Alzheimer's disease classification method based on channel residual attention

Abstract

The application discloses an Alzheimer's disease classification method based on channel residual attention. The method classifies Alzheimer's disease by fusing the characteristics of residual modules and channel attention mechanisms. The network is composed of an input module, a channel separation residual module, a channel attention module and an output module. Firstly, aiming at the problems existing in the current research data set, a scientific division strategy of the data set is formulated under the guidance of a clinician; then the processed data are sent into the channel separation residual module to extract the shallow and deep features of the network and prevent repeated gradient information; then the extracted features are sent into the channel attention module to adjust the weight between channels and obtain more accurate classification features; finally, the feature matrix is sent into a linear classification layer to output a classification result. The method has broad application prospects in the field of medical images and Alzheimer's disease classification.

Description

TECHNICAL FIELD

[0001] The application designs an Alzheimer's disease classification method based on channel residual attention, and relates to the technical field of computer vision image classification. BACKGROUND

[0002] Alzheimer's disease (AD) is a common central nervous system degenerative disease, and is the most common cause of dementia. It is characterized by progressive deterioration of cognitive and memory abilities, and seriously affects the daily life of patients. According to relevant research, by 2050, about 640 million people will suffer from Alzheimer's disease. With the increasing severity of global aging, Alzheimer's disease with a long course will gradually become an urgent problem in the health field. Magnetic resonance imaging (MRI) is crucial for early diagnosis, differential diagnosis and evaluation of disease progression of Alzheimer's disease because it can accurately and intuitively show brain substructures, atrophy and abnormal lesions, has high image resolution and low cost. In recent years, convolutional neural networks (CNN) have shown great potential in computer-aided cognitive disease diagnosis because of their strong automatic feature extraction capabilities, which can extract a large number of detailed features and lesion locations related to diseases from complex data.

[0003] In the era of deep learning, data is as important as algorithms, and deep learning models require a large amount of data for training. Most current research directly extracts a portion of data from public Alzheimer's disease databases for research, but the lack of uniformity in data collection standards and division standards and different preprocessing methods leads to a lack of credibility in comparing different algorithms. By randomly shuffling the samples and dividing the training set and test set, the brain image data of the same person at different times may appear in the training set and test set at the same time, resulting in a false high classification accuracy in the experiment and a sudden drop in accuracy in practice. In addition, most current MRI image-based research simply divides the data set by MRI image results without considering other parameters such as scales, ages, and genders. There may be extreme cases in the data set or the data distribution between the training set and the test set may be too different, which may prevent the data set from showing the general disease process and characteristics of Alzheimer's disease, ultimately affecting the classification results and the generalization of the model. Therefore, a standard classification data set without data leakage and data balance needs to be established before the data is input into the CNN model. SUMMARY

[0004] The application firstly aims at the problems in the construction of the data set in the current research, and formulates a scientific division strategy for the data set under the guidance of the clinician; then, aiming at the characteristics of the brain MRI of AD, a channel residual attention-based Alzheimer's disease classification network is built based on the 3D-ResNet as the basic network, combined with the characteristics of the channel attention mechanism and the residual module, which mainly consists of an input layer, a feature extraction module, a channel attention module and an output module. The input module adopts 3*3*3 convolution to extract shallow features according to the characteristics that the MRI medical image resolution is not high and more attention should be paid to the detailed features, so as to avoid losing effective information; the feature extraction module is composed of four consecutive channel separation residual sub-modules, the channel separation residual sub-module mainly performs feature extraction, fully combines low-dimensional and high-dimensional features and reduces repeated gradient information, the channel attention module mainly adjusts the degree of dependence between channels by allocating the weight between channels, so as to make the model pay attention to the channel features with the largest amount of information and suppress some unimportant features, thereby helping the network to extract more accurate features. The output module first adopts a fully connected layer for feature dimension reduction, and then uses Softmax to obtain the probability of prediction classification, and selects the maximum probability as the result output.

[0005] The application achieves the above-mentioned purposes through the following technical solutions:

[0006] Step one: based on the public ADNI data set, select the subjects with MRI images, age, gender and cognitive test scale data to construct a data set, and scientifically and evenly divide the data set;

[0007] Step two: pre-process the MRI image data set constructed in step one to obtain the MRI images after registration, sampling and normalization;

[0008] Step three: send the pre-processed MRI images into a 3*3*3 convolution layer to extract shallow features, and then send them into four consecutive stacked channel separation residual modules to extract and combine shallow and deep features;

[0009] Step four: send the extracted features into a channel attention module to adjust the weight between channels and obtain more accurate classification features;

[0010] Step five: take the dimension-reduced feature matrix as the input information of the linear classification layer of the whole network.

[0011] As the scientific and balanced division of the data set described in step one, the following is explained:

[0012] The experimental data of the application is selected from the MRI images of 1.5T T1 weighting of ADNI 1, ADNI 2 and ADNIGO programs, with a total number of 8600. According to the disease process of Alzheimer's disease, it is divided into three categories: cognitive impairment (Control Normal, CN), mild cognitive impairment (Mild Cognitive Impairment, MCI) and (Alzheimer's Disease, AD). The division strategy is mainly divided according to the individual ID, avoiding the same person's data at different times appearing in the training set and test set at the same time, and being divided into AD, MCI and CN three categories respectively, each category contains balanced training set and test set.

[0013] The ADNI data set collects various test indicators for each patient, with more than 100 test indicators, including age, disease, test date, scale score and other related data, and dementia is divided into five types of senile dementia, vascular dementia, frontotemporal dementia, Lewy body dementia and Parkinson's dementia. The application mainly studies Alzheimer's disease, i.e. senile dementia, so other parameters need to be introduced to exclude other types of dementia as much as possible and ensure that the data set can show the general disease process and characteristics of AD. Under the guidance of clinicians, the following three statistical parameters of the sample are selected for balanced processing:

[0014] 1) Gender: In AD patients, compared with male patients, female patients have longer average life expectancy, lower prevalence rate, and longer transition period from MCI to AD, and the change process is slower. Studies have shown that this is because sex chromosomes have a certain role in regulating Alzheimer's disease-related aspects. Therefore, the male and female ratio should tend to 1:1 when dividing data.

[0015] 2) Age: Age is one of the biggest risk factors for Alzheimer's disease. The World Alzheimer Report 2021 Journey through the diagnosis of dementia points out that more than 80% of Alzheimer's patients are 65 years old and above. With the increase of age, the risk of Alzheimer's disease increases sharply. The brain structure in imaging will change with age, and brain atrophy does not necessarily prove that Alzheimer's disease is present. Therefore, the degree of brain atrophy cannot be used to divide the data set, and the age distribution should also be an important basis for dividing the data set.

[0016] 3) Cognitive test scale: Cognitive test scale is a neuropsychological test used by clinical neurologists to assess patients' general cognitive abilities, memory, language expression, logical judgment, vision, etc. By asking informants and patients themselves, each item can be scored accurately to check the overall cognitive ability of the subjects, which is a powerful indicator for diagnosing AD. Imaging can assess the brain from the perspective of structural and functional changes, but it cannot assess patients from the perspective of cognitive assessment. Therefore, the data set should not be divided one-sidedly from the perspective of imaging, but also from the perspective of cognitive test scale results. The scale referred to in this paper is the Clinical Dementia Rating (CDR), the Clinical Dementia Rating Scale-sum of Boxes (CDR-SOB) and the Mini-mental State Examination (MMSE).

[0017] As the channel separation residual module described in step three, the description is as follows:

[0018] The design of this module is based on the characteristics of Alzheimer's disease research in MRI, which requires attention to both nonlinear high-dimensional features and low-dimensional features such as hippocampal size and gray matter white matter distribution. The module contains a separation layer, a convolution layer, a normalization layer, an activation function layer and a feature fusion layer. This module inherits the advantage of residual module that can avoid gradient explosion and gradient disappearance in the process of deepening the network, and through the separation of channels, the gradient flow propagates in different network paths, effectively preventing excessive repeated gradient information. The extracted features of the two parts are fused, the low-dimensional and high-dimensional features are fully combined, thereby greatly improving the learning ability of the network, maintaining the stability of the network, and significantly reducing the computational cost.

[0019] The specific process is: the pre-processed MRI image is sent into a 3x3x3 convolution layer to extract shallow features, and then sent into a continuous stack of four channel separation residual modules. The features retain the advantage of residual modules that can avoid gradient explosion and gradient disappearance in the process of deepening the network, and through the separation of feature extraction paths, the low-dimensional and high-dimensional features are fully combined to improve the learning ability of the network; first, the feature map x is input, after the separation layer, it is divided into and two feature extraction paths, as shown in equation (1), where c is the number of channels;

[0020] (1)

[0021] Then through the convolution layer, normalization and activation function, abstract nonlinear high-level features are extracted, and No operation is performed, and the low-level features are retained; finally, the features extracted from the two paths are fused according to formula (2) to obtain the output Y;

[0022] (2)

[0023] in Indicates to and The feature fusion operation between the two paths, F represents the feature fusion operation between the two paths. Convolution, normalization, and activation function operations are performed, and W represents the weight parameters to extract and combine shallow and deep features.

[0024] The channel separation residual module primarily performs feature extraction, fully combining low-dimensional and high-dimensional features while reducing redundant gradient information. The four consecutive channel separation residual modules (R1, R2, R3, R4) differ in that the number of channels is halved sequentially. This structure is adopted for two reasons: First, shallow features in MRI images are crucial in clinical research on Alzheimer's disease, and to extract sufficient shallow features, more convolutional kernels are needed in the early stages of feature extraction. Second, setting a smaller number of channels during high-dimensional feature extraction achieves dimensionality reduction and regularization, thus avoiding overfitting. Attached Figure Description

[0025] Figure 1 This is the channel separation residual module of the present invention.

[0026] Figure 2 This is the channel attention mechanism module of the present invention.

[0027] Figure 3 This invention provides a classification model framework for Alzheimer's disease based on channel residual attention. Detailed Implementation

[0028] Table 1 shows the distribution of the standard dataset after scientific and balanced partitioning according to this invention.

[0029] Table 2 shows the distribution of the non-standard dataset that was randomly shuffled in this invention.

[0030] Table 3 shows the comparison results between the model proposed in this invention and the basic model on the standard dataset.

[0031] Table 4 shows the comparison results between the model proposed in this invention and the basic model on non-standard datasets.

[0032] Table 5 shows the ablation experiment results of the optimal embedding position of the channel attention module in this invention.

[0033] Table 1. Distribution of the Standard-Dataset dataset

[0034]

[0035] Where Sub represents the number of samples, M indicates the number of males, F represents the number of females; Age represents the average age and standard deviation of the samples; MMSE, CDR and CDR-SOB respectively represent the average scores and standard deviations of the corresponding scales of the samples.

[0036] In order to facilitate the comparison of the frontier algorithm, verify the effectiveness of the algorithm in this paper and show the necessity of scientific division of the data set, it is necessary to construct a data set Non-standard Dataset which can be roughly qualitative analysis. Using the pre-processing method of most researchers, the training set and test set are divided after randomly shuffling the samples, and the data set has relative data leakage. The data distribution is shown in Table 2:

[0037] Table 2 Distribution of Non-standard Dataset data set

[0038]

[0039] Table 3 Comparison of classification accuracy of different networks (%)

[0040]

[0041] Table 3 is to explore the effectiveness of the model in this paper under the Standard-Dataset data set, reproduce open source algorithms and mainstream networks. Liu et al. proposed a feature extraction module composed of 3D convolution, instance regularization, ReLU activation function and maximum pooling, and stacked four layers of the module for feature extraction, and then used a fully connected layer for classification output, which is similar to the structure of this paper. By comparing the results of the classification task, it can be concluded that the classification accuracy of the network model proposed in this paper is higher than that of the basic network, and the binary classification accuracy is higher than that of Liu's method, and the three classification accuracy is slightly lower than that of Liu's method. It shows that the method proposed in this paper has good effect on classification.

[0042] Table 4 Comparison of classification accuracy of frontier networks (%)

[0043]

[0044] The methods compared in Table 4 are all in the presence of a certain degree of data leakage. As can be seen from the results, the accuracy of the three classification AD / MCI / CN of the method of the present application on the Non-standard Dataset dataset reached 81.94%, and the accuracy of the two classification AD / CN and MCI / CN reached 95.27% and 85.07%, which were higher than those of some algorithms in the forefront; compared with the data of the Standard-Dataset dataset, the accuracy of each classification task of the method of the present application on the Non-standard Dataset dataset was greatly improved, among which the accuracy of three classification and MCI / CN classification increased by 20%, and the accuracy of AD / CN classification increased by nearly 10%, which was sufficient to illustrate the necessity of scientific division of data sets.

[0045] Table 5 Channel attention module embedding position results comparison (%)

[0046]

[0047] As can be seen from Table 5, comparison experiments 1 and 2-6 can be obtained, the introduction of the channel attention module is effective for improving the classification accuracy of the model, comparison experiments 2 to 5 can be obtained, the channel attention module is introduced after the channel separation residual module R4, the R4 module is the last module of feature extraction, and the extracted features can better reflect the differences of the AD / CN classification task, cooperate with the channel attention module, adjust the dependence between channels, make the network pay more attention to important features, extract more accurate classification features, and thus further improve the classification accuracy. In addition, experiments show that the channel attention module can be added to R2 and R4 to achieve good results, therefore, experiment 6 is designed, the channel attention module is added after R2 and R4, and each index is not greatly improved compared with experiment 5, and more data operation amount is introduced, and memory is consumed. According to the Occam razor principle, do not increase entities if not necessary. According to the experimental results and theoretical proof, embedding the channel attention module after the module R4 can achieve better results.

Claims

1. A channel residual attention-based Alzheimer's disease classification model, characterized in that Comprising the following steps: Step one: Based on the public ADNI dataset, select subjects with MRI images, age, gender and cognitive test scale data to construct the dataset, and first divide the data set according to patient ID to avoid the same ID from different periods of longitudinal samples from appearing in the training set or test set at the same time, and divide into AD, MCI and CN three categories respectively, each category contains training set and test set; Then under the guidance of clinicians, select three statistical parameters of sample age, gender and cognitive test scale for data balancing processing to ensure that the data set can show the general disease process and characteristics of Alzheimer's disease; Step two: Preprocess the MRI image dataset constructed in step one to obtain the MRI image after registration, sampling and normalization; Step three: the pre-processed MRI image is sent into a 3x3x3 convolution layer to extract shallow features, and then sent into a continuous stack of four channel separation residual modules. The residual modules can avoid gradient explosion and gradient disappearance in the process of deepening the network, and fully combine low-dimensional and high-dimensional features through the separation feature extraction path to improve the learning ability of the network. First, the input feature map x is separated into and two feature extraction paths as shown in equation (1), where c is the number of channels. (1) Then Through the convolution layer, normalization and activation function, the abstract nonlinear high-level features are extracted, and Without any operation, the low-level features are reserved; finally, the features extracted by the two paths are fused in the channel according to formula (2) to obtain the output Y; (2) in Indicates to and The feature fusion operation between the two paths, F represents the feature fusion operation between the two paths. Perform convolution, normalization, and activation function operations, where W represents the weight parameters to extract and combine shallow and deep features; Step four: the extracted features are sent into the further fusion step three channel separation residual module extracted features to obtain more accurate classification features; first, the length, width, height and channel number are LxWxHxC feature maps After input, first, the length, width and height based on the feature map are globally averaged to pool the spatial features to 1x1x1xC feature map and obtain the global receptive field, as shown in equation (3): （3） Then the non-linear relationship between the individual channels is learned using two fully connected layers and a non-linear activation function ReLU, resulting in ; as shown in equation (4): (4) wherein , ; and finally, the normalized weights are obtained by passing the Sigmoid activation function, and then multiplied by each channel of the original feature map to complete the channel attention to the original feature re-rating; as shown in equation (5): (5) Finally, the output feature map is obtained ; Step five: The feature matrix after dimension reduction is used as the input information of the linear classification layer of the whole network.

Images