Cognitive disorder prediction method and device and electronic equipment
By segmenting and feature extraction of 3D nuclear magnetic images of the human brain, combining feature weighting and fusion modules, accurate prediction of cognitive impairment types is achieved, and the problem of inaccurate prediction in the prior art is solved.
Patent Information
- Application Number
- CN202510290975.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-12
- Publication Date
- 2025-06-10
AI Technical Summary
The prior art cannot accurately predict cognitive impairment types, especially when using 3D structural nuclear magnetic images, it is difficult to fully extract effective features.
The 3D nuclear magnetic images of the human brain were segmented through the AAL-90 standard brain template to obtain 90 3D brain regions images, and the multi-brain region feature extraction module, feature weighting module and feature fusion module were used to perform 3D convolution operations, correlation weighting and feature fusion, and output prediction results.
By fully extracting the effective brain region features in each small-scale image, the accurate prediction ability of cognitive impairment types is improved, and the problem of inaccurate prediction in the prior art is solved.
Smart Images
Figure CN120125912A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of medical artificial intelligence, and particularly to a method, device, and electronic device for predicting cognitive impairment. Background Art
[0002] Alzheimer's disease is a neurodegenerative disease with complex pathology and incurable, and there is a lack of effective treatment means to prevent or reverse the development of the disease. Currently, in clinical detection and treatment, neuropsychological assessment and imaging examinations (such as Magnetic Resonance Imaging, Positron Emission Computed Tomography) are used to evaluate the cognitive function and behavioral performance of patients, and observe the structural and functional changes in the brains of patients to determine the stage of the disease. Among them, SMRI can clearly image the structural changes of the brain, and can effectively judge the structural conditions of regions of interest such as the hippocampus, which is beneficial to judge the current state of the patient. However, there are still challenges in the prediction of AD (Alzheimer's Disease) based on 3D SMRI images.
[0003] Most people use 2D structural nuclear magnetic resonance image slices for prediction. Although this makes it easy for the neural network to process data and the network complexity is relatively low, a lot of feature information will be lost. Another part of people use 3D structural nuclear magnetic resonance images for prediction, but because it is difficult to fully extract the effective features of 3D structural nuclear magnetic resonance images, it is impossible to accurately predict the type of cognitive impairment. For the problem that the existing technology cannot accurately predict the type of cognitive impairment, no effective solution has been proposed yet. Summary of the Invention
[0004] In the present invention, a prediction model and training method for Alzheimer's disease are provided to solve the problem that the existing technology cannot accurately predict the type of cognitive impairment.
[0005] In a first aspect, the present invention provides a method for predicting cognitive impairment, including:
[0006] Obtaining a 3D structural nuclear magnetic resonance image of a patient's human brain;
[0007] Segmenting the 3D nuclear magnetic resonance image of the human brain through the AAL-90 standard brain template to obtain 90 3D brain region images;
[0008] Identifying and predicting the 90 3D brain region images through a cognitive impairment prediction model to obtain a prediction result;
[0009] Among them, the cognitive impairment prediction model includes a multi-brain region feature extraction module, a feature weighting module, and a feature fusion module. The multi-brain region feature extraction module is used to perform 3D convolution operations on 90 3D brain region images respectively through a feature extraction network to obtain 90 brain region features. The feature weighting module is used to perform correlation weighting on the 90 brain region features. The feature fusion module is used to fuse the weighted 90 brain region features and output a prediction result according to the fusion result.
[0010] In some embodiments, segmenting the 3D nuclear magnetic resonance image of the human brain through the AAL-90 standard brain template to obtain 90 3D brain region images includes:
[0011] Align the AAL-90 standard brain template with the 3D nuclear magnetic resonance image of the human brain, and use the masking method to segment out 90 3D brain region images.
[0012] In some embodiments, the feature extraction network is a 3D ResNet network, and the 3D ResNet network performs 3D convolution operations on 90 3D brain region images respectively by sharing network parameters.
[0013] In some embodiments, performing correlation weighting on 90 brain region features includes:
[0014] Construct the 90 brain region features into an input matrix F = [f 1 , f 2 , …, f 90 , where f i represents the i-th brain region feature. Perform a linear transformation on the input matrix to obtain feature matrices P q , P k , P v . Perform a dot product operation on P q and the transpose k of P to obtain a feature correlation matrix. Pass the feature correlation matrix through the SoftMax layer and perform weighted summation with P v . Send the result of the weighted summation to the classifier to obtain a contribution matrix S. Weight the contribution matrix S with the input matrix F to obtain the weighted 90 brain region features.
[0015] In some embodiments, fusing the weighted 90 brain region features and outputting a prediction result according to the fusion result includes:
[0016] Fuse the weighted 90 brain region features and the patient's clinical evaluation score, and output a prediction result according to the fusion result.
[0017] Outputting a prediction result according to the fusion result includes:
[0018] The fusion result is input into a multi-layer perceptron, and the output prediction result is obtained through the multi-layer perceptron. The prediction result includes having Alzheimer's disease, being mildly ill, and being a normal person.
[0019] The loss function of the cognitive impairment prediction model is:
[0020]
[0021] where L is the loss function, y i is the i-th true label, and P i is the probability that the prediction result of the model for the i-th sample is the positive class, and N is the number of samples.
[0022] In a second aspect, a cognitive impairment prediction device is provided in the present invention, including:
[0023] An image acquisition module, configured to acquire a 3D structural MRI image of a patient's human brain;
[0024] An image segmentation module, configured to segment the 3D MRI image of the human brain through an AAL-90 standard brain template to obtain 90 3D brain region images;
[0025] An image prediction module, configured to perform recognition and prediction on the 90 3D brain region images through a cognitive impairment prediction model to obtain a prediction result;
[0026] where the cognitive impairment prediction model includes a multi-brain region feature extraction module, a feature weighting module, and a feature fusion module. The multi-brain region feature extraction module is configured to perform 3D convolution operations on the 90 3D brain region images respectively through a feature extraction network to obtain 90 brain region features. The feature weighting module is configured to perform correlation weighting on the 90 brain region features. The feature fusion module is configured to fuse the weighted 90 brain region features and output a prediction result according to the fusion result.
[0027] In a third aspect, an electronic device is provided in the present invention, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the cognitive impairment prediction method described in the first aspect.
[0028] In a fourth aspect, a computer-readable storage medium is provided in the present invention, on which a computer program is stored. When the computer program is executed by a processor, the steps of the cognitive impairment prediction method described in the first aspect are implemented.
[0029] Compared with the related art, the present invention has the following beneficial effects:
[0030] 1. The present invention provides a method for predicting cognitive impairment. By using the AAL-90 standard brain template to segment the 3D MRI images of a patient's human brain into 90 3D brain region images, the large-scale image is segmented into multiple small-scale images, and then the multi-brain region feature extraction module performs 3D convolution operations on the 90 3D brain region images respectively to extract 90 brain region features, so that the effective brain region features contained in each small-scale image are fully extracted, thereby solving the problem that the prior art cannot accurately predict the type of cognitive impairment.
[0031] 2. The present invention uses a feature weighting module to perform correlation weighting on the 90 brain region features, increasing the weight allocation of important features and reducing the weight allocation of non-critical features, so that the neural network pays attention to important brain regions. Then, the feature fusion module fuses the 90 weighted brain region features, and at the same time incorporates the clinical score as an additional weighted feature, making it a global information to improve the prediction accuracy of the model.
[0032] Details of one or more embodiments of the present application are presented in the following drawings and description to make other features, objectives, and advantages of the present application more concise and understandable. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 is a flowchart of the steps of the method for predicting cognitive impairment in this embodiment;
[0034] Figure 2 is an overall architecture diagram of the cognitive impairment prediction model in this embodiment;
[0035] Figure 3 is a brain region weight diagram of the first sample in this embodiment;
[0036] Figure 4 is a brain region weight diagram of the second sample in this embodiment;
[0037] Figure 5 is a brain region weight diagram of the third sample in this embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0038] To more clearly understand the purpose, technical solution, and advantages of the present application, the present application will be described and explained below with reference to the drawings and embodiments.
[0039] Unless otherwise defined, technical terms or scientific terms involved in this application shall have the general meanings understood by those with ordinary skills in the technical field to which this application belongs. In this application, words such as "a", "an", "one kind", "the", "these" and the like do not indicate a limitation in quantity, and they can be singular or plural. The terms "including", "comprising", "having" and any variants thereof involved in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent in these processes, methods, products or devices. The words such as "connected", "coupled" and the like involved in this application do not limit to physical or mechanical connections, but may include electrical connections, whether directly or indirectly. The "plurality" involved in this application means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" may represent: A exists alone, A and B exist simultaneously, and B exists alone. Usually, the character " / " represents an "or" relationship between the associated objects before and after. The terms "first", "second", "third" and the like involved in this application only distinguish similar objects and do not represent a specific sorting of the objects.
[0040] In this embodiment, a method for predicting cognitive impairment is provided. Figure 1 It is a flowchart of the method for predicting cognitive impairment provided in this embodiment. Refer to Figure 1 As shown, this process includes: step S1, step S2, and step S3.
[0041] Step S1, obtain the 3D structural MRI image of the patient's human brain.
[0042] Exemplarily, the 3D structural MRI image of the human brain can be extracted by a professional device, and then preprocessing operations such as rotation, registration, and normalization are performed on the image. The SPM tool is used to correct the head movement of the image data of the 3D structural MRI image, and operations such as intensity layer normalization and smoothing are taken to reduce the influence brought by individual differences of the subjects.
[0043] Step S2, segment the 3D MRI image of the human brain by the AAL-90 standard brain template to obtain 90 3D brain region images.
[0044] Further, in this embodiment, the AAL-90 standard brain template is aligned with the 3D MRI image of the human brain, and then 90 3D brain region images are segmented by a masking method. Among them, each brain region is marked with a number in the AAL-90 standard brain template, and then each 3D MRI image of the human brain is respectively subjected to a Hadamard product operation with the ALL-90 standard meninges template to segment 90 brain regions, specifically as follows:
[0045] R i =X⊙AAL i ,i = 1,2,…90
[0046] Where X is the 3D MRI image of the human brain, R i is the i-th 3D brain region image, and AAL i is the i-th brain region template in the AAL-90 standard brain template.
[0047] Step S3, the 90 3D brain region images are recognized and predicted by the cognitive impairment prediction model to obtain a prediction result. Referring to Figure 2 shown, among them, the cognitive impairment prediction model includes a multi-brain region feature extraction module, a feature weighting module, and a feature fusion module. The multi-brain region feature extraction module is used to extract 90 brain region features by performing multi-layer convolution operations on the 90 3D brain region images through a feature extraction network respectively. The feature weighting module is used to weight the 90 brain region features, and the feature fusion module is used to fuse the weighted 90 brain region features and output a prediction result according to the fusion result.
[0048] The 3D MRI image of the human brain of the patient is segmented by the AAL-90 standard brain template to obtain 90 3D brain region images, and the large-range image is segmented into multiple small-range images. Then, the multi-brain region feature extraction module performs 3D convolution operations on the 90 3D brain region images respectively to extract 90 brain region features, so that the effective brain region features contained in each small-range image are fully extracted, thus solving the problem that the prior art cannot accurately predict the type of cognitive impairment.
[0049] Furthermore, in this embodiment, the feature extraction network is a 3D ResNet network with shared network parameters. Specifically, the 3D ResNet network is used to perform 3D convolution operations on 90 3D brain region images. Due to the feature of parameter sharing, the parameters of the convolution kernels are shared across the entire input data, and the convolution kernels of different layers can use the same parameters, that is, a 3D convolution layer (with a kernel size of 7 and a stride of 2 for all) is respectively applied to each 3D brain region image and the feature maps are output. Then, a series of operations such as layer normalization, activation, downsampling, and residual connection are performed on these feature maps to gradually process the data to extract useful features, and finally 90 brain region features are obtained. Sharing parameters in the feature extraction network can also comprehensively consider all features within the modality, enabling the model to focus on the discriminative features within the modality, helping the model better understand and classify the data, and contributing to improving the model's perception ability when extracting features.
[0050] Furthermore, in this embodiment, performing correlation-weighting on the 90 brain region features includes constructing the 90 brain region features into an input matrix F = [f 1 , f 2 , …, f 90 , where f i represents the i-th brain region feature. A linear transformation is performed on the input matrix to obtain feature matrices P q , P k , P v . The dot product operation is performed between P q and the transpose of P k to obtain a feature correlation matrix. The feature correlation matrix is weighted-summed with P v through a SoftMax layer, and the result of the weighted sum is sent to a classifier to obtain a contribution matrix S. The contribution matrix S is weighted with the input matrix F to obtain the 90 weighted brain region features. Specifically, the contribution matrix S is:
[0051]
[0052] where d is the dimension of the feature matrix.
[0053] For each input vector in the input matrix F, the corresponding query vector q i , key vector k i , and value vector v i are calculated respectively through three different linear transformation matrices W Q , W k , W V , that is, q i = x i W Q , k i = x i Wk , v i = x i W V . By performing correlation weighting on 90 brain region features, 90 brain region features with contribution weighting are obtained. This can not only focus on the degree of association between multiple brain regions, but also enhance the brain region features corresponding to the AD lesion area. Further, the weighted 90 brain region features are fused with the patient's clinical assessment score, and the prediction result is output according to the result. The clinical assessment score is a preliminary judgment value of a doctor's observation of a patient, and is a value between 0 and 1. It can be used to represent the degree of illness or the probability of illness. Exemplarily, a clinical assessment score of 1 can indicate that the patient has severe Alzheimer's disease, or it can indicate that the patient definitely has Alzheimer's disease; and, a clinical assessment score of 0.5 can indicate that the patient has mild illness, or it can indicate that the patient has a 50% probability of having Alzheimer's disease. The clinical assessment score is used as 1 weight feature. The feature fusion module performs layer normalization on the weighted 90 brain region features and then fuses them with the above 1 weight feature through an attention mechanism. The fused features are again subjected to layer normalization and then input into a feed-forward neural network for processing to obtain a fusion result.
[0054] Further, outputting the prediction result according to the fusion result includes inputting the fusion result into a multi-layer perceptron, and obtaining the output prediction result through the multi-layer perceptron. The prediction result includes having Alzheimer's disease, mild illness, and normal people. Specifically, in this embodiment, the multi-layer perceptron is used for image classification. After inputting the fusion result, the multi-layer perceptron maps it to different classification categories respectively, and then obtains 3 groups of binary classifications: AD (having Alzheimer's disease) and CN (normal people), MCI (mild illness) and CN, AD and MCI.
[0055] The cognitive impairment prediction model needs to be trained before use.
[0056] Further, as shown in Table 1, in this embodiment, the dataset used for training the cognitive impairment prediction model is randomly selected from the ADNI public dataset with 317 cases of AD, 798 cases of MCI, and 730 cases of NC, a total of 1845 3D structural MRI images of the human brain. After performing preprocessing operations such as rotation, registration, and normalization on them, a dataset is formed. Then, 20% of the images are evenly sampled from it according to gender and age distribution to form a test set, and the remaining images are used as a training set (the images in the training set are samples) to ensure that the categories are balanced during training.
[0057] Table 1 Training dataset of the cognitive impairment prediction model
[0058]
[0059] Further, the loss function of the cognitive impairment prediction model is as follows:
[0060]
[0061] where L is the loss function, y i is the i-th true label, and P i is the probability that the prediction result of the model for the i-th sample is the positive class, and N is the number of samples.
[0062] Training the cognitive impairment prediction model with the above loss function can enable the cognitive impairment prediction model to accurately give the prediction result.
[0063] As above, this is the cognitive impairment prediction method provided in this embodiment. Below, the effectiveness of the cognitive impairment prediction method in this embodiment will be illustrated by some experimental data.
[0064] In this embodiment, the experimental settings and evaluation metrics used are first introduced, then ablation experiments are conducted to analyze the effectiveness of the proposed method, and finally, a comparison and analysis are made with existing AD prediction algorithms.
[0065] The cognitive impairment prediction model proposed in this embodiment is programmed under the PyTorch framework and trained on an NVIDIA GeForce RTX 4090. During the training process, the batch size is set to 2, the maximum number of training epochs is set to 50, the Adam optimizer is used, the initial learning rate is set to 0.00001, the weight decay is set to 0.8, and the learning rate is adjusted using exponential decay.
[0066] The evaluation metrics of the cognitive impairment prediction model are as follows: accuracy (Accuracy, ACC), sensitivity (Sensitivity, SEN), specificity (Specificity, SPE), and the area under the receiver operating characteristic curve (Area Under Receiver Operating Characteristic Curve, AUC). Their calculation formulas are as follows:
[0067]
[0068] where TP, FP, TN, and FN represent true positive, false positive, true negative, and false negative, respectively, and x i and y i are SEN and 1 - SPE obtained during the prediction process, respectively.
[0069] In this embodiment, the following experiments for classifying AD, MCI, and NC are designed for different components in the cognitive impairment prediction model. First, only the multi-brain region feature extraction module is used to extract 90 brain region features and classify them. Then, the feature weighting module and the feature fusion module are used to verify that the association weighting of brain region features and the fusion of brain region features are beneficial to improving the classification accuracy. The results are shown in Table 2. After classifying using 90 brain region features, the effect is 0.8872. The classification accuracy using the association weighting of brain region features and attention fusion is 0.9552, an increase of about 7%. This shows that the classification accuracy is greatly improved through the association weighting of brain region features and attention fusion.
[0070] Table 2 Classification accuracy in different cases
[0071]
[0072] In this embodiment, three samples with relatively high brain region feature weighting values are randomly selected. The brain region weights shown by the three samples respectively correspond to Figure 3 , Figure 4 and Figure 5 As shown, after feature extraction of 90 brain regions through the 3D ResNet network respectively, the brain region features are correlated with each other to obtain a contribution matrix. According to the contribution values in the contribution matrix, the brain region features are weighted respectively. The main regions with relatively high correlation degrees (weight values) among the brain region features in the randomly selected samples are concentrated in the olfactory cortex, superior frontal gyrus, medial, superior frontal gyrus, medial orbital, anterior cingulate and paracingulate gyri, and hippocampus and other regions. In addition to the above brain region features, some also focus on regions such as the thalamus and temporal cortex. The regions concerned by these brain region features are all AD-related lesion regions, indicating that the 3D ResNet network designed in this embodiment can accurately identify AD-related lesion regions.
[0073] As shown in Table 3, compared with the global feature extraction algorithms proposed by Zhang et al. (2020) and Lian et al. (2022), the classification accuracy of the cognitive impairment prediction model in this embodiment has increased by approximately 4%. Compared with the feature extraction algorithms based on fixed patches or multi-planes proposed by Qiu et al. (2020) and Liu et al. (2023), the classification accuracy of the cognitive impairment prediction model used in this embodiment has increased by approximately 1%. The pathological process of AD is very complex, which can cause structural changes in multiple brain regions, and the manifestations of structural lesions in different brain regions are inconsistent. In this case, using a single size for feature extraction will make it difficult for the model to comprehensively capture the key feature information of the same lesion area, and it is easy to cause the omission of important information. To solve the above problems, this embodiment proposes an innovative feature extraction method, that is, feature extraction based on 90 brain region images. This method can accurately extract the feature information of each brain region and effectively avoid the loss of relevant lesion information during the feature fusion process. At the same time, considering the overfitting problem that may be caused by insufficient data volume, this embodiment introduces a residual module. The residual module is used for residual connection. By constructing skip connections, the cognitive impairment prediction model can more easily learn the complex features in the data, so that it can maintain good generalization ability even under limited data conditions, effectively improving the classification accuracy. Further, compared with the ROICSE features extracted in the frequency domain and the SVM classifier used by Feng et al. (2022), the machine learning method based on bilateral hippocampal imaging proposed by Gao et al. (2024), and the SMRI slice image prediction network based on regional attention proposed by Zhang et al. (2022), the classification accuracy of the feature weighting module and the feature fusion module proposed in this embodiment has increased by approximately 2%, 1%, and 5% respectively. Because splicing the features of multiple brain regions will increase the spatial complexity of the features and make it impossible to focus on the key areas, resulting in a decline in the network generalization performance. The feature weighting module and the feature fusion module used in this embodiment perform adaptive weight allocation, allocating more weights to the lesion areas that are beneficial to the classification task. At the same time, the weights of these areas are optimized according to the information of other brain regions, so that the cognitive impairment prediction model can comprehensively consider the information of all brain regions.
[0074] Table 3 Results of the classification processes of different methods
[0075]
[0076] In summary, in this embodiment, a method for predicting cognitive impairment is provided. The 3D MRI images of the human brain of a patient are segmented by the AAL-90 standard brain template to obtain 90 3D brain region images, and the large-range image is segmented into multiple small-range images. Then, the multi-brain region feature extraction module performs 3D convolution operations on the 90 3D brain region images respectively to extract 90 brain region features, so that the effective brain region features contained in each small-range image are fully extracted, thus solving the problem that the prior art cannot accurately predict the type of cognitive impairment.
[0077] Furthermore, in this embodiment, the feature weighting module is used to perform correlation weighting on the 90 brain region features, increase the weight allocation of important features, and reduce the weight allocation of non-critical features, so that the neural network pays attention to important brain regions. Then, the feature fusion module fuses the 90 weighted brain region features, and at the same time incorporates the clinical score as an additional weighted feature, making it a global information to improve the prediction accuracy of the model.
[0078] In this embodiment, a device for predicting cognitive impairment is also provided. This device is used to implement the method for predicting cognitive impairment provided in this embodiment, and those that have been described will not be repeated. The terms "module", "unit", "sub-unit", etc. used hereinafter can be a combination of software and / or hardware that can achieve a predetermined function. Although the device described in the following embodiments is preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.
[0079] A device for predicting cognitive impairment provided in this embodiment includes an image acquisition module, an image segmentation module, and an image prediction module.
[0080] The image acquisition module is used to acquire the 3D structural MRI images of the human brain of a patient. The image segmentation module is used to segment the 3D MRI images of the human brain by the AAL-90 standard brain template to obtain 90 3D brain region images. The image prediction module is used to perform recognition and prediction on the 90 3D brain region images through a cognitive impairment prediction model to obtain a prediction result. Among them, the cognitive impairment prediction model includes a multi-brain region feature extraction module, a feature weighting module, and a feature fusion module. The multi-brain region feature extraction module is used to perform 3D convolution operations on the 90 3D brain region images respectively through a feature extraction network to obtain 90 brain region features. The feature weighting module is used to perform correlation weighting on the 90 brain region features. The feature fusion module is used to fuse the 90 weighted brain region features and output a prediction result according to the fusion result.
[0081] It should be noted that the above-mentioned respective modules can be functional modules or program modules, and can be implemented either by software or by hardware. For the modules implemented by hardware, the above-mentioned respective modules can be located in the same processor; or the above-mentioned respective modules can also be located in different processors in any combined form.
[0082] In this embodiment, an electronic device is further provided, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the cognitive impairment prediction method in this embodiment.
[0083] In this embodiment, a computer-readable storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the cognitive impairment prediction method in this embodiment are implemented.
[0084] It should be understood that the specific embodiments described herein are only used to explain this application, rather than to limit it. According to the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of this application.
[0085] Obviously, the drawings are only some examples or embodiments of this application. For those of ordinary skill in the art, this application can also be applied to other similar situations based on these drawings without creative work. Additionally, it can be understood that although the work done during the development process here may be complex and time-consuming, for those of ordinary skill in the art, certain design, manufacturing, or production changes based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient disclosure of this application.
Claims
1. A method for predicting cognitive impairment, characterized in that: include: Obtain 3D structural MRI images of the patient's brain; The human brain 3D MRI images were segmented using the AAL-90 standard brain template to obtain 90 3D brain region images; The cognitive impairment prediction model was used to identify and predict 90 3D brain region images and obtain prediction results; Among them, the cognitive impairment prediction model includes a multi-brain region feature extraction module, a feature weighting module and a feature fusion module. The multi-brain region feature extraction module is used to perform 3D convolution operations on 90 3D brain region images through a feature extraction network to obtain 90 brain region features. The feature weighting module is used to weight the correlation of the 90 brain region features. The feature fusion module is used to fuse the weighted 90 brain region features and output the prediction results based on the fusion results.
2. The method for predicting cognitive impairment according to claim 1, characterized in that: The AAL-90 standard brain template is used to segment the human brain 3D MRI image to obtain 90 3D brain region images including: The AAL-90 standard brain template was aligned with the 3D MRI image of the human brain, and 90 3D brain region images were segmented using the mask method.
3. The cognitive impairment prediction method according to claim 1, characterized in that: The feature extraction network is a 3D ResNet network with shared network parameters.
4. The cognitive impairment prediction method according to claim 1, characterized in that: The correlation weighting of 90 brain region features includes: The 90 brain region features are constructed as an input matrix F = [f1,f2,…,f 90 ], f i Represents the characteristics of the i-th brain region, and the input matrix is linearly transformed to obtain the feature matrix P q , P k , P v , P q With P k Transpose Perform dot product operation to obtain the feature association matrix, and pass the feature association matrix through the SoftMax layer and P v Perform weighted summation, send the result of weighted summation to the classifier to obtain contribution matrix S, weight the contribution matrix S and input matrix F to obtain weighted 90 brain region features.
5. The method for predicting cognitive impairment according to claim 4, characterized in that: The weighted 90 brain region features are fused and the prediction results are output based on the fusion results, including: The weighted 90 brain region features and the patient's clinical assessment scores are fused and the prediction results are output based on the fusion results.
6. The method for predicting cognitive impairment according to claim 5, characterized in that: The prediction results output based on the fusion results include: The fusion results are input into a multi-layer perceptron, and the output prediction results are obtained through the multi-layer perceptron. The prediction results include people with Alzheimer's disease, mildly ill people, and normal people.
7. The method for predicting cognitive impairment according to claim 1, characterized in that: The loss function of the cognitive impairment prediction model is: Among them, L is the loss function, y i is the i-th true label, P i is the probability that the prediction result of the model response to the i-th sample is positive, and N is the number of samples.
8. A cognitive impairment prediction device, characterized in that: include: An image acquisition module is used to obtain a 3D structural nuclear magnetic resonance image of the patient's human brain; An image segmentation module is used to segment the 3D MRI image of the human brain using the AAL-90 standard brain template to obtain 90 3D brain region images; The image prediction module is used to identify and predict 90 3D brain region images through the cognitive impairment prediction model to obtain prediction results; Among them, the cognitive impairment prediction model includes a multi-brain region feature extraction module, a feature weighting module and a feature fusion module. The multi-brain region feature extraction module is used to perform 3D convolution operations on 90 3D brain region images through a feature extraction network to obtain 90 brain region features. The feature weighting module is used to weight the correlation of the 90 brain region features. The feature fusion module is used to fuse the weighted 90 brain region features and output the prediction results based on the fusion results.
9. An electronic device, comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to execute the cognitive impairment prediction method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the cognitive impairment prediction method according to any one of claims 1 to 7 are implemented.