An ad determination method and system based on gene difference and image research

Abstract

The application discloses an AD judgment method and system based on gene difference and image research, the method filters out differential expression genes between AD patients and CN by using bioinformatics methods, carries out KEGG and GO analysis by selecting the most significant differential genes, simultaneously carries out double-sample T test on MRI images of AD and CN, filters out the highest expression brain tissue images of the differential expression genes from an HPA database, and determines that the gray matter is directly affected by the disease; the screened three-dimensional gray matter images are processed, and the processed images are taken as input data and input into a 3D-ECA-EfficientNet model; the model respectively obtains an accuracy of 97.51% and 93.36% in CN / AD and CN / MCI classification tasks. The method can more accurately understand early pathological characteristics of Alzheimer's disease and obtain higher classification accuracy.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence-assisted technology, specifically to an AD diagnosis method and system based on genetic differences and image studies. Background Technology

[0002] The rising incidence of dementia globally has placed considerable pressure on public healthcare systems, increasingly becoming a significant public health issue. Given the lack of effective treatments for Alzheimer's disease, timely and accurate diagnosis is crucial for multi-dimensional intervention and slowing disease progression. As a non-invasive and widely used neuroimaging method, brain MRI provides ample anatomical information, aiding in the interpretation and analysis of abnormal structural changes in the brain. Computer-aided MRI diagnosis has gained widespread attention due to its objectivity, timeliness, and cost-effectiveness.

[0003] With the development of deep learning algorithms, convolutional neural network (CNN)-based diagnostic methods integrate image processing, feature extraction, and diagnostic classification. However, CNN-based methods have limitations in modeling long-range dependent features. They can only increase the receptive field by stacking convolutional layers, making it difficult to effectively acquire multi-level and multi-dimensional pathology-related global features, resulting in a waste of computational resources. Furthermore, existing CNN-based diagnostic methods cannot adaptively focus on lesion areas and cannot effectively utilize abnormal structural features, leading to a loss of diagnostic accuracy.

[0004] Therefore, in order to solve the above problems, this paper proposes an AD diagnosis method and system based on genetic differences and image research. Summary of the Invention

[0005] The purpose of this invention is to provide an AD diagnosis method based on gene differences and image studies. This method involves screening for the most significantly different genes based on gene differences, identifying highly expressed brain tissue (gray matter structure) based on the gene expression of the most significantly different genes, processing images of such highly expressed brain tissue, and then using the processed three-dimensional MRI images of the brain gray matter as input data. By employing three-dimensional gray matter images that more directly represent the pathological state, this method aims to more accurately understand the early pathological characteristics of Alzheimer's disease and achieve higher classification accuracy.

[0006] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: an AD diagnosis method based on gene differences and image research, characterized by comprising the following steps:

[0007] S1. Screening of differentially expressed genes in Alzheimer's disease: Differentially expressed genes between Alzheimer's patients (AD) and those with normal cognitive function (CN) were screened in the GEO database using bioinformatics methods. Genes with the most significant differential expression and a P-value ≥ 17.5 were selected. Two-sample t-tests were performed on MRI images of AD and CN, and KEGG and GO analyses were conducted on the most significant differentially expressed genes. The results showed that the selected differentially expressed genes were closely related to Alzheimer's disease. Then, these differentially expressed genes were screened from the HPA database, and the brain tissues with the highest expression of the selected differentially expressed genes were analyzed. It was found that the brain tissues with the highest expression of these differentially expressed genes were directly composed of gray matter or indirectly affected the gray matter structure.

[0008] S2. Through the above KEGG enrichment pathway analysis and GO enrichment analysis of differentially expressed genes, and the comparison of significant differences between MRI images in the AD and CN groups, MRI images collected from the ADNI database were screened out, and the images were preprocessed using the MATLAB additional toolkit SPM12, with the cat12 toolkit added to the toolbox of the toolkit.

[0009] S3. Perform spatial registration on the preprocessed image: register it to the standard MNI152 brain template space for comparison; then use the CAT12 toolkit to process the image to obtain gray and white matter images of the brain;

[0010] S4. Construct a dataset based on the images obtained in S3, and divide the dataset into a training set and a validation set according to a training set:validation set ratio of 8:2; then perform rotation, flipping, and random noise augmentation on the training set.

[0011] S5. Input the images from the dataset into the 3D-ECA-EfficientNet model, which is a combination of a lightweight ECA attention mechanism and EfficientNet.

[0012] S6. Supervised training of the 3D-ECA-EfficientNet model using the training set of the dataset;

[0013] S7: After preprocessing the images of AD patients to be diagnosed in steps S2 and S3, the trained 3D-ECA-EfficientNet model is used for diagnosis and classification.

[0014] Furthermore, in S1, there are 16 genes with the most significant differences, and the gene names are SLC35E1, PXDC1, MIR9-1HG, TUBB, PSMB3, TUBB4B, ATP5F1C, ATP5F1B, PCYOX1L, PSMB2, COPG2IT1, NME1, PSMB7, P4HTM, SARS1, and IDH3G.

[0015] Furthermore, in S1, KEGG pathway enrichment analysis showed that these differentially expressed genes were primarily enriched in pathways associated with neurodegenerative diseases; GO enrichment analysis showed that the differentially expressed genes were functionally closely related to the gray matter region of the brain.

[0016] Furthermore, in S1, a two-sample t-test was performed on the MRI images of AD and CN using the SPM12 toolkit, with P < 0.05 as the significance threshold.

[0017] Furthermore, in S2, the preprocessing includes:

[0018] AC correction is performed on the orientation of the scanned MRI images, which means positioning the MRI head image of the brain to a specific point of the head anatomical structure so that the various MRI images can be compared in the same spatial coordinate system.

[0019] The original image after AC correction is subjected to craniotomy, which is the process of removing the skull and other tissues from the brain image, leaving only the brain tissue. The brain image obtained after craniotomy is obtained.

[0020] Furthermore, the 3D-ECA-EfficientNet model is a three-dimensional modification of the EfficientNet convolutional neural network, which increases its ability to utilize spatial information and introduces a more efficient ECA attention mechanism. Specifically, the SE channel attention module of the MBConv module in the EfficientNet convolutional neural network is replaced with an ECA attention mechanism module; (wherein the original MBConv module consists of a 1×1 ordinary convolution, a k×k depthwise separable convolution, an SE channel attention module, and a Dropout layer).

[0021] Furthermore, the ECA module first performs global average pooling on the input feature map to generate a global description vector for each channel. Then, it replaces the fully connected layer in the SE module with a local one-dimensional convolution operation with an adaptive kernel size to capture the interdependencies between channels. The one-dimensional convolution operation performs convolution along the channel dimension, and the kernel size is adaptively determined according to the number of channels C. Then, the weight vector obtained by the one-dimensional convolution operation is multiplied element-wise with the original feature map to adjust the importance of each channel. The ECA module reduces the number of parameters and computational complexity through local one-dimensional convolution, avoids the unnecessary dimensionality reduction brought by the fully connected layer, and adaptively adjusts the kernel size to adapt to inputs with different numbers of channels. Through these steps, the ECA module implements an efficient and effective channel attention mechanism, which can significantly reduce computational and parameter overhead while maintaining or improving network performance.

[0022] Another objective of this invention is to provide an AD diagnostic system based on gene differences and image studies, characterized in that it includes: a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute an AD judgment method based on gene differences and image studies as described in the above technical solution.

[0023] The beneficial effects of this invention are:

[0024] This invention screens out the most significantly different genes based on gene differences, and then selects brain tissues (gray matter structures) with high expression based on the gene expression of the most significantly different genes. The images of these highly expressed brain tissues are then processed, and the processed three-dimensional MRI images of the brain gray matter are used as input data. By using three-dimensional gray matter images that can more directly show the pathological state, it is possible to more accurately understand the early pathological characteristics of Alzheimer's disease and obtain higher classification accuracy.

[0025] This invention introduces the principle of three-dimensional convolution and combines a lightweight ECA attention mechanism with EfficientNet to build a 3D-ECA-EfficientNet model. This model is used to perform two binary classification tasks, CN / MCI and CN / AD, on gray matter images of Alzheimer's disease subjects. In the CN / AD classification task, the model achieved a classification accuracy of 97.51%, which is an improvement in both accuracy and sensitivity compared to the unmodified model. In the CN / MCI classification task, since MCI is a pathological state between AD and CN, the differences in pathological structures may not be obvious. However, even so, a classification accuracy of 93.36% was still achieved, which is an improvement in both accuracy and specificity compared to the unmodified model. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a volcano diagram of differentially expressed genes screened out by this invention;

[0028] Figure 2 This is a schematic diagram showing the KEGG analysis of the differentially expressed genes screened by this invention;

[0029] Figure 3 This is a GO enrichment analysis diagram of the differentially expressed genes screened by this invention;

[0030] Figure 4 These are the brain regions that show significant differences in MRI images between the AD and CN groups in this invention;

[0031] Figure 5 This is the preprocessed brain-peeling image in Embodiment 1 of the present invention;

[0032] Figure 6 This invention uses the MNI152 standard template and spatially registered brain-peeling images.

[0033] Figure 7 These are gray and white matter images after preprocessing according to the present invention;

[0034] Figure 8 This is a schematic diagram of the ECA-MBConv module structure of the present invention;

[0035] Figure 9 This is a line graph showing the NC / MCI accuracy of this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1

[0038] An AD diagnosis method based on genetic differences and image studies includes the following steps:

[0039] Step 1: Gene Screening

[0040] 1. Using the GEO database, high-throughput sequencing datasets of AD patients were retrieved using the keywords "Alzheimer's disease" and "Homosapiens". The gene chip dataset GSE5281 was selected, and its details are shown in Table 1.

[0041] Table 1. Gene Dataset Information

[0042]

[0043] 2. Using R, gene IDs in the GSE5281 dataset were converted to gene symbols. The original gene expression data in the dataset underwent a log2 transformation to a logarithmic scale, which helps normalize the data, reduce the impact of outliers, and remove samples with duplicate symbols. Then, R was used to analyze differentially expressed genes (DEGs) in AD patients and healthy subjects. The selection criteria were |log2FoldChange|≥1 and P<0.05. A total of 1142 DEGs were found in the GSE5281 dataset (364 upregulated genes and 778 downregulated genes). Figure 1 The volcano plot shown represents the differentially expressed genes. Blue indicates downregulated genes, and red indicates upregulated genes. The most significantly differentially expressed genes were selected based on a −log10P.value ≥ 17.5. Table 2 shows the names, LogFC values, −log10P.values, and corresponding groupings of the most significantly differentially expressed genes (upregulated and downregulated genes).

[0044] Table 2: Some differentially expressed genes between people with AD and healthy individuals

[0045]

[0046] 3. Use RStudio to perform KEGG pathway analysis on the screened differentially expressed genes, and visualize the top ten pathways in KEGG, such as... Figure 2 As shown, various disease pathways of neurodegenerative diseases, including amyotrophic lateral sclerosis (ALS), Alzheimer's disease, Parkinson's disease, Huntington's disease, prion diseases, oxidative phosphorylation, phagosomes, gap junctions, and proteosomal enzymes, are the main enrichment pathways. Among them, the first six enrichment pathways are the most significant and are all the same as those in Alzheimer's disease, indicating neurodegenerative diseases. This verifies that the screened differentially expressed genes are closely related to neurodegenerative diseases such as Alzheimer's disease.

[0047] 4. Based on the above conclusions, differentially expressed genes were screened from the existing HPA database, and the brain tissues with the highest expression of the screened differentially expressed genes were analyzed, as shown in Table 3:

[0048] Table 3. Partial Genetic Differences Between AD and Healthy Individuals

[0049]

[0050] It can be seen that the brain tissues with the highest expression of these differentially expressed genes are all directly composed of gray matter or indirectly affect the structure of gray matter; this further verifies that the screened differentially expressed genes affect the gray matter part of the brain.

[0051] 5. Using RStudio, differentially expressed genes between AD and normal subjects were selected, and Gene Ontology (GO) analysis was performed on these 16 differentially expressed genes. The results of the GO enrichment analysis are as follows: Figure 3 As shown, the results indicate that differentially expressed genes are mainly involved in biological processes such as purine ribonucleoside triphosphate metabolism, proton transmembrane transport, cellular respiration, and aerobic respiration. Significantly enriched cellular components include the proton transport dual-domain ATPase complex, the proteasome core complex, the β-subunit complex, the mitochondrial protein complex, and the mitochondrial inner membrane protein complex. Molecular functions are mainly enriched in cytoskeleton structural components, proton transport ATPase activity, HMCI-like protein binding, ATPase activity, coupled ion transmembrane movement, and ATPase-coupled ion transmembrane transport activity.

[0052] 6. Simultaneously, to further verify the differences between AD and CN on MRI images, a two-sample t-test was performed using the SPM12 toolkit on the MRI images of AD subjects and CN subjects used in this study to identify inter-group differences in MRI images between AD and CN. Using P < 0.05 as a significance threshold, the following activation values ​​were obtained. Figure 4 .from Figure 4 The study showed that the brain structures with significant differences between the AD and CN groups were mainly distributed in the cerebral cortex, hippocampus, and thalamus. This overlaps with the brain regions with high expression of the differentially expressed genes, further verifying that the parts with significant differences between AD and CN are closely related to gray matter.

[0053] Step 2, Image Processing:

[0054] 1. First, we need to perform image preprocessing on the 439 MRI images collected from the ADNI database. Since the MRI images directly obtained from ADNI are initial images, they contain irrelevant information such as the skull and ears. Also, the images scanned by different MRI machines will have different resolutions. In order to reduce the influence of irrelevant factors, reduce differences, and improve classification accuracy, we use the MATLAB-based additional toolkit SPM12. We add the cat12 toolkit to the toolbox of this toolkit to preprocess the images.

[0055] ① Firstly, due to head movements during the scanning process, the MRI images may have inconsistent orientations. In this case, AC correction is required, which involves positioning the MRI head image of the brain to a specific point in the anatomical structure of the head so that the various MRI images can be compared in the same spatial coordinate system.

[0056] ② Next, a cranial dissection operation is performed on the original image after AC correction. This involves removing the skull and other tissues from the brain image, retaining only the brain tissue. The resulting image after cranial dissection is shown below. Figure 5 As shown.

[0057] 2. Perform spatial registration on the preprocessed image: Register it to the standard MNI152 brain template space for comparison, such as... Figure 6 As shown, Figure 6 Image A is the MNI152 brain template, and image B is the brain stripping image after spatial registration. The images are then processed using the CAT12 toolkit to obtain images of the brain's gray and white matter, as shown below. Figure 7 As shown; Figure 7 Image A shows the white matter of the brain, and image B shows the gray matter of the brain, which more clearly shows tissues such as the cerebral cortex, hippocampus, thalamus, and cerebellum. Based on the KEGG pathway enrichment analysis and GO enrichment analysis of differentially expressed genes, as well as the results of a two-sample t-test on the MRI images of AD and CN subjects, the quality images were obtained.

[0058] Step 3: Model Building

[0059] In the EfficientNet network architecture, the ECA attention mechanism is selected to replace the SE module in the core module MBConv of the EfficientNet network architecture, resulting in the ECA-MBConv module (i.e., the 3D-ECA-EfficientNet model).

[0060] The structure of the ECA-MBConv module is as follows: Figure 8As shown, this module consists of a 1×1×1 ordinary convolution, a k×k×k depthwise separable convolution, an ECA channel attention module, and a Dropout layer. The first 1×1×1 convolution increases dimensionality, while the last 1×1×1 convolution reduces dimensionality. DepwiseConv effectively reduces the number of model parameters. The ECA attention mechanism reduces the number of parameters and computational complexity through local one-dimensional convolutions, avoiding unnecessary dimensionality reduction from fully connected layers, and reducing computational and parameter overhead while maintaining or improving network performance. When the shape of the input feature map is the same as the shape of the output feature map, a side shortcut operation is performed.

[0061] Specifically, the ECA module first performs global average pooling on the input feature map to generate a global description vector for each channel. Then, it replaces the fully connected layer in the SE module with a local one-dimensional convolution operation with an adaptive kernel size to capture the inter-channel dependencies. The one-dimensional convolution operation performs convolution along the channel dimension, and the kernel size is adaptively determined based on the number of channels C. The weight vector obtained from the one-dimensional convolution operation is then element-wise multiplied with the original feature map to adjust the importance of each channel. The ECA module reduces the number of parameters and computational complexity through local one-dimensional convolution, avoiding unnecessary dimensionality reduction from fully connected layers. Simultaneously, the adaptive adjustment of the kernel size can adapt to inputs with different numbers of channels. Through these steps, the ECA module implements an efficient and effective channel attention mechanism, significantly reducing computational and parameter overhead while maintaining or improving network performance.

[0062] Step 4: Model Training

[0063] Based on the images obtained from the above steps, a dataset is constructed, and the dataset is divided into a training set and a validation set according to a ratio of 8:2. The training set is then subjected to rotation, flipping, and random noise data augmentation.

[0064] The 3D-ECA-EfficientNet model was trained under supervision using the training set of the dataset.

[0065] Step 5: Model Usage

[0066] After preprocessing the images of AD patients to be diagnosed in steps S2 and S3, the trained 3D-ECA-EfficientNet model is used for diagnostic classification; the details are as follows:

[0067] The size changes of the processed input (113×113×137) 3D grayscale MRI image after passing through the above convolutional model are shown in the following table. In the first stage, after processing with a (3×3×3) 3D convolutional kernel, the input data resolution becomes (57×57×69), and the number of channels is 32. After the second stage, the number of channels is halved. After the third stage, the resolution is (29×29×35), and the number of channels is 24. The subsequent 3-8 stages repeat the above operations. The specific changes in image resolution and number of channels are shown in the table. In Stage 9, global average pooling is applied to reduce the spatial dimension of the feature map to 1x1 while preserving the channel dimension. The output of the global pooling operation is converted into the final prediction result through a fully connected layer, as shown in Table 4.

[0068] Table 43 D-ECA-EfficientNet Network Framework

[0069]

[0070] Example 2

[0071] The practical application of the AD diagnosis method based on gene differences and image research includes the following steps:

[0072] An AD judgment method based on gene differences and image research is implemented in the processor and memory. Specifically, the hardware selected for the experiment is shown in Table 5. The deep learning framework is based on the Python PyTorch environment, and training is performed using GPU.

[0073] Table 5 Hardware Configuration

[0074]

[0075] Experimental parameters: In the model, the batch size was set to 2, the initial learning rate was set to 0.0001, the optimizer used was the stochastic gradient descent optimizer, and to avoid too many iterations that would prevent the learning of new features later, the number of iterations was set to 50, and the loss function was chosen as the cross-entropy loss function.

[0076] Experimental content: Two binary classification tasks, CN / AD and CN / MCI, were performed on three categories: normal cognitive function (CN), mild cognitive impairment (MCI), and Alzheimer's disease (AD). The segmented grayscale images were expanded. 157, 170, and 162 processed grayscale images were selected for CN, MCI, and AD, respectively. In both classification tasks, the training set and validation set were divided in a ratio of 8:2. That is, in the CN / AD classification task, the training set had 255 grayscale images and the validation set had 64 images. In the CN / MCI classification task, the training set had 262 grayscale images and the validation set had 65 images.

[0077] Experimental Results: The accuracy of CN / AD classification using the 3D-ECA-EfficientNet model varies between the training and validation sets as follows: Figure 9 As shown, the accuracy improved significantly with increasing iterations. On the training set, fluctuations only occurred in the first 25 iterations. In the first 10 iterations, the accuracy fluctuated dramatically between 0.8 and 0.6. After the 11th iteration, the accuracy began to fluctuate between 0.9 and 1, and after the 25th iteration, the accuracy converged to 1. On the validation set, the fluctuations were smaller in the first 10 iterations, and the convergence was faster.

[0078] To evaluate the performance of the improved 3D-ECA-EfficientNet model used, Table 6 compares the accuracy achieved with several other algorithms in the Alzheimer's classification task. It can be seen that the model used in this chapter achieves a CN / AD classification accuracy of 97.51% and a CN / MCI classification accuracy of 93.36%, which is significantly better than other models in terms of accuracy.

[0079] Table 6. Accuracy comparison of 3D-ECA-EfficientNet with other methods

[0080]

Claims

1. A method for diagnosing Alzheimer's disease (AD) based on genetic differences and image analysis, characterized in that, Includes the following steps: S1. Screening of differentially expressed genes in Alzheimer's disease: Differentially expressed genes between Alzheimer's patients (AD) and those with normal cognitive function (CN) were screened in the GEO database using bioinformatics methods. Genes with the most significant differential expression and a P-value ≥ 17.5 were selected. Two-sample t-tests were performed on MRI images of AD and CN, and KEGG and GO analyses were conducted on the most significant differentially expressed genes. The results showed that the selected differentially expressed genes were closely related to Alzheimer's disease. Then, these differentially expressed genes were screened from the HPA database, and the brain tissues with the highest expression of the selected differentially expressed genes were analyzed. It was found that the brain tissues with the highest expression of these differentially expressed genes were directly composed of gray matter or indirectly affected the gray matter structure. S2. Through the above KEGG enrichment pathway analysis and GO enrichment analysis of differentially expressed genes, and the comparison of significant differences between MRI images in the AD and CN groups, MRI images collected from the ADNI database were screened out, and the images were preprocessed using the MATLAB additional toolkit SPM12, with the cat12 toolkit added to the toolbox of the toolkit. S3. Perform spatial registration on the preprocessed image: register it to the standard MNI152 brain template space for comparison; then use the CAT12 toolkit to process the image to obtain gray and white matter images of the brain; S4. Construct a dataset based on the images obtained in S3, and divide the dataset into a training set and a validation set according to a training set:validation set ratio of 8:2; then perform rotation, flipping, and random noise augmentation on the training set. S5. Input the images from the dataset into the 3D-ECA-EfficientNet model, which is a combination of a lightweight ECA attention mechanism and EfficientNet. S6. Supervised training of the 3D-ECA-EfficientNet model using the training set of the dataset; S7: After preprocessing the images of AD patients to be diagnosed in steps S2 and S3, the trained 3D-ECA-EfficientNet model is used for diagnosis and classification.

2. The AD diagnosis method based on gene differences and image research according to claim 1, characterized in that, In S1, there are 16 genes with the most significant differences, and their names are SLC35E1, PXDC1, MIR9-1HG, TUBB, PSMB3, TUBB4B, ATP5F1C, ATP5F1B, PCYOX1L, PSMB2, COPG2IT1, NME1, PSMB7, P4HTM, SARS1, and IDH3G.

3. The AD diagnosis method based on gene differences and image research according to claim 1, characterized in that, In S1, KEGG pathway enrichment analysis showed that these differentially expressed genes were mainly enriched in pathways associated with neurodegenerative diseases; GO enrichment analysis showed that the differentially expressed genes were functionally closely related to the gray matter of the brain.

4. The AD diagnosis method based on gene differences and image research according to claim 1, characterized in that, In S1, a two-sample t-test was performed on the MRI images of AD and CN. Specifically, the SPM12 toolkit was used to perform a two-sample t-test on the MRI images of AD subjects and CN subjects, with P < 0.05 as the significance threshold.

5. The AD diagnosis method based on gene differences and image research according to claim 1, characterized in that, In S2, the preprocessing includes: AC correction is performed on the orientation of the scanned MRI images, which means positioning the MRI head image of the brain to a specific point of the head anatomical structure so that the various MRI images can be compared in the same spatial coordinate system. The original image after AC correction is subjected to craniotomy, which is the process of removing the skull and other tissues from the brain image, leaving only the brain tissue. The brain image obtained after craniotomy is obtained.

6. The AD diagnosis method based on gene differences and image research according to claim 1, characterized in that, The 3D-ECA-EfficientNet model is a three-dimensional modification of the EfficientNet convolutional neural network, which increases its ability to utilize spatial information and introduces a more efficient ECA attention mechanism. Specifically, the SE channel attention module of the MBConv module in the EfficientNet convolutional neural network is replaced with the ECA attention mechanism module.

7. The AD diagnosis method based on gene differences and image research according to claim 6, characterized in that, The ECA module first performs global average pooling on the input feature map to generate a global description vector for each channel. Then, it replaces the fully connected layer in the SE module with a local one-dimensional convolution operation with an adaptive kernel size to capture the inter-channel dependencies. The one-dimensional convolution operation performs convolution along the channel dimension, with the kernel size adaptively determined based on the number of channels C. The weight vector obtained from the one-dimensional convolution operation is then element-wise multiplied with the original feature map to adjust the importance of each channel. The ECA module reduces the number of parameters and computational complexity through local one-dimensional convolution, avoiding unnecessary dimensionality reduction from fully connected layers. Simultaneously, the adaptive adjustment of the kernel size accommodates inputs with different numbers of channels. Through these steps, the ECA module implements an efficient and effective channel attention mechanism, significantly reducing computational and parameter overhead while maintaining or improving network performance.

8. An AD diagnostic system based on gene differences and image studies, characterized in that, include: The processor and memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the AD judgment method based on genetic differences and image research as described in any one of claims 1-6.

Images