An alzheimer's disease diagnosis method based on fNIRS

Abstract

The present application relates to the field of artificial intelligence, and in particular to a kind of Alzheimer's disease diagnosis method based on fNIRS.The specific algorithm steps of a kind of Alzheimer's disease diagnosis method based on fNIRS include collecting fNIRS data, constructing data expansion network, training data expansion network, constructing self-supervised model, training self-supervised model, fine-tuning model, for increasing the specificity of data set.The present application proposes to use the way based on causal convolutional neural network to do data expansion for fNIRS data, to make up for the lack of data quantity, and proposes to use the way of self-supervised learning+model fine-tuning to train fNIRS data, to make up for the lack of label training data.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, G06K9 / 62, and specifically to a method for diagnosing Alzheimer's disease based on fNIRS. Background Technology

[0002] Currently, cognitive dysfunction in the elderly is a common neurological disorder, second only to cerebrovascular diseases in neurological health. Alzheimer's disease (AD) is the most common symptom of cognitive impairment in the elderly. Cognitive function assessment scales are commonly used domestically and internationally as initial screening tools for AD; however, due to significant individual variability and subjective components, their diagnostic effectiveness is relatively poor. With technological advancements, various imaging techniques are widely used in the diagnosis and research of AD and other brain diseases, such as positron emission tomography (PET), magnetic resonance imaging (MRI), and diffusion tensor imaging (DTI). However, these imaging techniques are typically time-consuming and expensive, hindering large-scale rapid screening. Some researchers use techniques such as electroencephalography (EEG), ERP, eye tracking, and electroencephalography (EEG) to explore the cognitive function of AD patients, but their drawbacks include poor spatial resolution and difficulty in locating signal sources. Furthermore, functional magnetic resonance imaging (fMRI), while offering extremely high spatial resolution, has relatively poor temporal resolution and imposes significant restrictions on subject activity.

[0003] Therefore, there is an urgent need to develop a multimodal brain function signal recognition method that can quickly reflect the activation status of the cerebral cortex, while exhibiting good noise resistance during data acquisition, tolerating moderate movement of the subject during data collection, and possessing advantages such as high resolution and portability. Patent application CN202010847474.5 provides a method and system for screening AD feature parameters based on deep neural networks. It filters data that causes significant changes in the model's output value through node stimulation, thus identifying AD feature parameters that can more accurately classify AD symptoms. Patent application CN201810204007.3 discloses a method for diagnosing Alzheimer's disease based on auxiliary domain transfer learning. By clustering medical images and natural images to obtain a subset of the natural image dataset, and using these subsets to train the network, the convolutional neural network can indirectly learn the features of medical images. Further transfer learning is then used to train the medical image data, improving the accuracy of automatic Alzheimer's disease identification. However, it suffers from problems such as insufficient data collection and training data. Summary of the Invention

[0004] To address some problems existing in the prior art, this invention provides an Alzheimer's disease diagnosis method based on fNIRS. The specific algorithm steps include: collecting fNIRS data for subsequent model building and analysis; constructing a data augmentation network based on the collected fNIRS data to avoid overfitting, and transmitting it to the training data augmentation network to build a self-supervised model; training the self-supervised model to match common data features in the fNIRS data, and performing feature fine-tuning model calculations on the trained self-supervised model to increase the specificity of the dataset.

[0005] Preferably, the fNIRS data acquisition uses near-infrared light with wavelengths of 780nm and 850nm, and measures HbO and HbR according to different absorption rates of the cerebral cortex; the distance between the near-infrared light and the detector can be divided into four types, namely 1.5cm, 2.12cm, 3cm, and 3.35cm, and 204 channels are established in the prefrontal cortex of each subject to obtain the subject's test data and transmit it to the computer via serial communication.

[0006] Preferably, the method for training data augmentation network includes: after receiving the test data of the subjects, constructing a two-dimensional matrix of size 204x650 and inputting it into a causal convolutional layer to predict the brain data that the subjects may form based on the obtained test data of the subjects, so that the brain data that the subjects may form is closer to the actual value, and transmitting it to the feature extraction module for multiple extractions.

[0007] Preferably, the feature extraction module comprises multiple parts, which are connected end to end, and each feature extraction module additionally outputs a side feature data with a size of 102x325.

[0008] Preferably, the side branch feature data is input into the ReLU activation function to increase the nonlinearity of the neural network model, and the feature data size is amplified by deconvolution to make it consistent with the size of the input experimental data. The output data is normalized to between positive and negative 1 by the sigmoid activation function.

[0009] Preferably, the training data augmentation network trains the output data, and the specific training data parameters correspond one-to-one with the available training data volume, and the data augmentation network generates two data X1 and X2 that are similar to the input data.

[0010] Preferably, the method for constructing a self-supervised model includes: obtaining f1 and f2 by extracting features from data X1 and X2 through an encoder block network structure, and then obtaining g1 and g2 by performing nonlinear transformations through fully connected and ReLU activation functions respectively. Finally, the goal of the self-supervised model is to maximize the similarity between g1 and g2.

[0011] Preferably, the method for constructing the encoder block network structure includes: the input passes through several dilated convolutional layers and finally outputs one-dimensional feature data; residual connections are used between each layer; a different dilation rate is selected for each layer; in addition, each layer divides the channels of the input matrix into 20 groups and performs convolution on each group; each group outputs a matrix with 5 channels; and the outputs of the 20 convolutional groups are merged together and fed into the next layer. The number of dilated convolutional layers is determined according to the size of the input two-dimensional matrix.

[0012] Preferably, the self-supervised training model calculates the corresponding loss by maximizing the similarity between g1 and g2, so that the loss value is no longer updated.

[0013] Preferably, the fine-tuning model extracts the encoder block module and the parameters of the dense layer of the trained self-supervised model, connects them to a softmax layer to form an independent classification network, and then performs fine-tuning training on the collected original real data to finally obtain a diagnostic model for Alzheimer's disease.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] (1) This invention proposes to augment fNIRS data using a causal convolutional neural network to compensate for the insufficient amount of collected data.

[0016] (2) It is proposed to use self-supervised learning + model fine-tuning to train the fNIRS data in order to make up for the lack of label training data. Attached Figure Description

[0017] Figure 1 Data augmentation network structure algorithm model;

[0018] Figure 2 Flowchart of the featureblock module algorithm in the data augmentation network structure;

[0019] Figure 3 Self-supervised model algorithm structure model;

[0020] Figure 4 The encoder block network structure in a self-supervised model. Detailed Implementation

[0021] This invention provides a method for diagnosing Alzheimer's disease based on fNIRS. The specific algorithm steps include collecting fNIRS data for subsequent model building and analysis; and constructing a data augmentation network based on the collected fNIRS data. The specific algorithm flow is as follows: Figure 1 As shown, the algorithm flowchart for the feature block module is as follows: Figure 2 As shown, this is used to avoid overfitting of the data and is then fed into the training data augmentation network to build a self-supervised model. The specific algorithm structure is as follows: Figure 3 As shown, a self-supervised model is trained to match common data features in the fNIRS data, and feature fine-tuning model calculation is performed on the trained self-supervised model to increase the specificity of the dataset.

[0022] In one embodiment, the fNIRS data acquisition uses near-infrared light with wavelengths of 780nm and 850nm, and measures HbO and HbR based on different absorption rates of the cerebral cortex; the distance between the near-infrared light and the detector can be divided into four types, namely 1.5cm, 2.12cm, 3cm, and 3.35cm, and 204 channels are established in the prefrontal cortex of each subject to acquire the subject's test data and transmit it to the computer via serial communication.

[0023] In one implementation, the method for training a data augmentation network includes: after receiving the test data of the subjects, constructing a two-dimensional matrix of size 204x650 and inputting it into a causal convolutional layer to predict the brain data that the subjects may form based on the obtained test data, so that the brain data that the subjects may form is closer to the actual value, and transmitting it to a feature extraction module for multiple extractions.

[0024] In one embodiment, the feature extraction module comprises multiple parts, which are connected end to end, and each feature extraction module additionally outputs a side feature data with a size of 102x325.

[0025] In one implementation, the side branch feature data is input into the ReLU activation function to increase the nonlinearity of the neural network model, and the feature data size is amplified by deconvolution to make it consistent with the size of the input experimental data. The output data is normalized to the range (-1, 1) using the sigmoid activation function.

[0026] In one implementation, the training data augmentation network trains the output data, and the specific training data parameters correspond one-to-one with the available training data volume. The data augmentation network then generates two data sets, X1 and X2, that are similar to the input data.

[0027] In one implementation, the method for constructing a self-supervised model includes: obtaining f1 and f2 by extracting features from data X1 and X2 through an encoder block network structure, and then obtaining g1 and g2 by performing nonlinear transformations through fully connected and ReLU activation functions respectively. Finally, the goal of the self-supervised model is to maximize the similarity between g1 and g2.

[0028] In one implementation, the method for constructing the encoder block network structure includes: the input passes through several dilated convolutional layers (3x3 convolutional kernels) and finally outputs one-dimensional feature data; residual connections are used between each layer; a different dilation rate is selected for each layer; in addition, each layer divides the channels of the input matrix into 20 groups and performs convolution on each group; each group outputs a matrix with 5 channels; and the outputs of the 20 groups of convolutions are merged together and fed into the next layer. The number of dilated convolutional layers is determined according to the size of the input two-dimensional matrix.

[0029] In one implementation, the self-supervised training model calculates the corresponding loss by maximizing the similarity between g1 and g2, so that the loss value is no longer updated.

[0030] In one implementation, the fine-tuning model extracts the encoder block module and the parameters of the dense layer of the trained self-supervised model, connects them to a softmax layer to form an independent classification network, and then performs fine-tuning training on the collected original real data to finally obtain a diagnostic model for Alzheimer's disease.

[0031] In a preferred embodiment, the experimental procedure involves collecting fNIRS data from 40 subjects according to the diagnostic criteria for Alzheimer's disease (AD) in the Diagnostic and Statistical Manual of Mental Disorders, Revised IV (DSM-IV-R), of which 20 are AD patients and the other 20 are healthy elderly individuals matched for age, sex, and education level with the AD patients.

[0032] All participants underwent laboratory and imaging examinations, including complete blood count, blood biochemistry, liver function, kidney function, comprehensive surgical examination, folic acid, vitamin B12, CT scan, and MRI, to exclude patients with other serious illnesses. Each participant performed three data collection experiments, each consisting of one 1-back task (30 seconds) and one 2-back task (30 seconds). After completing both the 1-back and 2-back tasks, participants were placed in a 30-second resting state.

[0033] Each data acquisition experiment lasted 2 minutes. A portion of the resting-state data was discarded (the last 20 seconds of each of the two resting-state data sets were discarded), leaving 80 seconds of data. The near-infrared device had a sampling rate of 8.138 Hz and 204 channels, so each acquisition experiment yielded one 204x650 two-dimensional matrix. Each participant ultimately obtained three 204x650 two-dimensional matrices, resulting in a total of 120 two-dimensional matrices for 40 participants.

[0034] It is understood that those skilled in the art can make equivalent substitutions or modifications to the technical solution and inventive concept of the present invention, and all such substitutions or modifications should fall within the protection scope of the appended claims.

Claims

1. A method for diagnosing Alzheimer's disease based on fNIRS, characterized in that, fNIRS data is collected for subsequent model building and analysis. Based on the collected fNIRS data, a data augmentation network is constructed to avoid overfitting and is then fed into the training data augmentation network to build a self-supervised model. The self-supervised model is trained to match common data features in the fNIRS data, and feature fine-tuning model calculations are performed on the trained self-supervised model to increase the specificity of the dataset. The fNIRS data acquisition uses near-infrared light with wavelengths of 780 nm and 850 nm, and measures HbO and HbR based on different absorption rates of the cerebral cortex. The distance between the near-infrared light and the detector can be divided into four types: 1.5 cm, 2.12 cm, 3 cm, and 3.35 cm. 204 channels are established in the prefrontal cortex of each subject to acquire the subject's test data and transmit it to the computer via serial communication. The method for training data augmentation networks includes: after receiving the test data of the subjects, constructing a two-dimensional matrix of size 204x650 and inputting it into a causal convolutional layer to predict the brain data that the subjects may form based on the obtained test data, making the brain data that the subjects may form closer to the actual value, and then sending it to the feature extraction module for multiple extractions. The feature extraction module comprises multiple parts, which are interconnected. Each feature extraction module additionally outputs a 102x325 branch feature data. The side branch feature data is input into the ReLU activation function to increase the nonlinearity of the neural network model, and the feature data size is enlarged by deconvolution to make it consistent with the size of the input experimental data. The output data is normalized to between positive and negative 1 by the sigmoid activation function. The training data augmentation network is trained on the output data. The specific training data parameters correspond one-to-one with the available training data volume, and the data augmentation network generates two data X1 and X2 that are similar to the input data. The method for constructing a self-supervised model includes: obtaining data X1 and X2, extracting features through an encoder block network structure to obtain f1 and f2, and then performing nonlinear transformations through fully connected and ReLU activation functions to obtain g1 and g2. Finally, the goal of the self-supervised model is to maximize the similarity between g1 and g2. The method for constructing the encoder block network structure includes: the input passes through several dilated convolutional layers and finally outputs one-dimensional feature data. Residual connections are used between each layer, and a different dilation rate is selected for each layer. In addition, each layer divides the channels of the input matrix into 20 groups and performs convolution on them separately. Each group outputs a matrix with 5 channels. The outputs of the 20 convolutional groups are then merged and fed into the next layer. The number of dilated convolutional layers is determined according to the size of the input two-dimensional matrix. The self-supervised training model calculates the corresponding loss by maximizing the similarity between g1 and g2, so that the loss value is no longer updated. The fine-tuning model extracts the encoder block module and the parameters of the dense layer of the trained self-supervised model, connects them to a softmax layer to form an independent classification network, and then performs fine-tuning training on the collected original real data to finally obtain a diagnostic model for Alzheimer's disease.

Images