, CA2 / 3, and CA4 are integrated into one region; the dentate gyrus granule cell layer, hippocampal molecular cell layer, hippocampal fimbria, and hippocampal fissure are integrated into one region; the hippocampal tail has little connection with other subregions, so it is a separate region; the specific statistics of the number of vertices in the integrated surface of the hippocampal subregions are shown in Table 2.

[0089] Table 2

[0090]

[0091]

[0092] Step 2: To effectively extract surface feature information from the hippocampal surface, the model first performs graph convolution on the hippocampal surface. The features of each node on the hippocampal surface include: the three-dimensional coordinates of the vertex, the discrete curvature of the vertex, and the template distance of the vertex (a standard template is constructed using all CN samples, and the distance from each point to the corresponding point on the template is calculated as a feature to reflect the atrophy of the hippocampal surface at that point).

[0093] Step 3: After the graph convolution operation, the features of the vertices of each sub-region or integrated region are extracted according to the labels of the vertices of the hippocampal sub-region. Then, an MLP classifier is constructed for each sub-region, and the prediction results of each sub-region are obtained.

[0094] Step 4: Through the automatic weighted loss function (Automatic Weighted Loss), the loss functions of each sub-region are integrated to obtain the final loss and prediction results:

[0095]

[0096] The results show that:

[0097] For the divided training and test sets, classification experiments were conducted using MeshCNN, MeshNet, LB-CNN, GCN, and the improved model. The classification effect statistics on the test set are shown in the following table:

[0098] Table 3 Comparison of classification effects of various models in the AD-CN group

[0099]

[0100] Table 4 Comparison of classification effects of various models in the MCI-CN group

[0101]

[0102]

[0103] Table 5 Comparison of classification effects of various models in the AD-MCI group

[0104]

[0105] The results in Tables 3, 4, and 5 show that the proposed model has better classification effects than other classification models in the three groups of classification tasks. Not only has the accuracy been slightly improved, but the recall rate and F1 score are also higher than other models. It is worth noting that the MeshNet and MeshCNN models have a recall rate of 100% or 0% in the classification tasks. This is because the training effect of the model is poor. In the binary classification task, all targets are predicted as a certain category, resulting in a situation where the prediction accuracy is not high but the recall rate and F1 score are too high or too low.

[0106] Table 6 Statistics of classification effects of the four sub-areas on the hippocampal surface

[0107]

[0108] Among the results of different groups, the classification accuracy of the AD-CN group was still the highest. The AD-CN classification effect of the left hippocampus was slightly improved, with an accuracy of about 86-88%, while the classification effect of the right hippocampus was less improved, with an accuracy of about 80%; followed by the MCI-CN group. Compared with other models, the graph convolution-assisted learning model effectively improved the classification effect of this group, with the classification accuracy of the left hippocampus reaching about 73%, and the classification accuracy of the right hippocampus at about 68%; the overall classification effect of the AD-MCI group was average, with an accuracy improvement of only 3-4% compared to the graph convolution pooling model.

[0109] On the left and right sides, the graph convolution-assisted learning model is as follows Figure 3 As shown, the classification results for the left hippocampus were better than those for the right hippocampus. The difference was more pronounced in the AD-CN and MCI-CN groups. In the AD-CN group, the accuracy on the left side was approximately 7% higher than that on the right side, while in the MCI-CN group, the accuracy on the left side was approximately 5% higher than that on the right side. There was no significant difference in the classification results between the left and right sides in the AD-MCI group.

[0110] Different hippocampal subregions achieved different classification results and contributed differently to the overall model. The parasubiculum, presubiculum, subiculum, CA1, and CA2 / 3 subregions performed well on both left and right classification tasks, while the hippocampal tail performed moderately well on the right hippocampus.

Claims

1. A graph convolution-assisted learning method based on the surface of hippocampal subfields, It is characterized in that The following steps are taken: Step 1: Extract hippocampal voxel data from MRI images in the ADNI database, and then use the spherical harmonic mapping algorithm to obtain the hippocampal surface triangular mesh data, and segment the hippocampus and hippocampal sub-regions; The triangular mesh data consists of vertex information and face information. The vertex information includes the three-dimensional coordinates of the vertex and the hippocampal sub-area label of the vertex; the face information includes the indexes of the three vertices constituting the face; The surface of the hippocampus is divided into 12 sub-regions, namely: parasubiculum, anterior subiculum, subiculum, CA1, CA2 / 3, CA4, dentate gyrus granule cell layer, hippocampal molecular cell layer, hippocampal fimbria, hippocampal tail, and hippocampal fissure; Step 2: Integrate the sub-regions divided on the hippocampal surface; The labels of the vertices of the sub-regions on the hippocampal surface were integrated, and the following results were obtained: the parasubiculum, presubiculum, and subiculum were closely related and integrated into one region; CA1, CA2 / 3, and CA4 were integrated into one region; the granule cell layer of the dentate gyrus, the molecular cell layer of the hippocampus, the fimbria, and the hippocampal fissure were integrated into one region; and the hippocampal tail was a separate region; Step 3: Add a pooling module to the existing graph convolutional neural network model to form an improved graph convolutional pooling model, and train the hippocampal surface triangular mesh data after integrating the sub-regions; The specific process is: First, all the hippocampal surface triangular mesh data were randomly divided into training sets and test sets in different categories; A standard template was constructed using the NC sample. For the point data in the triangular mesh data, the distance from each point to the corresponding point on the template was calculated as a feature to reflect the atrophy of the hippocampal surface at that point. The template distance combined with the three-dimensional coordinates of the vertex formed a 4-dimensional vertex feature. For the face data in the triangular mesh data, the coordinate information of the three vertices constituting the face is used to calculate the area of ​​each face, the inner product of the face normal and the angles of the three inner angles, and these features are combined into a 5-dimensional face feature; Then, two different graph convolutional network models, GCN-V and GCN-F, were used to train the point data and surface data. The corresponding pooling rules were used, and the vertex and surface features were pooled three times. Each specification of data was first processed by the graph convolution layer and then pooled. Step 4: After graph convolution and pooling, the trained point features and surface features are mapped to the same latitude for splicing, and the MLP classifier completes the classification task and obtains the final prediction result; After the graph convolution operation, the features of the vertices of each sub-region or integrated region are extracted according to the labels of the vertices of the hippocampal sub-regions. Then, an MLP classifier is constructed for each sub-region, and the prediction results of each sub-region are obtained. Through the automatic weighted loss function, the loss functions of each sub-region are integrated to obtain the final loss and prediction results: In the formula, c i is a trainable parameter, 2·c i 2 As the weighted coefficient of each auxiliary task Loss, the weight of Loss is automatically optimized through network training; The prediction result is: judging whether it belongs to CN, MCI, or AD based on the atrophy of the hippocampus.

2. A graph convolution-assisted learning method based on the surface of hippocampal subregions as claimed in claim 1, It is characterized in that The process of segmenting the hippocampus and hippocampal sub-regions using the spherical harmonic mapping algorithm includes six steps: MRI preprocessing, skull dissection, volume labeling, intensity normalization, white matter segmentation and hippocampal sub-region segmentation.

3. A graph convolution-assisted learning method based on the surface of hippocampal subregions as claimed in claim 1, It is characterized in that The MLP includes three fully connected layers: an input layer, a hidden layer and an output layer; the number of hidden layers and the number of neurons are selected according to specific circumstances.

Images