Alzheimer's disease prediction model training method, Alzheimer's disease prediction method and device

Abstract

The present application provides a method for training an Alzheimer's disease prediction model, a method for predicting Alzheimer's disease, and a device. The training method includes: using a ventricle contour point cloud of a magnetic resonance imaging sample and a first deep learning model to obtain first Alzheimer's disease prediction result data; using ventricle volume ratio data corrected by covariates and a second deep learning model to obtain second Alzheimer's disease prediction result data; and centrally training an SVM model with the first Alzheimer's disease prediction result data, the second Alzheimer's disease prediction result data, and classification labels to predict the Alzheimer's disease prediction fusion result. The present application can focus on the morphological features of the ventricles and achieve accurate characterization of the ventricle shape, can effectively eliminate interference from irrelevant factors, can use multi-dimensional ventricle data for integrated learning, and can effectively improve the classification performance and generalization ability of the Alzheimer's disease prediction model, thereby effectively improving the efficiency and accuracy of Alzheimer's disease prediction.

Description

Technical Field

[0001] The present application relates to the technical field of disease-assisted diagnosis, and in particular to an Alzheimer's disease prediction model training method, an Alzheimer's disease prediction method and a device. Background Art

[0002] Alzheimer's disease (AD) is the most common type of dementia and is a progressive neurodegenerative disease. The disease is characterized by the appearance of amyloid plaques and tau protein tangles in the brain. These abnormal structures can cause neuronal damage and death, ultimately causing cognitive dysfunction in patients and affecting their ability to function in daily life. Early diagnosis is crucial to slowing the progression of AD and improving the quality of life of patients. However, the early symptoms of AD are often subtle and easily confused with age-related cognitive decline, which poses a challenge to timely diagnosis. Magnetic Resonance Imaging (MRI) is a non-invasive medical imaging technology that can provide high-resolution brain structural information and is widely used in the diagnosis and research of AD. By analyzing MRI images, brain structural changes associated with AD, such as brain atrophy and white matter lesions, can be identified.

[0003] The existing MRI-based AD diagnosis method can adopt a method based on manual analysis: experienced radiologists observe MRI images with the naked eye to identify brain structural changes related to AD, such as hippocampal atrophy and ventricular enlargement. However, this method relies on the subjective judgment of doctors, is easily affected by differences between observers, and has low efficiency. Therefore, a method based on computer-aided diagnosis is proposed: using computer programs to automatically analyze MRI images, extract image features related to AD, and establish a classification model for diagnosis. Commonly used image features include brain tissue volume, cortical thickness, white matter integrity, etc. Commonly used classification models include support vector machines (SVM), random forests (RF), deep learning, etc.

[0004] However, existing computer-assisted MRI-based AD diagnostic methods do not fully utilize information from specific brain regions and usually require analysis of whole-brain MRI images, which is computationally intensive and time-consuming, making it unfavorable for clinical application. They also have high requirements for image quality, and whole-brain image analysis is easily affected by artifacts and noise, such as head movement, metal artifacts, etc., resulting in a decrease in diagnostic accuracy. Summary of the invention

[0005] In view of this, the embodiments of the present application provide an Alzheimer's disease prediction model training method, an Alzheimer's disease prediction method and a device to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present application provides a method for training an Alzheimer's disease prediction model, comprising:

[0007] Inputting the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample into a preset first deep learning model, so that the first deep learning model outputs first Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample;

[0008] And, inputting the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample into a preset second deep learning model, so that the second deep learning model outputs second Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample;

[0009] The first Alzheimer's disease prediction result data, the second Alzheimer's disease prediction result data and the preset Alzheimer's disease classification label corresponding to each of the magnetic resonance imaging samples are respectively used as centralized learning samples corresponding to each of the magnetic resonance imaging samples, and a preset SVM model is trained based on each of the centralized learning sample training sets, so that the SVM model is used to output the Alzheimer's disease prediction fusion result data of the magnetic resonance imaging sample according to the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data of the magnetic resonance imaging sample, and together with the first deep learning model and the second deep learning model, constitute an Alzheimer's disease prediction model.

[0010] In some embodiments of the present application, before inputting the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample into a preset first deep learning model, the method further includes:

[0011] Extract the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample;

[0012] The ventricle contour point cloud data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label are respectively used as the first training samples corresponding to each of the magnetic resonance imaging samples, and a preset first deep learning model is trained based on each of the first training samples, so that the first deep learning model is used to output the first Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricle contour point cloud data corresponding to the magnetic resonance imaging sample.

[0013] In some embodiments of the present application, before inputting the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample into the preset second deep learning model, the method further includes:

[0014] Extract the covariate-corrected ventricular volume ratio data corresponding to each MRI sample;

[0015] The ventricular volume ratio data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label are respectively used as second training samples corresponding to each of the magnetic resonance imaging samples, and a preset second deep learning model is trained based on each of the second training samples, so that the first deep learning model is used to output the second Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricular volume ratio data corresponding to the magnetic resonance imaging sample.

[0016] In some embodiments of the present application, the first deep learning model includes: a feature encoder, a hierarchical feature extraction model, a global feature aggregation layer, and a classifier connected in sequence;

[0017] The feature encoder is used to extract the feature code of each point in the currently input ventricle contour point cloud data;

[0018] The hierarchical feature extraction model is used to perform multi-layer grouping feature extraction on the ventricle contour point cloud data according to the feature coding of each point in the ventricle contour point cloud data, so as to obtain an initial global feature vector corresponding to the ventricle contour point cloud data;

[0019] The global feature aggregation layer is used to perform feature aggregation processing on the initial global feature vector to obtain a global feature vector corresponding to the ventricle contour point cloud data;

[0020] The classifier is used to output first Alzheimer's disease prediction result data corresponding to the ventricle contour point cloud data according to the global feature vector corresponding to the ventricle contour point cloud data.

[0021] In some embodiments of the present application, the second deep learning model includes: a logistic regression model;

[0022] The logistic regression model is used to output the predicted probability of Alzheimer's disease corresponding to the input ventricular volume ratio data after covariate correction based on the activation function, so as to use the predicted probability of Alzheimer's disease as the second Alzheimer's disease prediction result data of the magnetic resonance imaging sample corresponding to the ventricular volume ratio data.

[0023] In some embodiments of the present application, extracting the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample includes:

[0024] Acquire ventricle segmentation image data corresponding to each magnetic resonance imaging data in a preset magnetic resonance image library;

[0025] Based on a breadth-first search algorithm, the boundary points of the ventricle region in each of the ventricle segmentation image data are respectively extracted to form initial ventricle contour point cloud data corresponding to each magnetic resonance imaging data;

[0026] The farthest point sampling algorithm is adopted to sample each of the initial ventricle contour point cloud data respectively according to a preset sampling point threshold value, so as to obtain the ventricle segmentation image data corresponding to each magnetic resonance imaging data.

[0027] In some embodiments of the present application, the step of acquiring ventricle segmentation image data corresponding to each magnetic resonance imaging data in a preset magnetic resonance image library includes:

[0028] Selecting a preset number of magnetic resonance imaging data from a preset magnetic resonance image library to construct a spectrogram library, and receiving and storing ventricle segmentation image data corresponding to each of the magnetic resonance imaging data in the spectrogram library;

[0029] and performing skull image removal processing on each magnetic resonance imaging data in the magnetic resonance image library to obtain skull-removed image data corresponding to each magnetic resonance imaging data in the magnetic resonance image library;

[0030] A preset multi-atlas registration and voting strategy is adopted, and based on the ventricle segmentation image data corresponding to each magnetic resonance imaging data in the atlas library, ventricle segmentation is performed on the skull-removed image data corresponding to each magnetic resonance imaging data that has not currently undergone ventricle segmentation in the magnetic resonance image library, so as to obtain the ventricle segmentation image data corresponding to each magnetic resonance imaging data in the magnetic resonance image library.

[0031] In some embodiments of the present application, the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample includes:

[0032] Determining the initial ventricular volume ratios corresponding to the respective magnetic resonance imaging samples according to the ventricular volumes and the whole brain volumes corresponding to the respective magnetic resonance imaging samples;

[0033] Based on the pre-fitted correspondence between the individual age corresponding to the magnetic resonance imaging and the initial ventricular volume ratio, the initial ventricular volume ratio corresponding to each magnetic resonance imaging sample is respectively corrected for the individual age to obtain the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample.

[0034] In some embodiments of the present application, the feature encoder includes: an MLP model, and each layer in the MLP model includes a connected linear transformation layer and a nonlinear activation function layer;

[0035] Among them, the linear transformation layer is used to multiply the vector of each point in the currently input ventricular contour point cloud data by the weight matrix respectively, and add a bias term to output the linear transformation result data corresponding to each point in the currently input ventricular contour point cloud data; the nonlinear activation function is used to perform nonlinear feature learning on each linear transformation result data output by the linear transformation layer of the same layer based on the activation function, so that the MLP model obtains the feature coding of each point in the ventricular contour point cloud data according to the linear transformation layer and nonlinear activation function layer of each layer.

[0036] In some embodiments of the present application, the hierarchical feature extraction model includes: a PointNet++ network, and the PointNet++ network includes: a first feature extraction layer, a second feature extraction layer, and a third feature extraction layer connected in sequence;

[0037] The first feature extraction layer is used to select a first number of first sampling points in the ventricle contour point cloud data by adopting the farthest point sampling algorithm, and respectively divide the points in the spherical neighborhoods of different radii centered on each of the first sampling points to form point sets corresponding to different radii of each of the first sampling points, respectively extract the feature codes of each point in each of the point sets by adopting the PointNet network to obtain feature data corresponding to each of the points in each of the point sets, and then concatenate and fuse the point sets corresponding to different radii of each of the first sampling points to obtain first fused feature vectors corresponding to each of the first sampling points;

[0038] The second feature extraction layer is used to select a second number of second sampling points from each of the first sampling points in the ventricle contour point cloud data by using the farthest point sampling algorithm, where the second number is smaller than the first number, and respectively divide the points in the spherical neighborhoods of different radii centered on each of the second sampling points to form point sets corresponding to different radii of each of the second sampling points, respectively extract features from the first fused feature vectors of each of the first sampling points in each of the point sets by using the PointNet network to obtain feature data corresponding to each of the first sampling points in each of the point sets, and then concatenate and fuse the point sets corresponding to different radii of each of the second sampling points to obtain second fused feature vectors corresponding to each of the second sampling points;

[0039] The third feature extraction layer is used to perform feature extraction on the second fused feature vectors corresponding to each of the second sampling points, so as to obtain an initial global feature vector corresponding to the ventricle contour point cloud data.

[0040] In some embodiments of the present application, the classifier includes: a first fully connected layer, a second fully connected layer, a third fully connected layer, and an activation function layer connected in sequence;

[0041] The first fully connected layer is used to perform linear transformation and nonlinear activation on the global feature vector corresponding to the ventricle contour point cloud data output by the global feature aggregation layer, so as to output a first feature vector corresponding to the ventricle contour point cloud data;

[0042] The second fully connected layer is used to perform linear transformation and nonlinear activation on the first feature vector to output a second feature vector corresponding to the ventricle contour point cloud data;

[0043] The third fully connected layer is used to perform linear transformation and nonlinear activation on the second eigenvector to output a third eigenvector corresponding to the ventricle contour point cloud data;

[0044] The activation function layer is used to convert the third eigenvector corresponding to the ventricle contour point cloud data into a probability distribution based on the activation function, so as to obtain the first Alzheimer's disease prediction result data corresponding to the ventricle contour point cloud data.

[0045] A second aspect of the present application provides a method for predicting Alzheimer's disease, comprising:

[0046] Inputting the ventricle contour point cloud data corresponding to the target magnetic resonance imaging data into a first deep learning model in an Alzheimer's disease prediction model, so that the first deep learning model outputs first Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data, wherein the Alzheimer's disease prediction model is pre-trained based on the Alzheimer's disease prediction model training method;

[0047] And, inputting the covariate-corrected ventricular volume ratio data corresponding to the target magnetic resonance imaging data into the second deep learning model in the Alzheimer's disease prediction model, so that the second deep learning model outputs second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data;

[0048] The first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data are input into the SVM model in the Alzheimer's disease prediction model, so that the SVM model outputs the Alzheimer's disease prediction fusion result data corresponding to the target magnetic resonance imaging data.

[0049] The third aspect of the present application provides an Alzheimer's disease prediction model training device, comprising:

[0050] A first base learner application module is used to input the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample into a preset first deep learning model, so that the first deep learning model outputs first Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample;

[0051] and, a second base learner application module, for inputting the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample into a preset second deep learning model, so that the second deep learning model outputs second Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample;

[0052] A meta-learner training module is used to use the first Alzheimer's disease prediction result data, the second Alzheimer's disease prediction result data and the preset Alzheimer's disease classification label corresponding to each of the magnetic resonance imaging samples as the centralized learning samples corresponding to each of the magnetic resonance imaging samples, and train a preset SVM model based on each of the centralized learning sample training sets, so that the SVM model is used to output the Alzheimer's disease prediction fusion result data of the magnetic resonance imaging sample according to the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data of the magnetic resonance imaging sample, and together with the first deep learning model and the second deep learning model, form an Alzheimer's disease prediction model.

[0053] In some embodiments of the present application, the Alzheimer's disease prediction model training device further includes:

[0054] A shape information extraction module is used to extract the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample;

[0055] The first base learner training module is used to use the ventricle contour point cloud data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label as the first training samples corresponding to each of the magnetic resonance imaging samples, and train a preset first deep learning model based on each of the first training samples, so that the first deep learning model is used to output the first Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricle contour point cloud data corresponding to the magnetic resonance imaging sample.

[0056] In some embodiments of the present application, the Alzheimer's disease prediction model training device further includes:

[0057] A volume information extraction module extracts the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample;

[0058] The second base learner training module is used to use the ventricular volume ratio data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label as the second training samples corresponding to each of the magnetic resonance imaging samples, and train a preset second deep learning model based on each of the second training samples, so that the first deep learning model is used to output the second Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricular volume ratio data corresponding to the magnetic resonance imaging sample.

[0059] The fourth aspect of the present application provides an Alzheimer's disease prediction device, comprising:

[0060] A first base learner application module is used to input the ventricle contour point cloud data corresponding to the target magnetic resonance imaging data into a first deep learning model in the Alzheimer's disease prediction model, so that the first deep learning model outputs first Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data, wherein the Alzheimer's disease prediction model is pre-trained based on the Alzheimer's disease prediction model training method;

[0061] and, a second base learner application module, for inputting the covariate-corrected ventricular volume ratio data corresponding to the target magnetic resonance imaging data into a second deep learning model in the Alzheimer's disease prediction model, so that the second deep learning model outputs second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data;

[0062] The meta-learner application module is used to input the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data into the SVM model in the Alzheimer's disease prediction model, so that the SVM model outputs the Alzheimer's disease prediction fusion result data corresponding to the target magnetic resonance imaging data.

[0063] The fifth aspect of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the Alzheimer's disease prediction model training method is implemented, and / or the Alzheimer's disease prediction method is implemented.

[0064] The sixth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the Alzheimer's disease prediction model training method and / or the Alzheimer's disease prediction method.

[0065] The seventh aspect of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the Alzheimer's disease prediction model training method and / or implements the Alzheimer's disease prediction method.

[0066] The Alzheimer's disease prediction model training method provided in the present application comprises the following steps: inputting the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample into a preset first deep learning model, so that the first deep learning model outputs the first Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample; and inputting the covariate-corrected ventricle volume ratio data corresponding to each magnetic resonance imaging sample into a preset second deep learning model, so that the second deep learning model outputs the second Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample; using the first Alzheimer's disease prediction result data, the second Alzheimer's disease prediction result data and the preset Alzheimer's disease classification label corresponding to each of the magnetic resonance imaging samples as the concentrated learning samples corresponding to each of the magnetic resonance imaging samples, and training a preset SVM model based on the training set of each of the concentrated learning samples, so that the SVM model is used to predict the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data according to the magnetic resonance imaging sample. The Alzheimer's disease prediction result data is outputted corresponding to the Alzheimer's disease prediction fusion result data of the magnetic resonance imaging sample, and together with the first deep learning model and the second deep learning model, the Alzheimer's disease prediction model is formed, focusing on the specific brain area of ​​the ventricle, which can avoid the large amount of calculation and high image quality requirements brought by the whole brain image analysis, making the analysis process more efficient and faster, and easier to apply to clinical practice; the deep learning model is used to extract and classify the ventricle shape information, which can automatically learn complex shape features, avoid the limitations of manually designed features, and improve the classification accuracy; in view of the influence of age and individual brain volume differences on the ventricle volume, the covariate correction is used to eliminate the interference of irrelevant factors on the classification results, so that the model pays more attention to the changes in the ventricle volume itself, thereby improving the classification accuracy; the multi-dimensional data of the ventricle is used for integrated learning, and the classification performance and generalization ability of the Alzheimer's disease prediction model can be effectively improved, the calculation speed is fast, and it can meet the needs of clinical applications, thereby effectively improving the efficiency and accuracy of Alzheimer's disease prediction.

[0067] Additional advantages, purposes, and features of the present application will be partially described in the following description, and will become partially apparent to those skilled in the art after studying the following, or may be learned from the practice of the present application. The purposes and other advantages of the present application can be achieved and obtained by the structures specifically pointed out in the specification and the drawings.

[0068] Those skilled in the art will understand that the purposes and advantages that can be achieved by the present application are not limited to the above specific description, and the above and other purposes that can be achieved by the present application will be more clearly understood based on the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The drawings described herein are used to provide a further understanding of the present application, constitute a part of the present application, and do not constitute a limitation of the present application. The components in the drawings are not drawn to scale, but are only for the purpose of illustrating the principles of the present application. In order to facilitate the illustration and description of some parts of the present application, the corresponding parts in the drawings may be enlarged, that is, they may become larger relative to other components in the exemplary device actually manufactured according to the present application. In the drawings:

[0070] Figure 1 Schematic diagram of a first flow chart of a method for training an Alzheimer's disease prediction model in an embodiment of the present application.

[0071] Figure 2 Schematic diagram of a second flow chart of the Alzheimer's disease prediction model training method in one embodiment of the present application.

[0072] Figure 3 Schematic diagram of the architecture of the first deep learning model in one embodiment of the present application.

[0073] Figure 4 Schematic diagram of the architecture of the second deep learning model in one embodiment of the present application.

[0074] Figure 5 Schematic diagram of a third process of the Alzheimer's disease prediction model training method in one embodiment of the present application.

[0075] Figure 6 This is a fourth flow chart of the Alzheimer's disease prediction model training method in one embodiment of the present application.

[0076] Figure 7 This is a schematic diagram of the execution logic of the first deep learning model in one embodiment of the present application.

[0077] Figure 8 FIG. 1 is a schematic diagram of the architecture of a feature encoder in an embodiment of the present application.

[0078] Fig. 9 Schematic diagram of the architecture of a hierarchical feature extraction model in one embodiment of the present application.

[0079] Fig.10 Schematic diagram of the architecture of a classifier in one embodiment of the present application.

[0080] Fig.11 Schematic diagram of the process of predicting Alzheimer's disease in one embodiment of the present application.

[0081] Fig.12 This is a flowchart of a deep learning-based Alzheimer's disease prediction model training and classification method in an application example of the present application.

[0082] Fig.13 This is a schematic diagram of the application process of the second basis learner in an application example of the present application.

[0083] Fig.14 This is a schematic diagram of the first structure of an Alzheimer's disease prediction model training device in one embodiment of the present application.

[0084] Fig.15 This is a second structural schematic diagram of the Alzheimer's disease prediction model training device in one embodiment of the present application.

[0085] Fig.16 Schematic diagram of the structure of an Alzheimer's disease prediction device in one embodiment of the present application. DETAILED DESCRIPTION

[0086] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the implementation modes and the accompanying drawings. Here, the illustrative implementation modes and descriptions of the present application are used to explain the present application, but are not intended to limit the present application.

[0087] It should also be noted here that in order to avoid obscuring the present application due to unnecessary details, only the structures and / or processing steps closely related to the scheme according to the present application are shown in the accompanying drawings, while other details that are not very relevant to the present application are omitted.

[0088] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0089] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0090] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0091] AD classification models based on deep learning, such as using convolutional neural networks (CNNs) to classify whole-brain MRI images. Previous studies have shown that using 3D CNNs to classify MRI images in the ADNI database can achieve a high accuracy rate. However, these methods still rely on whole-brain information, and have problems such as high computational complexity and high image quality requirements. Specifically, existing MRI-based AD diagnostic methods have the following limitations:

[0092] (1) Dependence on whole-brain information: Most methods require analysis of whole-brain MRI images, which is computationally intensive and time-consuming, making it difficult for clinical applications.

[0093] (2) High requirements for image quality: Whole-brain imaging analysis is easily affected by artifacts and noise, such as head movement and metal artifacts, which leads to a decrease in diagnostic accuracy.

[0094] (3) Underutilization of information from specific brain regions: Some studies have shown that specific brain regions, such as the hippocampus and cerebral ventricles, undergo significant changes in the early stages of AD, while existing methods often ignore the information from these specific brain regions.

[0095] In order to solve the above-mentioned problems existing in the existing MRI-based AD diagnostic methods, the embodiments of the present application respectively provide an Alzheimer's disease prediction model training method, an Alzheimer's disease prediction model training device for executing the Alzheimer's disease prediction model training method, an Alzheimer's disease prediction method, an Alzheimer's disease prediction device for executing the Alzheimer's disease prediction method, an electronic device, a computer-readable storage medium and a computer program product, which can solve the problems of some MRI-based AD diagnostic methods, such as large amount of calculation, long time consumption, high requirements for image quality and insufficient utilization of specific brain area information.

[0096] The details are described in detail through the following examples.

[0097] Based on this, the present application embodiment provides an Alzheimer's disease prediction model training method that can be implemented by an Alzheimer's disease prediction model training device, see Figure 1 The Alzheimer's disease prediction model training method specifically includes the following contents:

[0098] Step 100: Input the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample into a preset first deep learning model, so that the first deep learning model outputs the first Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample.

[0099] In one or more embodiments of the present application, the magnetic resonance imaging sample refers to a sample corresponding to magnetic resonance imaging data generated after performing magnetic resonance imaging (MRI) on a human brain.

[0100] It can be understood that the first Alzheimer's disease prediction result data, the second Alzheimer's disease prediction result data and the Alzheimer's disease prediction fusion result number mentioned in one or more embodiments of the present application refer to Alzheimer's disease prediction result data output by different models, and the Alzheimer's disease prediction result data is used to indicate whether the subject corresponding to the magnetic resonance imaging sample suffers from Alzheimer's disease, and specifically can be composed of a probability value of suffering from Alzheimer's disease, a probability value of not suffering from Alzheimer's disease and a probability value that cannot be determined.

[0101] That is to say, step 100 focuses on the ventricle, a specific brain area, avoiding the large amount of calculation and high image quality requirements brought by whole-brain image analysis, making the analysis process more efficient and faster, and easier to apply in clinical practice. In addition, step 100 is based on deep learning to achieve accurate characterization of the ventricle shape: this application adopts a first deep learning model to extract and classify ventricle shape information, which can automatically learn complex shape features, avoid the limitations of manually designed features, and improve classification accuracy.

[0102] Step 200: Input the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample into a preset second deep learning model, so that the second deep learning model outputs second Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample.

[0103] In step 200, in view of the fact that ventricular volume is affected by age and individual brain volume differences, the present application introduces a covariate correction method for age, which can eliminate the interference of irrelevant factors on the classification results, so that the second deep learning model pays more attention to the changes in the ventricular volume itself, thereby improving the classification accuracy.

[0104] Step 300: The first Alzheimer's disease prediction result data, the second Alzheimer's disease prediction result data and the preset Alzheimer's disease classification label corresponding to each of the magnetic resonance imaging samples are respectively used as the centralized learning samples corresponding to each of the magnetic resonance imaging samples, and a preset SVM model is trained based on each of the centralized learning sample training sets, so that the SVM model is used to output the Alzheimer's disease prediction fusion result data of the magnetic resonance imaging sample according to the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data of the magnetic resonance imaging sample, and together with the first deep learning model and the second deep learning model, form an Alzheimer's disease prediction model.

[0105] In order to more comprehensively utilize the ventricular information, step 300 adopts an integrated learning method to fuse the prediction results of the first deep learning model and the second deep learning model, and uses a support vector machine (SVM model) as a meta-learner to further improve the classification performance and generalization ability of the model.

[0106] From the above description, it can be seen that the Alzheimer's disease prediction model training method provided in the embodiment of the present application focuses on the specific brain area of ​​the ventricle, which can avoid the large amount of calculation and high image quality requirements brought by the whole brain image analysis, making the analysis process more efficient and faster, and easier to apply in clinical practice; the deep learning model is used to extract and classify the ventricle shape information, which can automatically learn complex shape features, avoid the limitations of manually designed features, and improve the classification accuracy; in view of the influence of age and individual brain volume differences on the ventricle volume, the interference of irrelevant factors on the classification results is eliminated through covariate correction, so that the model pays more attention to the changes in the ventricle volume itself, thereby improving the classification accuracy; the multi-dimensional data of the ventricle is used for integrated learning, and the classification performance and generalization ability of the Alzheimer's disease prediction model can be effectively improved, the calculation speed is fast, and it can meet the needs of clinical applications, thereby effectively improving the efficiency and accuracy of Alzheimer's disease prediction.

[0107] In order to further improve the application effectiveness and training efficiency of the first deep learning model, in the Alzheimer's disease prediction model training method provided in the embodiment of the present application, see Figure 2 , the Alzheimer's disease prediction model training method further specifically includes the following contents before step 100:

[0108] Step 010: extracting the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample.

[0109] Step 020: The ventricle contour point cloud data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label are respectively used as the first training samples corresponding to each of the magnetic resonance imaging samples, and a preset first deep learning model is trained based on each of the first training samples, so that the first deep learning model is used to output the first Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricle contour point cloud data corresponding to the magnetic resonance imaging sample.

[0110] In order to further improve the application effectiveness and reliability of the second deep learning model, in the Alzheimer's disease prediction model training method provided in the embodiment of the present application, see Figure 2 , the Alzheimer's disease prediction model training method further specifically includes the following contents before step 200:

[0111] Step 030: extracting the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample.

[0112] Step 040: The ventricular volume ratio data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label are respectively used as second training samples corresponding to each of the magnetic resonance imaging samples, and a preset second deep learning model is trained based on each of the second training samples, so that the first deep learning model is used to output the second Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricular volume ratio data corresponding to the magnetic resonance imaging sample.

[0113] In order to further improve the application effectiveness and reliability of the first deep learning model, in the Alzheimer's disease prediction model training method provided in the embodiment of the present application, see Figure 3 The first deep learning model in the Alzheimer's disease prediction model training method specifically includes the following contents:

[0114] The feature encoder 1, the hierarchical feature extraction model 2, the global feature aggregation layer 3 and the classifier 4 are connected in sequence;

[0115] The feature encoder 1 is used to extract the feature code of each point in the currently input ventricle contour point cloud data;

[0116] The hierarchical feature extraction model 2 is used to perform multi-layer grouping feature extraction on the ventricle contour point cloud data according to the feature coding of each point in the ventricle contour point cloud data, so as to obtain an initial global feature vector corresponding to the ventricle contour point cloud data;

[0117] The global feature aggregation layer 3 is used to perform feature aggregation processing on the initial global feature vector to obtain a global feature vector corresponding to the ventricle contour point cloud data;

[0118] The classifier 4 is used for outputting first Alzheimer's disease prediction result data corresponding to the ventricle contour point cloud data according to the global feature vector corresponding to the ventricle contour point cloud data.

[0119] In order to further improve the application effectiveness and reliability of the first deep learning model, in the Alzheimer's disease prediction model training method provided in the embodiment of the present application, see Figure 4 The second deep learning model in the Alzheimer's disease prediction model training method specifically includes the following contents:

[0120] Logistic regression model 5; the logistic regression model is used to output the Alzheimer's disease prediction probability corresponding to the ventricular volume ratio data based on the activation function according to the input ventricular volume ratio data corrected by covariates, so as to use the Alzheimer's disease prediction probability as the second Alzheimer's disease prediction result data of the magnetic resonance imaging sample corresponding to the ventricular volume ratio data.

[0121] Specifically, the second deep learning model is actually a volume classification model in terms of function. When constructing the volume classification model, the logistic regression model 5 is selected as the volume classification model. The logistic regression model 5 is a simple and efficient linear classification model, which is particularly suitable for processing binary classification problems with a single numerical feature. In this application, the corrected ventricular volume ratio is used as a feature to train a logistic regression model to predict AD.

[0122] Its core is to convert the output of the linear model into a probability value between 0 and 1, indicating the probability that the sample belongs to the positive class. The expression of the sigmoid function is as follows:

[0123]

[0124] Where: x is the input feature, which is the corrected ventricular volume ratio in this application; θ is the model parameter, including weight coefficients and bias terms; is the transpose of the parameter vector θ It is the output of the model, indicating the probability that the sample belongs to AD.

[0125] The training goal of the logistic regression model is to find the optimal model parameter θ so that the model can accurately predict the category y of the sample (0 represents normal and 1 represents AD) based on the input feature x.

[0126] In order to measure the accuracy of model prediction, a loss function is used to evaluate the difference between the model prediction value and the true label. The commonly used loss function for logistic regression is the binary cross-entropy loss function (Binary Cross-Entropy Loss), J(θ), which is expressed as follows:

[0127]

[0128] Where m is the number of training samples; is the characteristic of the i-th sample; is the true label of the i-th sample; is the predicted probability of the model for the i-th sample.

[0129] In order to minimize the loss function, the gradient descent method is used to update the model parameters θ. The basic idea of ​​the gradient descent method is to update the parameters in the opposite direction of the gradient of the loss function, thereby gradually reducing the value of the loss function. The parameter update formula is as follows:

[0130]

[0131] in, is the jth parameter in the model parameter vector θ; “:=” is an assignment symbol, which is used to define a new symbol; α is the learning rate, which controls the step size of each parameter update; is the loss function over parameters The partial derivative of , indicating the influence of this parameter on the loss function;

[0132] By continuously iterating the training data and updating the model parameters using the gradient descent method, we can eventually find a set of optimal model parameters θ, which enables the model to accurately predict AD based on the corrected ventricular volume ratio.

[0133] In order to further improve the application effectiveness and reliability of the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample, in a method for training an Alzheimer's disease prediction model provided in an embodiment of the present application, see Figure 5 Step 010 in the Alzheimer's disease prediction model training method specifically includes the following contents:

[0134] Step 011: Acquire ventricle segmentation image data corresponding to each magnetic resonance imaging data in a preset magnetic resonance image library.

[0135] Step 012: Based on a breadth-first search algorithm, the boundary points of the ventricle region in each of the ventricle segmentation image data are extracted respectively to form initial ventricle contour point cloud data corresponding to each magnetic resonance imaging data.

[0136] Specifically, the breadth-first search algorithm (BFS) is applied to extract the boundary points of the ventricle area and construct the contour point cloud data of the ventricle.

[0137] Step 013: adopting the farthest point sampling algorithm, sampling each of the initial ventricle contour point cloud data respectively according to a preset sampling point threshold, so as to obtain the ventricle segmentation image data corresponding to each magnetic resonance imaging data.

[0138] Specifically, to ensure the consistency of the model input data, each ventricle contour point cloud data is sampled and a fixed number of points (e.g., 4096) are selected. The farthest point sampling method (FPS) is used to simplify the number of input points while ensuring that the selected points can be distributed as evenly as possible on the ventricle surface, so as to better represent the overall shape information of the ventricle.

[0139] In order to further improve the application effectiveness and reliability of the ventricle segmentation image data corresponding to each magnetic resonance imaging data, in a method for training an Alzheimer's disease prediction model provided in an embodiment of the present application, see Figure 6 Step 011 in the Alzheimer's disease prediction model training method specifically includes the following contents:

[0140] Step 0111: Select a preset number of magnetic resonance imaging data from a preset magnetic resonance image library to construct a spectrum library, and receive and store the ventricle segmentation image data corresponding to each of the magnetic resonance imaging data in the spectrum library.

[0141] And, step 0112: performing skull image removal processing on each magnetic resonance imaging data in the magnetic resonance image library to obtain skull-removed image data corresponding to each magnetic resonance imaging data in the magnetic resonance image library.

[0142] Step 0113: Adopting the preset multi-atlas registration and voting strategy, based on the ventricle segmentation image data corresponding to each magnetic resonance imaging data in the atlas library, ventricle segmentation is performed on the skull-removed image data corresponding to each magnetic resonance imaging data that has not currently undergone ventricle segmentation in the magnetic resonance image library, so as to obtain the ventricle segmentation image data corresponding to each magnetic resonance imaging data in the magnetic resonance image library.

[0143] In a specific implementation of step 0111 to step 1113, the specific process of ventricle segmentation is as follows:

[0144] a. Prepare training data: Select about 20 MRI images of individuals of different age groups and different ventricular shapes from the ADNI database, and manually mark the ventricular regions as the atlas library.

[0145] b. Preprocessing: Perform skull removal on all image data in the ADNI dataset. The specific operations are as follows:

[0146] b1. Estimate basic image and brain parameters: by processing the intensity histogram, ignoring the long tails at both ends, find the "lowest" (t2) and "highest" (t98) values ​​of intensity, calculate t, try to roughly distinguish brain matter from background, for all voxels with intensity greater than t, use their intensity (mass) to calculate the center of gravity (COG) of the image, count all voxels with intensity greater than t, assuming the brain is spherical, calculate the average radius of the brain / head.

[0147] b2. Surface model and initialization: Spherical mesh initial model: Starting from an icosahedron, each triangle is iteratively subdivided into four smaller triangles to form a surface that is as spherical as possible. This mesh is initially centered at the centroid and has a radius set to half the estimated brain / head radius.

[0148] b3. Main iteration loop: taking the vector product between adjacent vertices and calculating the local surface normal ; Calculate the average position of all adjacent vertices to obtain the difference vector s and decompose it into a component perpendicular to the local surface and a tangential component; Adjust the vertex positions so that they are evenly distributed and smooth within the surface. For areas of high curvature, use a nonlinear function for stronger smoothing; Calculate the local threshold tl by searching for the minimum and maximum intensity values ​​along the normal pointing to the inside of the brain. Depending on the tl value, decide whether the surface moves inward or outward.

[0149] b4. Increase smoothness (self-intersection detection): Check if the surface model is self-intersecting. If so, rerun the algorithm and apply a higher smoothness constraint in the first 75% of iterations.

[0150] b5. Estimation of the external skull surface: For each voxel located on the brain surface, search along a line perpendicular to the local surface and pointing outward, record the maximum and minimum intensities and their positions, and determine the location of the external skull surface.

[0151] c. Ventricular segmentation: Apply the atlas of 20 selected standard data, use ITK's standard image space definition for the image after skull removal, and use affine transformation (Affine Transformation), including translation, rotation, scaling and shearing, to obtain the proportional change and deformation between the image and the standard image data. The process of finding the best parameters uses a gradient-based optimization algorithm, and uses similarity metrics (mutual information) to evaluate the similarity between images. Transformation models and similarity metrics are used at different scales to improve the accuracy of registration, and voting is performed after the registration is completed. For each coordinate position, if more than half (i.e. ten) of the atlas registration results believe that this place belongs to the ventricular area, it will be marked, otherwise it will not be marked, and the segmentation of the ventricular area will be completed.

[0152] In order to further improve the application effectiveness and reliability of the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample, in a method for training an Alzheimer's disease prediction model provided in an embodiment of the present application, see Figure 5 Step 030 in the Alzheimer's disease prediction model training method specifically includes the following contents:

[0153] Step 031: According to the ventricle volume and the whole brain volume corresponding to each magnetic resonance imaging sample, respectively determine the initial ventricle volume ratio corresponding to each magnetic resonance imaging sample.

[0154] Specifically, the ventricular volume ratio was calculated: To eliminate the influence of individual brain volume differences, the ratio of ventricular volume divided by total brain volume was used as a feature.

[0155] Step 032: Based on the pre-fitted correspondence between the individual age corresponding to the magnetic resonance imaging and the initial ventricular volume ratio, the initial ventricular volume ratio corresponding to each magnetic resonance imaging sample is respectively corrected for the individual age by a covariate, so as to obtain the ventricular volume ratio data corresponding to each magnetic resonance imaging sample after the covariate correction.

[0156] Specifically, covariate correction: To eliminate the effect of age on ventricular volume ratio, the following covariate correction was performed:

[0157] (1) Fitting the relationship between age and ventricular volume ratio: Using nonlinear regression method, fit the functional relationship between age (x) and ventricular volume ratio (y) to obtain the regression equation .

[0158] (2) Calculation of corrected ventricular volume ratio: For each sample, the expected ventricular volume ratio after eliminating the effect of age was calculated based on its age and the fitted regression equation. . Then, the corrected ventricular volume ratio was calculated: , where mean(y) is the average of the ventricular volume ratios of all samples.

[0159] In order to further improve the application effectiveness and reliability of the feature encoder, in the Alzheimer's disease prediction model training method provided in the embodiment of the present application, see Figure 7 and Figure 8 The feature encoder 1 in the Alzheimer's disease prediction model training method specifically includes the following contents:

[0160] The MLP model includes a linear transformation layer and a nonlinear activation function layer connected to each other.

[0161] Among them, the linear transformation layer is used to multiply the vector of each point in the currently input ventricular contour point cloud data by the weight matrix respectively, and add a bias term to output the linear transformation result data corresponding to each point in the currently input ventricular contour point cloud data; the nonlinear activation function is used to perform nonlinear feature learning on each linear transformation result data output by the linear transformation layer of the same layer based on the activation function, so that the MLP model obtains the feature coding of each point in the ventricular contour point cloud data according to the linear transformation layer and nonlinear activation function layer of each layer.

[0162] In an example of feature encoder 1, when processing ventricle contour point cloud data, the first challenge is how to convert the three-dimensional coordinate information of each point into a more abstract and expressive feature representation for subsequent shape analysis and classification. To solve this problem, the present application uses a multi-layer perceptron (MLP model) to perform feature encoding on each point and extract its local geometric information.

[0163] Specifically, for each point , a three-layer MLP is used for feature transformation. Each layer of MLP contains two steps: linear transformation and nonlinear activation function:

[0164] (1) Linear transformation: multiply the input vector by the weight matrix and add the bias term. For example, the linear transformation of the first layer of MLP can be expressed as: ,in is the output vector of the first layer MLP; is the weight matrix; is the 3D coordinate of the point (input vector); is the bias term.

[0165] (2) Nonlinear activation function: Apply a nonlinear activation function to the output of the linear transformation, such as the ReLU function: By introducing the activation function, MLP can learn nonlinear features and better fit complex point cloud data.

[0166] Then apply the above two steps to the three-layer MLP model in sequence, and finally get the feature encoding of the point :

[0167]

[0168] in , and They are the weight matrices of the three-layer MLP, , and is the bias term and σ is the activation function.

[0169] Through the feature encoding of MLP, each point is mapped from the original three-dimensional coordinate space to a higher-dimensional feature space, thereby better capturing the local geometric information of the point cloud. In addition, the weight matrix and bias terms of MLP are learned during the model training process. By minimizing the loss function, the optimal MLP parameters can be found, so that the model can extract the most discriminative features, thereby improving the accuracy of AD classification.

[0170] In order to further improve the application effectiveness and reliability of the hierarchical feature extraction model, in the Alzheimer's disease prediction model training method provided in the embodiment of the present application, see Figure 7 and Fig. 9 The hierarchical feature extraction model 2 in the Alzheimer's disease prediction model training method specifically includes the following contents:

[0171] A PointNet++ network, wherein the PointNet++ network comprises: a first feature extraction layer, a second feature extraction layer, and a third feature extraction layer connected in sequence;

[0172] The first feature extraction layer is used to select a first number of first sampling points in the ventricle contour point cloud data by adopting the farthest point sampling algorithm, and respectively divide the points in the spherical neighborhoods of different radii centered on each of the first sampling points to form point sets corresponding to different radii of each of the first sampling points, respectively extract the feature codes of each point in each of the point sets by adopting the PointNet network to obtain feature data corresponding to each of the points in each of the point sets, and then concatenate and fuse the point sets corresponding to different radii of each of the first sampling points to obtain first fused feature vectors corresponding to each of the first sampling points;

[0173] The second feature extraction layer is used to select a second number of second sampling points from each of the first sampling points in the ventricle contour point cloud data by using the farthest point sampling algorithm, where the second number is smaller than the first number, and respectively divide the points in the spherical neighborhoods of different radii centered on each of the second sampling points to form point sets corresponding to different radii of each of the second sampling points, respectively extract features from the first fused feature vectors of each of the first sampling points in each of the point sets by using the PointNet network to obtain feature data corresponding to each of the first sampling points in each of the point sets, and then concatenate and fuse the point sets corresponding to different radii of each of the second sampling points to obtain second fused feature vectors corresponding to each of the second sampling points;

[0174] The third feature extraction layer is used to perform feature extraction on the second fused feature vectors corresponding to each of the second sampling points, so as to obtain an initial global feature vector corresponding to the ventricle contour point cloud data.

[0175] Specifically, this application optimizes the PointNet++ model and constructs a deep learning classification model with ventricle contour point cloud data as input to achieve the classification of Alzheimer's disease (AD). The input data is the preprocessed ventricle contour point cloud data, with a shape of [B, N, 3], where B is the batch size, N is the number of points (for example, the optimal number of 4096 determined by experiments), and 3 represents the three-dimensional coordinates (x, y, z) of each point.

[0176] In order to capture shape information at different scales, the PointNet++ used in this application uses the first feature extraction layer (PointNetSetAbstractionMsg_1), the second feature extraction layer (PointNetSetAbstractionMsg_2) and the third feature extraction layer (PointNetSetAbstraction).

[0177] The first feature extraction layer (PointNetSetAbstractionMsg_1) first uses the farthest point sampling (FPS) to select 512 points from the input point cloud, and then uses spherical neighborhoods with radii of 0.1, 0.2, and 0.4 to group each sampling point, and combines the points in each area into a point set. For each point set, the PointNet network is used for feature extraction, and the features at different radii are fused through splicing operations. The final output shape is a feature map of [B, 512, 320], where 512 is the number of sampling points and 320 is the feature dimension.

[0178] The second feature extraction layer (PointNetSetAbstractionMsg_2) is similar to the first layer. Its input is the output of the previous layer, the number of sampling points is 128, and the grouping radius is 0.2, 0.4, and 0.8. The output shape is a feature map of [B, 128, 640].

[0179] The third feature extraction layer (PointNetSetAbstraction) does not perform sampling and grouping, but directly extracts features from all points output by the previous layer, and finally outputs an initial global feature vector with a shape of [B, 1, 1024].

[0180] Then, the global feature aggregation layer 3 aggregates and splices the previously extracted features. Specifically, the initial global feature vector output by the third layer is reshaped to a size of [B, 1024] as the input of the subsequent classification layer.

[0181] In order to further improve the application effectiveness and reliability of the classifier, in the Alzheimer's disease prediction model training method provided in the embodiment of the present application, see Figure 7 and Fig.10 The classifier 4 in the Alzheimer's disease prediction model training method specifically includes the following contents:

[0182] The first fully connected layer, the second fully connected layer, the third fully connected layer and the activation function layer are connected in sequence.

[0183] The first fully connected layer is used to perform linear transformation and nonlinear activation on the global feature vector corresponding to the ventricle contour point cloud data output by the global feature aggregation layer, so as to output a first feature vector corresponding to the ventricle contour point cloud data.

[0184] The second fully connected layer is used to perform linear transformation and nonlinear activation on the first eigenvector to output a second eigenvector corresponding to the ventricle contour point cloud data.

[0185] The third fully connected layer is used to perform linear transformation and nonlinear activation on the second eigenvector to output a third eigenvector corresponding to the ventricle contour point cloud data.

[0186] The activation function layer is used to convert the third eigenvector corresponding to the ventricle contour point cloud data into a probability distribution based on the activation function, so as to obtain the first Alzheimer's disease prediction result data corresponding to the ventricle contour point cloud data.

[0187] Specifically, after completing feature encoding and hierarchical feature extraction, the PointNet++ model has converted each ventricle point cloud data into a global feature vector containing rich shape information. Next, these feature vectors are used for AD classification. The classifier is constructed using three fully connected layers and a softmax function. Each fully connected layer performs linear transformation and nonlinear activation on the input features, thereby gradually learning more advanced feature representations. Specifically, the process can be expressed by the following mathematical expression:

[0188] Specifically, the mathematical expression of the classification process is:

[0189]

[0190]

[0191]

[0192] where x is the global eigenvector of the input, and represents the output of the hidden layer in the neural network, , and are the weight matrices of the three fully connected layers, , and They are the bias items of the three fully connected layers. Output represents the final output of the model, which is a probability distribution vector indicating the probability that the input sample belongs to each category. ReLU is an activation function used to introduce nonlinear relationships. The softmax function can convert the final feature vector into a probability distribution, indicating the probability that each sample belongs to the AD and normal categories.

[0193] The final output shape is a probability distribution of [B, 2], where B is the batch size and 2 represents two categories (AD and normal).

[0194] In order to find the optimal model parameters (including the weight matrix of MLP, bias terms, and parameters of the fully connected layer), the model needs to be trained using training data. During the training process, the negative log-likelihood loss function is used to calculate the difference between the predicted probability and the true label. The smaller the value of the loss function, the closer the model's prediction result is to the actual situation. In order to minimize the loss function, the back propagation algorithm is used to update the model parameters. The back propagation algorithm calculates the gradient of each parameter based on the value of the loss function, and then updates the parameters in the opposite direction of the gradient, thereby gradually reducing the value of the loss function. By continuously iterating the training data and updating the model parameters using the back propagation algorithm, a converged PointNet++ model can be obtained, which can effectively learn the ventricle shape features and be used for the classification of AD.

[0195] Correspondingly, in order to more comprehensively utilize ventricular information for Alzheimer's disease (AD) classification and solve the problem that ventricular volume is affected by age and individual differences, this application proposes an Alzheimer's disease ventricle classification model based on deep learning and covariate correction. The model uses an integrated learning method to build a powerful classification model that comprehensively utilizes ventricular shape and volume information after covariate correction. Specifically, two base learners are constructed to extract ventricular features from different angles, respectively. The first deep learning model is used as one of the base learners, and the second deep learning model is used as the second base learner. The first deep learning model is functionally a point cloud-based shape classification model, which focuses on learning the overall shape information of the ventricles. Its input is the ventricular contour point cloud data, and its output is the predicted probability of AD.

[0196] Based on the aforementioned Alzheimer's disease prediction model training method embodiment, the present application also provides an Alzheimer's disease prediction method embodiment, see Fig.11, the Alzheimer's disease prediction method specifically includes the following contents:

[0197] Step 400: Inputting the ventricle contour point cloud data corresponding to the target magnetic resonance imaging data into the first deep learning model in the Alzheimer's disease prediction model, so that the first deep learning model outputs the first Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data.

[0198] The Alzheimer's disease prediction model is pre-trained based on the Alzheimer's disease prediction model training method provided in the aforementioned embodiment.

[0199] And, step 500: input the covariate-corrected ventricular volume ratio data corresponding to the target magnetic resonance imaging data into the second deep learning model in the Alzheimer's disease prediction model, so that the second deep learning model outputs second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data.

[0200] Step 600: Input the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data into the SVM model in the Alzheimer's disease prediction model, so that the SVM model outputs the Alzheimer's disease prediction fusion result data corresponding to the target magnetic resonance imaging data.

[0201] To further illustrate the above-mentioned Alzheimer's disease prediction model training method and the embodiment of the Alzheimer's disease prediction method, only the ventricle shape information is used to accurately classify AD patients and normal people, which has the advantages of fast calculation speed and low image quality requirements. This application also provides an application example of the Alzheimer's disease prediction model training and classification method based on deep learning, see Fig.12 The application examples include the following:

[0202] 1. Constructing the ventricle segmentation model

[0203] 1. Prepare training data: Select about 20 MRI images of individuals of different age groups and different ventricular shapes from the ADNI database, and manually mark the ventricular areas as the atlas library.

[0204] 2. Preprocessing: Perform skull removal on all image data in the ADNI dataset. The specific operations are as follows:

[0205] (1) Estimate basic image and brain parameters: by processing the intensity histogram, ignoring the long tails at both ends, find the “lowest” (t2) and “highest” (t98) values ​​of intensity, calculate t, try to roughly distinguish brain matter from background, for all voxels with intensity greater than t, use their intensity (mass) to calculate the center of gravity (COG) of the image, count all voxels with intensity greater than t, assuming the brain is spherical, calculate the average radius of the brain / head.

[0206] (2) Surface model and initialization: Spherical mesh initial model: Starting from an icosahedron, each triangle is iteratively subdivided into four smaller triangles to form a surface that is as spherical as possible. The initial center of this mesh is located at the center of gravity, and the radius is set to half the estimated brain / head radius.

[0207] (3) Main iterative loop: take the cross product of vectors between adjacent vertices and calculate the local surface normal n̂; calculate the average position of all adjacent vertices to obtain the difference vector s and decompose it into a component perpendicular to the local surface and a tangential component; adjust the vertex positions to make them uniformly distributed and smooth within the surface, and use nonlinear functions for stronger smoothing for areas of high curvature; calculate the local threshold tl by searching for the minimum and maximum intensity values ​​along the normal pointing to the inside of the brain, and decide whether the surface moves inward or outward based on the tl value.

[0208] (4) Increase smoothness (self-intersection detection): Check whether the surface model self-intersects. If so, rerun the algorithm and apply a higher smoothness constraint in the first 75% of iterations.

[0209] (5) Estimation of the external skull surface: For each voxel located on the brain surface, a search is performed along a line perpendicular to the local surface and pointing outward, and the maximum and minimum intensities and their positions are recorded to determine the location of the external skull surface.

[0210] (6) Ventricular segmentation: Apply the 20 selected standard data atlases, use ITK's standard image space definition for the image after skull removal, and use affine transformation (Affine Transformation), including translation, rotation, scaling and shearing, to obtain the proportional change and deformation between the image and the standard image data. The process of finding the best parameters uses a gradient-based optimization algorithm, and uses a similarity metric (mutual information) to evaluate the similarity between images. Transformation models and similarity metrics are used at different scales to improve the accuracy of registration, and voting is performed after the registration is completed. For each coordinate position, if more than half (i.e. ten) of the atlas registration results believe that this place belongs to the ventricular area, it will be marked, otherwise it will not be marked, and the ventricular area will be segmented.

[0211] It is understandable that the ventricle segmentation method based on deep learning segmentation methods (such as U-Net, V-Net, etc.) has achieved remarkable results in the field of medical image segmentation in recent years, and can achieve end-to-end automatic segmentation without complex image registration. However, the training of deep learning models requires a large amount of labeled data, and the generalization ability of the model is limited by the quality and quantity of training data. In addition, the computational complexity of deep learning models is high, and the requirements for hardware equipment are also higher, which limits its application in clinical practice to a certain extent. The segmentation method based on registration and voting adopted in this application does not require time-consuming model training, and has low requirements for image quality, has higher computational efficiency and robustness, and is more suitable for clinical application scenarios.

[0212] 2. Building a ventricle classification model

[0213] In order to more comprehensively utilize ventricle information for Alzheimer's disease (AD) classification and solve the problem that ventricle volume is affected by age and individual differences, this application proposes an Alzheimer's disease ventricle classification model based on deep learning and covariate correction. The model uses an ensemble learning method to build a powerful classification model that comprehensively utilizes ventricle shape and volume information after covariate correction.

[0214] Specifically, two base learners are constructed to extract ventricle features from different perspectives:

[0215] 1. First base learner: shape classification model based on point cloud

[0216] The model focuses on learning the overall shape information of the ventricles. Its input is the ventricle contour point cloud data, and its output is the predicted probability of AD.

[0217] (1) Data preprocessing: The ventricle segmentation model constructed above is used to perform ventricle segmentation on all MRI images to obtain the segmentation results of the ventricle region.

[0218] (2) Feature extraction:

[0219] 21) Obtain ventricle contour points: Apply the breadth-first search algorithm (BFS) to extract the boundary points of the ventricle area and form the contour point cloud data of the ventricle.

[0220] 22) Farthest Point Sampling: To ensure the consistency of the model input data, each ventricle contour point cloud data is sampled and a fixed number of points (for example, 4096) are selected. The farthest point sampling method (FPS) is used to simplify the number of input points while ensuring that the selected points can be distributed as evenly as possible on the ventricle surface, so as to better represent the overall shape information of the ventricle.

[0221] (3) Model training:

[0222] This application is based on the optimization of the PointNet++ model, and builds a deep learning classification model with ventricle contour point cloud data as input to achieve the classification of Alzheimer's disease (AD). The input data is the preprocessed ventricle contour point cloud data, with a shape of [B, N, 3], where B is the batch size, N is the number of points (for example, the optimal number of 4096 determined by experiments), and 3 represents the three-dimensional coordinates (x, y, z) of each point.

[0223] The training process of building a model can be divided into the following stages:

[0224] a. Feature encoder 1:

[0225] When processing ventricle contour point cloud data, the first challenge is how to convert the three-dimensional coordinate information of each point into a more abstract and expressive feature representation for subsequent shape analysis and classification. To solve this problem, this application uses a multi-layer perceptron (MLP) to encode the features of each point and extract its local geometric information.

[0226] Specifically, for each point, a three-layer MLP is used for feature transformation. Each layer of MLP contains two steps: linear transformation and nonlinear activation function:

[0227] a1) Linear transformation: multiply the input vector by the weight matrix and add the bias term.

[0228] a2) Nonlinear activation function: A nonlinear activation function is applied to the output of the linear transformation. By introducing the activation function, the MLP can learn nonlinear features and thus better fit complex point cloud data.

[0229] Then the above two steps are applied to the three-layer MLP in sequence, and finally the feature encoding of the point is obtained.

[0230] Through the feature encoding of MLP, each point is mapped from the original three-dimensional coordinate space to a higher-dimensional feature space, thereby better capturing the local geometric information of the point cloud. In addition, the weight matrix and bias terms of MLP are learned during the model training process. By minimizing the loss function, the optimal MLP parameters can be found, so that the model can extract the most discriminative features, thereby improving the accuracy of AD classification.

[0231] b. Hierarchical feature extraction model 2:

[0232] It is used to perform multi-layer grouping feature extraction on the ventricle contour point cloud data according to the feature coding of each point in the ventricle contour point cloud data, so as to obtain an initial global feature vector corresponding to the ventricle contour point cloud data.

[0233] c. Global feature aggregation layer 3:

[0234] Next, the previously extracted features are aggregated and spliced. Specifically, the initial global feature vector output by the third layer is reshaped to the size of [B, 1024] as the input of the subsequent classification layer.

[0235] d. Classifier 4:

[0236] After completing feature encoding and hierarchical feature extraction, the PointNet++ model has converted each ventricle point cloud data into a global feature vector containing rich shape information. Next, these feature vectors are used for AD classification. The classifier is built using three fully connected layers and a softmax function. Each fully connected layer performs linear transformation and nonlinear activation on the input features, thereby gradually learning more advanced feature representations. The final output shape is a probability distribution of [B, 2], where B is the batch size and 2 represents two categories (AD and normal).

[0237] Among them, the training process of the first base learner includes:

[0238] In order to find the optimal model parameters (including the weight matrix of MLP, bias terms, and parameters of the fully connected layer), the model needs to be trained using training data. During the training process, the negative log-likelihood loss function (F.nll_loss) is used to calculate the difference between the predicted probability and the true label. The smaller the value of the loss function, the closer the model's prediction result is to the actual situation. In order to minimize the loss function, the back propagation algorithm is used to update the model parameters. The back propagation algorithm calculates the gradient of each parameter based on the value of the loss function, and then updates the parameters in the opposite direction of the gradient, thereby gradually reducing the value of the loss function. By continuously iterating the training data and updating the model parameters using the back propagation algorithm, a converged PointNet++ model can be obtained, which can effectively learn the ventricle shape features and be used for the classification of AD.

[0239] 2. Second base learner: volume classification model based on logistic regression

[0240] Considering that the difference in ventricular shape may be affected by covariates, such as age and individual brain volume differences, a second base learner was constructed to focus on learning the ventricular volume information after correcting the covariates. Its input is the corrected ventricular volume ratio, and its output is the predicted probability of AD. The flowchart is shown in Figure 2. Fig.13 shown.

[0241] (1) Feature extraction:

[0242] Calculation of ventricular volume ratio: To eliminate the influence of individual brain volume differences, the ratio of ventricular volume to total brain volume was used as the feature.

[0243] Covariate adjustment: To eliminate the effect of age on ventricular volume ratio, the following covariate adjustment was performed:

[0244] Fitting the relationship between age and ventricular volume ratio: Using nonlinear regression method, fitting the functional relationship between age (x) and ventricular volume ratio (y), the regression equation is obtained .

[0245] Calculate the corrected ventricular volume ratio: For each sample, calculate its expected ventricular volume ratio after eliminating the effect of age based on its age and the fitted regression equation. Then, calculate the corrected ventricular volume ratio:

[0246]

[0247] Where mean(y) is the average of the ventricular volume ratios of all samples.

[0248] (2) When constructing the volume classification model, select logistic regression as the classifier. Logistic regression is a simple and efficient linear classification model, which is particularly suitable for processing binary classification problems with a single numerical feature. In this application, the corrected ventricular volume ratio is used as a feature to train a logistic regression model to predict AD.

[0249] Its core is to convert the output of the linear model into a probability value between 0 and 1, indicating the probability that the sample belongs to the positive class.

[0250] Where x is the input feature, which is the corrected ventricular volume ratio in this application; θ is the model parameter, including the weight coefficient and the bias term; It is the output of the model, indicating the probability that the sample belongs to AD.

[0251] (3) Model training:

[0252] The training goal of the logistic regression model is to find the optimal model parameter θ so that the model can accurately predict the category y of the sample (0 represents normal and 1 represents AD) based on the input feature x.

[0253] In order to measure the accuracy of model prediction, a loss function is used to evaluate the difference between the model prediction value and the true label. The commonly used loss function for logistic regression is the Binary Cross-Entropy Loss function.

[0254] In order to minimize the loss function, the gradient descent method is used to update the model parameters θ. The basic idea of ​​the gradient descent method is to update the parameters in the opposite direction of the gradient of the loss function, thereby gradually reducing the value of the loss function.

[0255] By continuously iterating the training data and updating the model parameters using the gradient descent method, we can eventually find a set of optimal model parameters θ, which enables the model to accurately predict AD based on the corrected ventricular volume ratio.

[0256] It is understandable that traditional three-dimensional shape descriptors such as shape histograms and spherical harmonics can be used to extract the shape features of the ventricles, but their expressive power is limited and it is difficult to capture complex shape changes. In addition, the calculation process of these descriptors often depends on manually set parameters and has a certain degree of subjectivity. The PointNet++ model used in this application can directly process point cloud data, automatically learn complex shape features, and does not require manual feature design, and has stronger expressive power and generalization ability.

[0257] In addition, although traditional machine learning models such as support vector machines (SVM) and random forests can be used for AD classification, their performance is limited by the quality of features. The ensemble learning method adopted in this application, combining the PointNet++ model and the logistic regression model, and using SVM as a meta-learner, can more comprehensively utilize the ventricle shape and volume information, effectively reduce the risk of overfitting, and improve the generalization ability of the model, thereby achieving better classification results.

[0258] 3. Ensemble learning:

[0259] In order to utilize ventricle information more accurately, it is necessary to consider both its shape and volume information. However, during the research process, the input features of the point cloud model for shape research are high-dimensional shape features, and the logistic regression model for volume research is a single numerical feature, which cannot be directly spliced ​​as the input of a single model.

[0260] Therefore, an ensemble learning method is adopted to take the prediction results of the two base learners as new features and input them into a meta-learner (support vector machine SVM model ensemble) for the final classification.

[0261] The specific process is as follows:

[0262] 1) Training base learner:

[0263] Use the training set data to train the point cloud model and logistic regression model respectively.

[0264] 2) Generate meta-learner training set:

[0265] Use the trained base learner to predict the training set data and obtain two prediction probabilities corresponding to each sample. The training set that constitutes the meta-learner , where each sample contains the point cloud model prediction probability ( Represents the AD probability predicted by the point cloud model for the i-th sample) and the predicted probability by the logistic regression model ( represents the AD probability predicted by the logistic regression model for the i-th sample), and the true label y of the sample, The true label of the i-th sample (1 for AD, 0 for normal).

[0266]

[0267] 3) Training the meta-learner:

[0268] Using the training set of the meta-learner Train a support vector machine (SVM) model.

[0269] 4) Find the optimal hyperplane:

[0270] The goal of the SVM algorithm is to find an optimal hyperplane that can separate samples of different categories as much as possible in high-dimensional space. This hyperplane can be expressed by the following equation:

[0271]

[0272] w is the weight vector; b is the bias term; φ() is the kernel function that maps the input data points to a higher-dimensional feature space.

[0273] 5) Maximize Intervals:

[0274] SVM finds the optimal hyperplane by maximizing the distance (i.e., the interval) between the two class samples and the hyperplane. The interval is defined as follows:

[0275]

[0276] Maximizing the interval can be transformed into minimizing ,in is the Euclidean norm of the weight vector w, while satisfying the following constraints (ensuring that all training samples are correctly classified and the distance to the hyperplane is at least 1:

[0277]

[0278] 6) Solve the optimization problem:

[0279] The training process of SVM can be transformed into a convex optimization problem: ;

[0280] At the same time, the following constraints are met:

[0281]

[0282] Get the optimal weight vector w and bias term b.

[0283] Compared with traditional methods, the above ensemble learning method has the following advantages:

[0284] 1) The two base learners describe the morphological features of the ventricles from different perspectives, and the relationship between them may be nonlinear. The SVM algorithm is very suitable for handling such low-dimensional classification problems that may have nonlinear relationships. It maps low-dimensional features to high-dimensional space through kernel functions, making it easier to find a linear hyperplane to separate data of different categories.

[0285] 2) SVM looks for the hyperplane that maximizes the interval between two categories, which makes the model have better generalization ability and can achieve good results even with a small number of training samples.

[0286] 3) Through ensemble learning, we can fully utilize the advantages of different base learners to obtain more accurate and robust classification results than a single model.

[0287] In summary, the deep learning-based Alzheimer's disease prediction model training and classification method provided in the application example of this application has the following characteristics:

[0288] 1. New ideas for diagnosing Alzheimer's disease based on ventricle shape information:

[0289] This application proposes for the first time to use ventricular shape information as the basis for AD diagnosis, breaking through the limitations of traditional methods that rely on whole-brain information or specific biomarkers.

[0290] A large number of studies have shown that ventriculomegaly is an important early imaging feature of AD, but traditional volume measurement methods cannot fully reflect the morphological changes of the ventricles. This application focuses on the overall shape information of the ventricles, which can more comprehensively and sensitively capture the morphological changes of the ventricles of AD patients, thereby improving the accuracy of diagnosis and the possibility of early diagnosis.

[0291] 2. Fast and accurate ventricle segmentation model based on registration and voting:

[0292] This application proposes an innovative ventricle segmentation model, which adopts a strategy based on multi-atlas registration and voting to achieve fast and accurate segmentation of the ventricle area.

[0293] Compared with traditional segmentation methods based on individualized graph construction or deep learning, the method proposed in this application has the following advantages:

[0294] Fast speed: No time-consuming individualized atlas construction or deep learning model training is required, and ventricle segmentation can be completed quickly.

[0295] High accuracy: Through multi-atlas registration and voting strategies, the impact of individual differences and noise can be effectively reduced and the segmentation accuracy can be improved.

[0296] Strong robustness: It has low dependence on image quality and can adapt to MRI data collected by different imaging parameters and equipment.

[0297] 3. Use two base learners and one meta learner to form the model architecture

[0298] The advanced point cloud PointNet++ model is used to extract and classify the extracted ventricle shape information to obtain a base learner. Compared with the traditional voxel-based or manually designed feature methods, the PointNet++ model has the following advantages:

[0299] Able to directly process point cloud data: The geometric information of the original point cloud is retained, avoiding information loss.

[0300] Ability to learn complex shape features: Ability to automatically learn complex shape features without the need for manual feature design, which improves the generalization ability of the model.

[0301] Insensitive to the order of point cloud: It can process disordered point cloud data and is not affected by the order of point cloud, making it more robust.

[0302] The second base learner was obtained by using a volume classification model based on logistic regression, which focused on learning the ventricular volume information after correcting for covariates. Its input was the corrected ventricular volume ratio and its output was the predicted probability of AD.

[0303] By using an ensemble learning method, the prediction results of the two base learners are used as new features and input into a meta-learner (support vector machine SVM model integration) for final classification, so as to more accurately utilize the shape and volume information of the ventricles. The advantages of the above ensemble learning method over the traditional method are as follows:

[0304] The SVM algorithm maps the low-dimensional features of the two base learners to a high-dimensional space through a kernel function, making it easier to find a linear hyperplane to separate data of different categories and giving the model better generalization capabilities.

[0305] Through ensemble learning, the advantages of different base learners are fully utilized to obtain more accurate and robust classification results than a single model.

[0306] In summary, compared with the existing MRI-based Alzheimer's disease (AD) diagnosis method, the above application examples and embodiments of the present application have the following significant advantages:

[0307] 1. Innovative use of ventricle shape information:

[0308] Breaking through the limitations of traditional methods: Existing MRI-based AD diagnostic methods mainly rely on whole-brain information analysis or specific biomarkers, such as hippocampal volume, degree of brain atrophy, etc. However, these methods often ignore the morphological changes of the ventricles, an important brain region. This application proposes for the first time to use ventricle shape information as a basis for AD diagnosis, providing a new perspective and idea for the early diagnosis of AD.

[0309] More comprehensive reflection of ventricular morphological changes: A large number of studies have shown that ventricular enlargement is an important early imaging feature of AD. However, traditional volume measurement methods cannot fully reflect the morphological changes of the ventricles, such as local expansion, asymmetry, etc. This application focuses on the overall shape information of the ventricles, which can capture the morphological changes of the ventricles of AD patients more comprehensively and sensitively, thereby improving the accuracy of diagnosis and the possibility of early diagnosis.

[0310] 2. Fast, accurate and robust ventricle segmentation:

[0311] Efficient registration and voting strategy: This application proposes an innovative ventricle segmentation model, which uses a multi-atlas registration and voting strategy to achieve fast and accurate segmentation of the ventricle region. Compared with deep learning models that require a lot of time to train, this method can quickly complete ventricle segmentation without complex model training, and is more suitable for clinical applications.

[0312] Multi-atlas improves segmentation accuracy: Through multi-atlas registration and voting strategies, the impact of individual differences and noise on segmentation results can be effectively reduced, and segmentation accuracy can be improved. Compared with methods that rely on a single atlas or model, this application can better adapt to MRI data from different individuals and different imaging conditions.

[0313] Low requirements for image quality: The segmentation method proposed in this application has low dependence on image quality, can adapt to MRI data collected by different imaging parameters and equipment, and has stronger robustness.

[0314] 3. Powerful shape feature learning and classification capabilities:

[0315] Direct processing of point cloud data: This application uses the advanced point cloud deep learning model PointNet++ to perform feature learning and classification on the extracted ventricle shape information. Compared with methods that require converting three-dimensional shapes into two-dimensional images or manually designing features, the PointNet++ model can directly process point cloud data, retaining the geometric information of the original point cloud and avoiding information loss.

[0316] Automatically learn complex shape features: The PointNet++ model can automatically learn complex shape features without the need for manual feature design, reducing the interference of human factors and improving the generalization ability of the model, enabling it to better adapt to different data sets and application scenarios.

[0317] Insensitive to the order of point cloud arrangement: The PointNet++ model is insensitive to the order of point cloud arrangement and can process disordered point cloud data. It is not affected by the order of point cloud arrangement and is more robust.

[0318] 4. Covariate correction and ensemble learning improve classification accuracy:

[0319] Effectively remove the influence of irrelevant factors: In the process of extracting volume features, this application takes into account the impact of age and individual brain volume differences on ventricular volume, and uses covariate correction methods to eliminate the influence of these irrelevant factors, so that the model pays more attention to the changes in the ventricular volume itself, thereby improving classification accuracy.

[0320] Integrate shape and volume information: In order to comprehensively utilize the shape and volume information of the ventricles, this application uses an integrated learning method to fuse the prediction results of the shape-based point cloud model and the volume-based logistic regression model. This multi-feature fusion strategy can more comprehensively characterize the morphological changes of the ventricles and further improve the classification accuracy.

[0321] SVM enhances ensemble learning capability: This application selects support vector machine (SVM) as a meta-learner to integrate the prediction results of the two base learners. The SVM algorithm is very suitable for processing low-dimensional classification problems in this application that may have nonlinear relationships. It can effectively capture the complex relationship between the prediction probabilities of the two base learners, thereby improving the final classification accuracy.

[0322] In other words, the deep learning diagnostic model based on ventricle shape proposed in this application effectively overcomes the limitations of existing technologies through innovative ideas, efficient algorithms and powerful models, and provides an accurate, fast and robust new method for the early diagnosis of AD, which has important clinical application value.

[0323] The present application also provides an Alzheimer's disease prediction model training device for executing all or part of the Alzheimer's disease prediction model training method, see Fig.14 The Alzheimer's disease prediction model training device specifically includes the following contents:

[0324] A first base learner application module 10 is used to input the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample into a preset first deep learning model, so that the first deep learning model outputs the first Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample;

[0325] And, a second base learner application module 20 is used to input the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample into a preset second deep learning model, so that the second deep learning model outputs second Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample;

[0326] The meta-learner training module 30 is used to use the first Alzheimer's disease prediction result data, the second Alzheimer's disease prediction result data and the preset Alzheimer's disease classification label corresponding to each of the magnetic resonance imaging samples as the concentrated learning samples corresponding to each of the magnetic resonance imaging samples, and train a preset SVM model based on each of the concentrated learning sample training sets, so that the SVM model is used to output the Alzheimer's disease prediction fusion result data of the magnetic resonance imaging sample according to the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data of the magnetic resonance imaging sample, and together with the first deep learning model and the second deep learning model, form an Alzheimer's disease prediction model.

[0327] The embodiment of the Alzheimer's disease prediction model training device provided in the present application can be specifically used to execute the processing flow of the embodiment of the Alzheimer's disease prediction model training method in the above-mentioned embodiment. Its functions will not be repeated here, and reference can be made to the detailed description of the above-mentioned Alzheimer's disease prediction model training method embodiment.

[0328] The part of the Alzheimer's disease prediction model training device that performs Alzheimer's disease prediction model training can be executed in the server or completed in the client device. The specific selection can be based on the processing power of the client device and the limitations of the user's usage scenario. This application is not limited to this. If all operations are completed in the client device, the client device may also include a processor for specific processing of the Alzheimer's disease prediction model training.

[0329] The client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side, and other implementation scenarios may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, or a server cluster consisting of multiple servers, or a server structure of a distributed device.

[0330] The server and the client device may communicate with each other using any suitable network protocol, including network protocols that have not yet been developed on the date of filing this application. The network protocols may include, for example, TCP / IP, UDP / IP, HTTP, HTTPS, etc. Of course, the network protocols may also include, for example, RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer) protocols used on top of the above protocols.

[0331] In one embodiment of the Alzheimer's disease prediction model training device, see Fig.15 The first base learner application module 10 in the Alzheimer's disease prediction model training device is also specifically connected to the following modules:

[0332] The shape information extraction module 01 is used to extract the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample.

[0333] The first base learner training module 02 is used to use the ventricle contour point cloud data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label as the first training samples corresponding to each of the magnetic resonance imaging samples, and train a preset first deep learning model based on each of the first training samples, so that the first deep learning model is used to output the first Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricle contour point cloud data corresponding to the magnetic resonance imaging sample.

[0334] In one embodiment of the Alzheimer's disease prediction model training device, see Fig.15 The second base learner application module 20 in the Alzheimer's disease prediction model training device is also specifically connected to the following modules:

[0335] The volume information extraction module 03 extracts the ventricular volume ratio data corresponding to each magnetic resonance imaging sample after covariate correction.

[0336] The second base learner training module 04 is used to use the ventricular volume ratio data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label as the second training samples corresponding to each of the magnetic resonance imaging samples, and train a preset second deep learning model based on each of the second training samples, so that the first deep learning model is used to output the second Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricular volume ratio data corresponding to the magnetic resonance imaging sample.

[0337] The present application also provides an Alzheimer's disease prediction device for executing all or part of the Alzheimer's disease prediction method. Fig.16 , the Alzheimer's disease prediction device specifically includes the following contents:

[0338] The first base learner application module 10 is used to input the ventricular contour point cloud data corresponding to the target magnetic resonance imaging data into the first deep learning model in the Alzheimer's disease prediction model, so that the first deep learning model outputs the first Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data, wherein the Alzheimer's disease prediction model is pre-trained based on the Alzheimer's disease prediction model training method.

[0339] And, the second base learner application module 20 is used to input the covariate-corrected ventricular volume ratio data corresponding to the target magnetic resonance imaging data into the second deep learning model in the Alzheimer's disease prediction model, so that the second deep learning model outputs the second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data.

[0340] The meta-learner application module 40 is used to input the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data into the SVM model in the Alzheimer's disease prediction model, so that the SVM model outputs the Alzheimer's disease prediction fusion result data corresponding to the target magnetic resonance imaging data.

[0341] The embodiment of the present application also provides an electronic device, which may include a processor, a memory, a receiver and a transmitter, wherein the processor is used to execute the Alzheimer's disease prediction model training method and / or the Alzheimer's disease prediction method mentioned in the above embodiment, wherein the processor and the memory may be connected via a bus or other means, with the bus connection being taken as an example. The receiver may be connected to the processor and the memory via a wired or wireless manner.

[0342] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0343] As a non-transient computer-readable storage medium, the memory can be used to store non-transient software programs, non-transient computer executable programs and modules, such as the program instructions / modules corresponding to the Alzheimer's disease prediction model training method and / or the Alzheimer's disease prediction method in the embodiments of the present application. The processor executes various functional applications and data processing of the processor by running the non-transient software programs, instructions and modules stored in the memory, that is, implementing the Alzheimer's disease prediction model training method and / or the Alzheimer's disease prediction method in the above method embodiments.

[0344] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0345] The one or more modules are stored in the memory, and when executed by the processor, perform the Alzheimer's disease prediction model training method and / or Alzheimer's disease prediction method in the embodiment.

[0346] In some embodiments of the present application, the user equipment may include a processor, a memory, and a transceiver unit, which may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected through a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0347] As an implementation method, the functions of the receiver and the transmitter in the present application can be considered to be implemented through a transceiver circuit or a dedicated chip for transceiver, and the processor can be considered to be implemented through a dedicated processing chip, a processing circuit or a general chip.

[0348] As another implementation method, it is possible to use a general-purpose computer to implement the server provided in the embodiment of the present application, that is, to store the program code for implementing the functions of the processor, receiver, and transmitter in a memory, and the general-purpose processor implements the functions of the processor, receiver, and transmitter by executing the code in the memory.

[0349] The embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the aforementioned Alzheimer's disease prediction model training method and / or Alzheimer's disease prediction method are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0350] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the Alzheimer's disease prediction model training method and / or implements the Alzheimer's disease prediction method.

[0351] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0352] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0353] In the present application, features described and / or illustrated for one embodiment may be used in the same manner or in a similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0354] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the embodiments of the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for training an Alzheimer's disease prediction model, characterized in that: include: Input the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample into a preset first deep learning model, so that the first deep learning model outputs the first Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample; the first deep learning model includes: a feature encoder, a hierarchical feature extraction model, a global feature aggregation layer and a classifier connected in sequence; the feature encoder is used to extract the feature code of each point in the currently input ventricle contour point cloud data; the hierarchical feature extraction model is used to perform multi-layer grouping feature extraction on the ventricle contour point cloud data according to the feature code of each point in the ventricle contour point cloud data, so as to obtain the initial global feature vector corresponding to the ventricle contour point cloud data; the global feature aggregation layer is used to perform feature aggregation processing on the initial global feature vector to obtain the global feature vector corresponding to the ventricle contour point cloud data; the classifier is used to output the first Alzheimer's disease prediction result data corresponding to the ventricle contour point cloud data according to the global feature vector corresponding to the ventricle contour point cloud data; And, the ventricular volume ratio data corrected by the covariates corresponding to each magnetic resonance imaging sample is input into a preset second deep learning model, so that the second deep learning model outputs the second Alzheimer's disease prediction result data corresponding to each magnetic resonance imaging sample; the second deep learning model includes: a logistic regression model; the logistic regression model is used to output the Alzheimer's disease prediction probability corresponding to the ventricular volume ratio data based on the activation function according to the input ventricular volume ratio data corrected by the covariates, so as to use the Alzheimer's disease prediction probability as the second Alzheimer's disease prediction result data of the magnetic resonance imaging sample corresponding to the ventricular volume ratio data; The first Alzheimer's disease prediction result data, the second Alzheimer's disease prediction result data and the preset Alzheimer's disease classification label corresponding to each of the magnetic resonance imaging samples are respectively used as centralized learning samples corresponding to each of the magnetic resonance imaging samples, and a preset SVM model is trained based on each centralized learning sample training set, so that the SVM model is used to output the Alzheimer's disease prediction fusion result data of the magnetic resonance imaging sample according to the first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data of the magnetic resonance imaging sample, and together with the first deep learning model and the second deep learning model, constitute an Alzheimer's disease prediction model.

2. The Alzheimer's disease prediction model training method according to claim 1, characterized in that: Before inputting the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample into the preset first deep learning model, the method further includes: Extract the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample; The ventricle contour point cloud data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label are respectively used as the first training samples corresponding to each of the magnetic resonance imaging samples, and a preset first deep learning model is trained based on each of the first training samples, so that the first deep learning model is used to output the first Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricle contour point cloud data corresponding to the magnetic resonance imaging sample.

3. The Alzheimer's disease prediction model training method according to claim 1, characterized in that: Before inputting the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample into the preset second deep learning model, the method further includes: Extract the covariate-corrected ventricular volume ratio data corresponding to each MRI sample; The ventricular volume ratio data corresponding to each of the magnetic resonance imaging samples and the preset Alzheimer's disease classification label are respectively used as second training samples corresponding to each of the magnetic resonance imaging samples, and a preset second deep learning model is trained based on each of the second training samples, so that the first deep learning model is used to output the second Alzheimer's disease prediction result data corresponding to the magnetic resonance imaging sample according to the ventricular volume ratio data corresponding to the magnetic resonance imaging sample.

4. The Alzheimer's disease prediction model training method according to claim 2, characterized in that: The step of extracting the ventricle contour point cloud data corresponding to each magnetic resonance imaging sample includes: Acquire ventricle segmentation image data corresponding to each magnetic resonance imaging data in a preset magnetic resonance image library; Based on a breadth-first search algorithm, the boundary points of the ventricle region in each of the ventricle segmentation image data are respectively extracted to form initial ventricle contour point cloud data corresponding to each magnetic resonance imaging data; The farthest point sampling algorithm is adopted to sample each of the initial ventricle contour point cloud data respectively according to a preset sampling point threshold value, so as to obtain the ventricle segmentation image data corresponding to each magnetic resonance imaging data.

5. The Alzheimer's disease prediction model training method according to claim 4, characterized in that: The step of acquiring the ventricle segmentation image data corresponding to each magnetic resonance imaging data in the preset magnetic resonance image library includes: Selecting a preset number of magnetic resonance imaging data from a preset magnetic resonance image library to construct a spectrogram library, and receiving and storing ventricle segmentation image data corresponding to each of the magnetic resonance imaging data in the spectrogram library; and performing skull image removal processing on each magnetic resonance imaging data in the magnetic resonance image library to obtain skull-removed image data corresponding to each magnetic resonance imaging data in the magnetic resonance image library; A preset multi-atlas registration and voting strategy is adopted, and based on the ventricle segmentation image data corresponding to each magnetic resonance imaging data in the atlas library, ventricle segmentation is performed on the skull-removed image data corresponding to each magnetic resonance imaging data that has not currently undergone ventricle segmentation in the magnetic resonance image library, so as to obtain the ventricle segmentation image data corresponding to each magnetic resonance imaging data in the magnetic resonance image library.

6. The Alzheimer's disease prediction model training method according to claim 3, characterized in that: The covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample includes: Determining the initial ventricular volume ratios corresponding to the respective magnetic resonance imaging samples according to the ventricular volumes and the whole brain volumes corresponding to the respective magnetic resonance imaging samples; Based on the pre-fitted correspondence between the individual age corresponding to the magnetic resonance imaging and the initial ventricular volume ratio, the initial ventricular volume ratio corresponding to each magnetic resonance imaging sample is respectively corrected for the individual age to obtain the covariate-corrected ventricular volume ratio data corresponding to each magnetic resonance imaging sample.

7. A method for predicting Alzheimer's disease, characterized in that: include: Inputting the ventricle contour point cloud data corresponding to the target magnetic resonance imaging data into a first deep learning model in an Alzheimer's disease prediction model, so that the first deep learning model outputs first Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data, wherein the Alzheimer's disease prediction model is pre-trained based on the Alzheimer's disease prediction model training method according to any one of claims 1 to 6; And, inputting the covariate-corrected ventricular volume ratio data corresponding to the target magnetic resonance imaging data into the second deep learning model in the Alzheimer's disease prediction model, so that the second deep learning model outputs second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data; The first Alzheimer's disease prediction result data and the second Alzheimer's disease prediction result data corresponding to the target magnetic resonance imaging data are input into the SVM model in the Alzheimer's disease prediction model, so that the SVM model outputs the Alzheimer's disease prediction fusion result data corresponding to the target magnetic resonance imaging data.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the Alzheimer's disease prediction model training method according to any one of claims 1 to 6, and / or implements the Alzheimer's disease prediction method according to claim 7.

Images