Diagnostic methods and correlation analysis of Alzheimer's disease based on multimodal data

Abstract

The present invention discloses a method for diagnosing Alzheimer's disease based on multimodal data and correlation analysis, which belongs to the technical field of Alzheimer's disease diagnosis. The steps include selecting two modal Alzheimer's disease data sets; preprocessing MPI data; preprocessing SNP data; building a model for processing MPI data and SNP data, and also discloses a correlation analysis of a method for diagnosing Alzheimer's disease based on multimodal data. The present invention adopts a method for diagnosing Alzheimer's disease based on multimodal data and correlation analysis, combining two-dimensional MRI slices, morphological features and single nucleotide polymorphisms, establishing a connection between imaging data and genetic data, and improving the sensitivity and accuracy of diagnosing Alzheimer's disease.

Description

Technical Field

[0001] The present invention relates to the technical field of Alzheimer's disease diagnosis, and in particular to an Alzheimer's disease diagnosis method and correlation analysis based on multimodal data. Background Art

[0002] Alzheimer's disease is a neurodegenerative disease that usually causes the cognitive ability of the elderly to gradually decline, affecting their daily life and social functions. With the continuous aging of the population, the incidence of Alzheimer's disease is also increasing, becoming an important challenge in the field of global public health. Accurate early diagnosis of Alzheimer's disease is crucial for timely intervention and treatment. The Alzheimer's Disease Neuroimaging Initiative (ADNI) is committed to determining the progression of Alzheimer's disease and collects data including MRI and PET images, genetics, cognitive tests and biomarkers. Researchers first used these single-modality data for research. Binette et al. used imaging data such as magnetic resonance imaging (MRI) to clarify that Alzheimer's disease is closely related to lesions in specific brain regions; Takeshi et al. analyzed that patients with Alzheimer's disease showed obvious abnormalities in specific biomarkers. However, Alzheimer's disease is affected by multiple factors, and single-modality data cannot fully reflect its complex pathophysiological process. Therefore, researchers began to explore the pathogenesis of Alzheimer's disease through the study of multimodal data, in order to use more comprehensive data to help diagnose Alzheimer's disease. Houria et al. used a fusion strategy of multimodal images (DTI and MRI) to detect white matter changes and gray matter atrophy in AD patients; Robi et al. used electroencephalogram, MRI and PET to diagnose Alzheimer's disease, and the classification performance of using each individual modality data alone can be improved by 10%-20%. In recent years, with the development of gene sequencing methods, researchers have found that single nucleotide polymorphisms (SNPs) of AD-related genes may seriously damage certain brain functions. Therefore, researchers began to combine imaging data and genetic data to more accurately diagnose and study Alzheimer's disease. Lu et al. proposed an Alzheimer's disease diagnosis model based on anatomical MRI measurements and SNPs; Wang et al. developed a deep learning method IGnet for automatic classification of AD using MRI and SNPs; Amini et al. proved that there is a causal relationship between SNPs and quantitative PET imaging traits.

[0003] For imaging data, magnetic resonance imaging (MRI) is a commonly used imaging technology for diagnosing Alzheimer's disease. Magnetic resonance images are pseudo-three-dimensional images composed of two-dimensional images. Folego et al. performed multi-class diagnosis of Alzheimer's disease through a 22-layer 3D CNN for sMRI. Due to the limited availability of training data and potential data, the implementation of deep learning technology in medical imaging faces challenges. Ahsan Bin et al. used data augmentation and 3D CNN to study the binary and multi-classification of AD. However, in the field of medical imaging, data augmentation may cause data distortion and affect the accuracy of diagnosis and treatment. In addition, the use of deep learning to analyze three-dimensional MRI requires a lot of computing resources and data.

[0004] SNPs in genetic data are a common form of genetic variation in the human genome, which refers to the variation of a single nucleotide in the genome, that is, a base in the DNA sequence is different from the common base in most people. This genetic variation is of great significance in human genetic research because it shows high variability between individuals and is associated with the risk and phenotypic characteristics of many diseases. In the case of Alzheimer's disease, genetic factors play a vital role in its pathogenesis, and gray matter atrophy is an important biomarker for early Alzheimer's disease. Therefore, we further explored the association between SNPs and gray matter volume to find genetic variations and genetic markers associated with the disease, providing a basis for early diagnosis and personalized treatment of the disease. Summary of the invention

[0005] The purpose of the present invention is to provide a diagnostic method for Alzheimer's disease based on multimodal data and correlation analysis, combining two-dimensional MRI slices, morphological characteristics and single nucleotide polymorphisms to establish a connection between imaging data and genetic data, thereby improving the sensitivity and accuracy of diagnosing Alzheimer's disease.

[0006] To achieve the above objectives, the present invention provides an Alzheimer's disease diagnosis method and correlation analysis based on multimodal data, comprising the steps of:

[0007] S1. Two modalities of Alzheimer's disease datasets are selected, namely magnetic resonance imaging (MPI) data and single nucleotide polymorphism (SNP) data. The dataset includes Alzheimer's disease patients (AD), early cognitive impairment patients (MCI) and healthy subjects (NC).

[0008] S2, preprocessing MPI data;

[0009] S3, preprocessing SNP data;

[0010] S4. Build a model to process MPI data and SNP data.

[0011] Preferably, the step S2 of preprocessing the MPI data comprises:

[0012] S21, preprocessing the MPI data using ADNI according to specific image correction steps;

[0013] S22, dividing the MPI data preprocessed in step S21 into two process preprocessing;

[0014] S221, use SPM to preprocess MRI, including AC-PC correction, registration, skull stripping and ROI calculation;

[0015] S222. Slice the three-dimensional MRI along the sagittal, coronal and axial planes, retain the three images at the center, and obtain two-dimensional structural magnetic resonance imaging and morphological features.

[0016] Preferably, the specific image correction steps in step S21 are: Gradwarp, B1non-uniformity and N3.

[0017] Preferably, the SNP data includes SNPs, and the step S3 of preprocessing the SNP data specifically includes:

[0018] S31. Perform quality control on SNPs data;

[0019] S32, fill in missing values ​​of SNPs data;

[0020] S33, Gene annotation of SNPs data;

[0021] S34. Select SNPs data features.

[0022] Preferably, the step S4 specifically includes:

[0023] S41, processing MPI data, dividing the MRI data into two channels, the first channel inputs the morphological features of the preprocessed structural magnetic resonance imaging (sMRI) data into a multi-layer perceptron (MLP) with two hidden layers; the second channel involves inputting the sMRI slices into a convolutional neural network (CNN) for feature extraction;

[0024] S42, processing SNP data, the pre-processed SNP data is input into the MLP with three hidden layers for processing;

[0025] S43, the sMRI data output in step S41 and the SNP data output in step S42 are input into a multi-layer attention feature fusion layer, and the local information and global information of different modalities are captured through the multi-layer attention mechanism and fused;

[0026] S44. Use the SoftMax function for diagnosis and output the class label.

[0027] Preferably, the class labels in step S44 include Alzheimer's disease patients AD, early cognitive impairment patients MCI and healthy subjects NC.

[0028] Correlation analysis of a diagnostic method for Alzheimer's disease based on multimodal data

[0029] This involves bivariate correlation analysis of gray matter volume and SNPs in different brain regions of patients with Alzheimer's disease. When both correlation analyses are significant, specific SNPs are screened out.

[0030] Preferably, the gray matter volumes of different brain regions in patients with Alzheimer's disease are obtained by inter-group difference analysis, which uses voxel-based morphological measurements through statistical parameter mapping to compare the sMRI of the AD group and the NC group to quantify the distribution of gray matter, white matter and cerebrospinal fluid in the brain.

[0031] Preferably, the two correlation analysis methods include Pearson correlation coefficient analysis and Spearman correlation coefficient analysis.

[0032] Therefore, the present invention adopts the above-mentioned Alzheimer's disease diagnosis method based on multimodal data and correlation analysis, which has the following beneficial effects:

[0033] (1) The present invention integrates two-dimensional imaging data, genetic data and morphological features, and applies them to deep learning models, significantly improving the training speed and accuracy of the models; it has the key features of multi-level data integration and machine learning model application, providing new prospects for the early diagnosis of Alzheimer's disease;

[0034] (2) 2D structural MRI and morphological features were obtained using 3D structural MRI and combined with genetic data, including gene variant information, to effectively identify the association between genetic factors and brain imaging features to better understand the interaction between genetic risk and brain structure and function;

[0035] (3) By identifying SNPs associated with gray matter volumes in different brain regions through correlation analysis, we can accurately identify pathogenic genes associated with Alzheimer's disease and better understand how gene variation affects brain structure and function. This provides an opportunity to further explore the relationship between genetics and brain imaging, and offers new ideas for the research and diagnosis of Alzheimer's disease.

[0036] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1A model framework diagram of an Alzheimer's disease diagnosis method and correlation analysis embodiment based on multimodal data of the present invention;

[0038] Figure 2 This is a gray matter comparison diagram of Alzheimer's disease patients and normal subjects in an embodiment of the Alzheimer's disease diagnosis method and correlation analysis based on multimodal data of the present invention. DETAILED DESCRIPTION

[0039] Example

[0040] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0041] Alzheimer's disease diagnosis method and correlation analysis based on multimodal data, including the following steps:

[0042] S1. Two modalities of Alzheimer's disease datasets were selected, namely magnetic resonance imaging (MPI) data and single nucleotide polymorphism (SNP) data. The datasets included 137 Alzheimer's disease patients (AD), 137 early cognitive impairment patients (MCI), and 137 healthy subjects (NC), as shown in Table 1:

[0043] Table 1

[0044]

[0045]

[0046] S2, preprocessing MPI data;

[0047] MPI data were preprocessed using ADNI following specific image correction steps: Gradwarp, B1non-uniformity and N3.

[0048] S22, dividing the MPI data preprocessed in step S21 into two flow preprocessing processes:

[0049] S221. Use SPM to preprocess MRI, including AC-PC correction, registration, skull stripping, and ROI calculation. Specifically, AC-PC correction was first performed using the MATLAB-based SPM12 toolkit. Then all MRIs were registered to the AAL3V1 template using SPM12, the three-dimensional size was unified to 113px×137px×113px, and the spatial resolution was unified to 1mm×1mm×1mm. The skull was stripped using the SPM12-based CAT12 package. Finally, for each ROI of the AAL3V1 template, its gray matter density was calculated and used for the final feature vector for each subject.

[0050] S222. Slice the three-dimensional MRI along the sagittal, coronal and axial planes, retain the three images at the center, and obtain two-dimensional structural magnetic resonance imaging and morphological features, which retain the characteristics and structure of the original three-dimensional image to the greatest extent.

[0051] S3, preprocessing SNP data;

[0052] There are about 3 million SNPs in each subject's original file, but not all SNPs are associated with Alzheimer's disease. Some SNPs are called low confidence, so we took four steps to significantly reduce the number of features: quality control, missing value filling, gene annotation, and feature selection.

[0053] S31. Quality control of SNPs

[0054] The SNPs were quality controlled using Plink software, with standard quality control criteria (minor allele frequency < 0.01, call rate > 95%, HWE < 10 -6 ) to filter the genotype data.

[0055] S32. Fill in missing values ​​of SNPs data

[0056] Beagle 5.0 was used to estimate missing values.

[0057] S33. Gene annotation of SNPs data

[0058] A list of all AD-related genes was downloaded from the AlzGene database, containing a total of 680 genes. These genes were then searched in the UCSC genome browser and 640 genes that matched the NCBI RefSeq annotation were retained. These gene regions were extracted from the Bed format RefSeqAnnotation (genomic reference sequence) (gff file) and used to further filter SNPs. Only genes located in these regions were retained. We then generated a genotype matrix from the file containing 0, 1, and 2 for all subjects, where the numbers represent the number of non-reference alleles.

[0059] S34. Select SNPs data features

[0060] The supervised feature selection method mRMR (minimum redundancy maximum relevance) was applied to further reduce the number of features. After mRMR feature selection, 300 SNPs features were retained.

[0061] S4. Build a model to process MPI data and SNP data, refer to Figure 1 ,include,

[0062] S41. Processing MPI data, the MRI data is divided into two channels. The first channel inputs the morphological features of the preprocessed structural magnetic resonance imaging (sMRI) data into a multi-layer perceptron (MLP) with two hidden layers; the second channel involves inputting the sMRI slices into a convolutional neural network (CNN) for feature extraction.

[0063] S42, processing SNP data, the pre-processed SNP data is input into an MLP with three hidden layers for processing.

[0064] S43, the sMRI data output in step S41 and the SNP data output in step S42 are input into a multi-layer attention feature fusion layer, and the local information and global information of different modalities are captured through the multi-layer attention mechanism and fused.

[0065] The multi-layer attention feature fusion layer includes two attention mechanisms: self-attention and cross-attention. In self-attention, the key, query, and value of each position are from the same sequence, while the key and value of cross-attention are from the same sequence, and the query is from another sequence.

[0066] S44. Use the SoftMax function for diagnosis and output class labels (NC, MCI, AD).

[0067] Correlation analysis of a diagnostic method for Alzheimer's disease based on multimodal data

[0068] This includes bivariate correlation analysis of gray matter volume and SNP in different brain regions of patients with Alzheimer's disease. When both correlation analyses are significant, specific SNPs are screened out, as shown in Table 2

[0069] Table 2 Significantly associated SNPs and corresponding genes

[0070]

[0071] The gray matter volumes of different brain regions in patients with AD were obtained by between-group difference analysis, which compared the sMRI of the AD and NC groups using voxel-based morphometric measurements through statistical parametric mapping to quantify the distribution of gray matter, white matter, and cerebrospinal fluid in the brain.

[0072] Statistical parametric mapping (SPM) refers to the process of constructing and evaluating spatial statistics for testing hypotheses in functional imaging data. It is a widely used technique in neuroimaging, which is mainly used to gain a deeper understanding of structural and functional differences in the brain.

[0073] Voxel-based morphometry (VBM) is a statistical method used to compare brain differences in neuroimaging. By comparing VBM results from different individuals or the same individual at different time points, changes in brain structure can be studied, such as discovering the association between disease and brain structure, or studying changes in brain structure during development.

[0074] The latest version of SPM was used to compare the sMRI of the Alzheimer's disease group and the normal control group using VBM to quantify the distribution of gray matter, white matter and cerebrospinal fluid in the brain. There were statistically significant differences in the performance of gray matter. 137 structural magnetic resonance imaging (sMRI) images were included. The analysis results refer to Figure 2, the difference was still less than 0.001 after FWE correction. The brain regions with significant differences include the superior temporal gyrus, middle temporal gyrus, inferior temporal gyrus, hippocampus, parahippocampal gyrus, fusiform gyrus, etc. These results strongly suggest that these brain regions may play a key role in the pathological process of Alzheimer's disease and are supported by existing medical research. The middle temporal gyrus and inferior temporal gyrus play an important role in language and semantic memory processing and visual intuition processing; the hippocampus plays a key role in memory and spatial navigation. Their atrophy is closely related to a variety of neurodegenerative diseases. The screening results are shown in Table 3. For example, CELF2 has an isoform that is mainly expressed in neurons, and CELF2 variation is associated with APOE e4 homozygotes who have the highest risk of AD; NXPH1 encodes a protein involved in promoting adhesion between dendrites and axons, which is a key factor in synaptic integrity, and the loss of synaptic integrity is a hallmark of AD; NRG1 complexes with ErbB4 can reduce memory formation, and NRG1 is related to cognitive evolution and may be involved in the pathophysiology of AD; CLC6, which is almost exclusively present in neurons, is involved in the regulation of chloride concentration inside and outside the cell, which is the main cause of neurological diseases, and its related manifestations include neurodegeneration, abnormal brain imaging, and speech disorders. These are all major symptoms of Alzheimer's disease; CTNNA3 may affect the nervous system through its role in maintaining glia in the central and peripheral nervous systems.

[0075] Table 3 Significant differences in brain regions and corresponding voxel sizes

[0076]

[0077]

[0078] The two correlation analysis methods include Pearson correlation coefficient analysis and Spearman correlation coefficient analysis.

[0079] The Pearson correlation coefficient, also known as the Pearson product-moment correlation coefficient, is used to measure the correlation (linear correlation) between two variables X and Y. It is defined as follows:

[0080]

[0081] in, is the mean of the sequence X, is the mean of sequence Y, and r is the Pearson correlation coefficient. The value of r is between [-1, 1]. The larger the absolute value of r, the higher the correlation between sequences X and Y; the smaller the absolute value of r, the lower the correlation between sequences X and Y.

[0082] The Spearman correlation coefficient is a nonparametric indicator that measures the dependence of two variables. It uses a monotonic equation to evaluate the correlation between two statistical variables. Assume that the two variables X and Y are n-dimensional vectors, expressed as: X = [X1, X2, ..., X n ],Y=[Y1,Y2,…,Y n ], sort X and Y in descending or ascending order, and take the order of the elements to form x = [x1, x2, ..., x n ],y=[y1,y2,…,y n ], then its definition is as follows:

[0083]

[0084] in, and Represents the mean of n-dimensional vectors x and y, n is the number of samples in the data, and ρ is the correlation coefficient between variables X and Y. The value range of ρ is between -1 and 1, where -1 indicates a complete negative correlation, 1 indicates a complete positive correlation, and 0 indicates no correlation.

[0085] Identification method experiment

[0086] Setting up the experiment

[0087] The model training and testing were completed on the eslaV100 GPU using the TensorFlow2.6.0 environment. The Adam optimization algorithm was used to iteratively update the parameters, with the batch size set to 16, 50 epochs trained, and the learning rate set to 0.001.

[0088] Evaluation indicators

[0089] The present invention is a multi-class classification, and four indicators are used to evaluate the model for each class: Accuracy, Precision, Recall, and F1 score (F1). True positives, true negatives, false positives, and false negatives are defined as TP, TN, FP, and FN, respectively. For example, these formulas are used to calculate the F1 score for each class and then calculate the macro-average F1 score, that is, the arithmetic mean of the F1 scores of each class in all classes, and the calculation formula is as follows:

[0090]

[0091]

[0092]

[0093]

[0094] result

[0095] Single modal data, multimodal data without morphological features, and multimodal data with morphological features are evaluated, and their indicators are shown in Table 4. The results show that multimodal data has a better level in sample classification, and adding morphological features can greatly improve the classification results.

[0096] Table 4 Results of the impact of different data on triple classification

[0097]

[0098] The proposed method is compared with four other state-of-the-art methods. These methods use multimodal data from the ADNI database and process images using 3D CNN. Table 5 lists the comparative experimental results of different classification methods. The results show that our results are superior to the other four advanced methods.

[0099] Table 5 Comparison of classification accuracy

[0100]

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for diagnosing Alzheimer's disease based on multimodal data, characterized in that: Includes steps: S1. Select two modalities of Alzheimer's disease datasets, the two modalities are magnetic resonance imaging MPI data and single nucleotide polymorphism SNP data; S2, preprocessing the MPI data, the steps include: S21, preprocessing the MPI data using ADNI according to specific image correction steps, the specific image correction steps are: Gradwarp, B1 non-uniformity and N3; S22, dividing the MPI data preprocessed in step S21 into two process preprocessing; S221, use SPM to preprocess MRI, including AC-PC correction, registration, skull stripping and ROI calculation; S222, slice the three-dimensional MRI along the sagittal, coronal and axial planes, retain the three images at the center, and obtain two-dimensional structural magnetic resonance imaging and morphological features; S3, preprocessing SNP data; S4. Build a model to process MPI data and SNP data, including: S41, processing MPI data, dividing the MRI data into two channels, the first channel inputs the morphological features of the preprocessed structural magnetic resonance imaging (sMRI) data into a multi-layer perceptron (MLP) with two hidden layers; the second channel involves inputting the sMRI slices into a convolutional neural network (CNN) for feature extraction; S42, processing SNP data, the pre-processed SNP data is input into the MLP with three hidden layers for processing; S43, the sMRI data output in step S41 and the SNP data output in step S42 are input into a multi-layer attention feature fusion layer, and the local information and global information of different modalities are captured through the multi-layer attention mechanism and fused; S44. Use the SoftMax function for diagnosis and output the class label.

2. The method for diagnosing Alzheimer's disease based on multimodal data according to claim 1, characterized in that: The SNP data includes SNPs, and the step S3 preprocesses the SNP data specifically including: S31. Perform quality control on SNPs data; S32, fill in missing values ​​of SNPs data; S33, Gene annotation of SNPs data; S34. Select SNPs data features.

3. The method for diagnosing Alzheimer's disease based on multimodal data according to claim 1, characterized in that: The class labels in step S44 include Alzheimer's disease patients AD, early cognitive impairment patients MCI and healthy subjects NC.

4. A correlation analysis of the multimodal data based Alzheimer's disease diagnosis method as claimed in claims 1-3, characterized in that: Bivariate correlation analysis was performed on the gray matter volume and SNPs in different brain regions of patients with Alzheimer's disease. When both correlation analyses were significant, specific SNPs were screened out.

5. The correlation analysis of the multimodal data based Alzheimer's disease diagnosis method according to claim 4, characterized in that: The gray matter volumes of different brain regions in patients with Alzheimer's disease were obtained by inter-group difference analysis, which used voxel-based morphometric measurement through statistical parameter mapping to compare the sMRI of the AD group and the NC group and quantify the distribution of gray matter, white matter and cerebrospinal fluid in the brain.

6. The correlation analysis of the multimodal data based Alzheimer's disease diagnosis method according to claim 4, characterized in that: The two correlation analysis methods include Pearson correlation coefficient analysis and Spearman correlation coefficient analysis.

Images