AD Early Diagnosis Method, System, Medium and Device Based on Deep Learning

Abstract

The present invention provides an early AD diagnosis method, system, medium and device based on deep learning, including: acquiring T1-MRI image data of AD and corresponding classification labels; constructing a discrete wavelet transform module to decompose the MRI image into low-frequency subbands and high-frequency subbands of different scales, and obtaining a feature map by means of a Concat operation on all subband images; constructing a feature extraction module to extract multi-scale features from the feature map; designing a multi-class classifier to classify the feature map; building a classification network for processing the classification task of T1-MRI images; and training the classification network on a test set to output a diagnosis result. The present invention uses T1-MRI images for feature extraction, and through the design of a multi-scale feature extraction module and a multi-class classifier, can accurately capture the subtle changes in the brains of MCI patients, significantly improving the accuracy of MCI diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and specifically, to an early AD diagnosis method, system, medium, and device based on deep learning. Background Art

[0002] As a common neurodegenerative disease, Alzheimer's disease (AD) severely affects the cognitive function and quality of life of patients. Early screening for AD can greatly delay the progression of the disease, improve the quality of life of patients, and reduce the social burden. As the prodromal stage of AD, Mild Cognitive Impairment (MCI) allows for accurate diagnosis and timely intervention during this period, which is expected to prevent or delay its conversion to AD. Therefore, the diagnosis of MCI occupies a core position in the AD early screening process. Among the numerous technologies for AD early screening, T1-weighted magnetic resonance imaging (T1-MRI) technology exhibits unique advantages. T1-MRI can clearly present the anatomical structure of the brain, providing a rich and accurate data basis for subsequent image analysis and feature extraction, helping to improve the diagnostic accuracy of MCI and providing strong support for AD early screening.

[0003] Currently, existing AD early screening methods are limited by the network structure and are unable to effectively extract key features when processing neuroimaging data, resulting in insufficient ability to capture subtle changes in the brains of MCI patients. This problem has kept the diagnostic accuracy of MCI at a relatively low level for a long time, severely restricting its application effect in clinical practice.

[0004] Therefore, there is an urgent need for an early AD diagnosis method based on T1-MRI to achieve precise screening of MCI, thereby assisting doctors in early AD diagnosis and clinical intervention.

[0005] Patent application document CN113208620A discloses an Alzheimer's disease screening method and system based on sleep staging. The method includes the following steps: a BCI device collects EEG signals during the user's sleep and transmits them to a sleep staging algorithm system; the sleep staging algorithm system extracts and analyzes the features of the user's EEG signals, discriminates the sleep state, and stages it; the early screening system extracts and analyzes the features of the sleep staging results to determine whether the user is in the high-risk stage of early onset of Alzheimer's disease. However, this patent cannot completely solve the existing technical problems and does not meet the requirements of the present invention. Summary of the Invention

[0006] Aiming at the defects in the prior art, the purpose of the present invention is to provide a method, system, medium and device for early diagnosis of AD based on deep learning.

[0007] According to the method for early diagnosis of AD based on deep learning provided by the present invention, it includes:

[0008] Step S1: Obtain the T1-weighted magnetic resonance imaging (T1-MRI) image data of Alzheimer's disease (AD) and the corresponding classification labels, denoted as and respectively. Divide into a training set and a test set according to a preset ratio, which are used for model training and performance evaluation and verification of the model respectively;

[0009] Step S2: Construct a discrete wavelet transform module (MWT) to decompose the MRI image into low-frequency sub-bands and high-frequency sub-bands of different scales, and obtain a feature map through a Concat operation on all sub-band images ;

[0010] Step S3: Construct a feature extraction module to extract multi-scale features from the feature map . The feature extraction module includes a Mamba module (MB) and a Dense module (DM). The MB module extracts global information through global modeling, and the DM module uses a densely connected convolutional structure to enhance the local feature representation ability;

[0011] Step S4: Design a multi-class classifier (MCC) to classify the feature map;

[0012] Step S5: Based on the constructed MWT, MB, DM and MCC modules, design and build a classification network ED-Mamba for processing the classification task of T1-MRI images;

[0013] Step S6: ED-Mamba is trained on until the loss function tends to be stable, save the final model parameters, output the diagnostic results, and finally test ED-Mamba on the test set .

[0014] Preferably, the step S1 includes: screening the obtained T1-MRI images, excluding preset low-quality images and data desensitization to ensure that all images meet the predetermined quality standards; using image processing technology to perform size resampling on the screened T1-MRI images, adjusting the size of each image to 224×224×16; dividing them into a training set and a test set according to a ratio of 8:2.

[0015] Preferably, the step S2 includes:

[0016] Step S2.1: Establish a scaling function and a wavelet function , as shown in the following formula:

[0017]

[0018]

[0019] where j is the scaling parameter and k is the translation parameter;

[0020] Step S2.2: For each sample , where are the length, width, and number of channels respectively; perform a one-dimensional wavelet transform operation on along the W direction. The low-frequency coefficient and the high-frequency coefficient are obtained by the following formula:

[0021]

[0022]

[0023] Step S2.3: Perform a one-dimensional wavelet transform in the H direction to obtain the low-frequency coefficient and the high-frequency coefficient, and finally obtain 4 sub-bands , ; each sub-band is filtered through a 1 1 convolution, and the wavelet features that fuse the global and local information of multiple sizes are obtained through the Concat operation , which is obtained by the following formula:

[0024]

[0025] where is the convolutional layer.

[0026] Preferably, the step S3 includes:

[0027] Step S3.1: Construct an MB module. The MB module includes a dual-branch VSS layer and a patch merging layer; the dual-branch VSS layer contains a convolutional branch and a spatial perception module SSM branch to extract global and local features; the patch merging layer halves the size of the feature map and doubles the number of channels; define the input of the th MB module as , and in the dual-branch VSS layer, perform a channel partitioning operation on to obtain and ; Next, a series of operations are performed on and to obtain the output of the dual-branch VSS layer , which can be obtained by the following formula:

[0028]

[0029] where represents the channel aggregation operation that restores the number of channels of the feature map to the original number of channels, represents the channel discretization operation that randomly shuffles the feature map at the channel level and preserves the information within the channels, is obtained from the convolutional branch of the feature map and is obtained by the following formula:

[0030]

[0031] where and represent the ReLU activation function and the batch normalization operation; is the channel aggregation operation used to restore the number of channels of the feature map to the original number of channels;

[0032] is obtained from the SSM branch of the feature map and is obtained by the following formula:

[0033]

[0034] represents the element-wise addition operation of the matrix, represents a linear transformation operation, and the feature maps and are obtained by the following formula:

[0035]

[0036]

[0037] where and are the SiLU activation function and the layer normalization operation, and the SSM operation enables the feature map to contain the information of the global receptive field, is the depthwise separable convolution;

[0038] Step S3.2: Construct the DM module. Each DM module contains 6 layers of dense convolutions, and each layer of dense convolution contains convolution operations, batch normalization operations, and ReLU activation functions. Define the input of the DM module as , and the output of the DM module is obtained by the following formula:

[0039]

[0040] Among them, and represent the ReLU activation function and the batch normalization operation.

[0041] Preferably, the step S4 includes:

[0042] Construct a multi-class classifier. When there are k classes, construct k binary classifiers , for learning the decision boundary between the k-th class and other classes; construct a dynamic threshold adjustment strategy to determine the threshold for each class according to the difficulty of distinguishing each class. The difficulty of distinguishing each class is determined by calculating the predicted value of the classifier for the sample, and the difficulty of distinguishing each class is defined as , and is calculated by the following formula:

[0043]

[0044] Among them, is the total number of samples, represents the predicted value of the sample for the sample at the t-th training round, The larger it is, the easier it is to distinguish the samples of the -th class; then normalize to [0,1], and use the normalized to scale the predefined threshold , and the formula is as follows:

[0045]

[0046] Among them, is the dynamic threshold of the -th class at the t-th training round, is the maximum value of the difficulty of all classes at the t-th training round.

[0047] Preferably, the step S5 includes:

[0048] Step S5.1: Build the modules proposed in steps S2 to S4 to obtain ED-Mamba. This network includes 1 MWT module, 4 MB modules, 4 DM modules and 1 MCC module; send the image into the MWT module to obtain the output feature map ;

[0049] Step S5.2: Send Input 1 MB module and 1 DM module to obtain the output feature map ;

[0050] Step S5.3: Repeat step S5.2 three times to obtain the feature map ;

[0051] Step S5.4: Feed into the MCC module to obtain the final output of the model , being the category of

[0052] Preferably, the said step S6 includes:

[0053] Step S6.1: Train ED-Mamba in , using the Adam optimizer with a learning rate of . After 400 iterations, the network converges, and save the model parameters at this time;

[0054] Step S6.2: Load the test set into the trained model parameters for prediction to obtain the classification result and calculate the accuracy rate;

[0055] Step S6.3: If the accuracy rate is lower than 0.9, repeat step S6.1 and step S6.2.

[0056] According to the AD early diagnosis system based on deep learning provided by the present invention, it includes:

[0057] A data acquisition module, used to acquire the T1-weighted magnetic resonance imaging T1-MRI image data of the patient and the corresponding classification labels;

[0058] A preprocessing module, used to screen the T1-MRI images, resample the size to 224×224×16, and divide them into a training set and a test set according to a ratio;

[0059] A discrete wavelet transform module MWT, used to decompose the T1-MRI image into multi-scale low-frequency sub-bands and high-frequency sub-bands, and generate multi-scale feature maps through convolution fusion;

[0060] A feature extraction module, including a Mamba module MB and a dense convolution module DM, used to extract global and local features from the multi-scale feature maps;

[0061] A multi-class classifier MCC, adopting a dynamic threshold adjustment strategy to adaptively adjust the decision boundary based on the classification difficulty;

[0062] A classification network ED-Mamba, integrating the MWT module, MB module, DM module and MCC module, and outputting the final diagnosis result.

[0063] A computer-readable storage medium storing a computer program according to the present invention, when the computer program is executed by a processor, implements the steps of the AD early diagnosis method based on deep learning.

[0064] An electronic device according to the present invention, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, when the computer program is executed by the processor, implements the steps of the AD early diagnosis method based on deep learning.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] (1) The present invention uses T1-MRI images for feature extraction. Through the design of a multi-scale feature extraction module and a multi-class classifier, it can accurately capture the subtle changes in the brains of MCI patients, significantly improving the accuracy of MCI diagnosis, with an accuracy rate reaching above 0.9.

[0067] (2) Through the MWT module and the MB module, multi-scale features in T1-MRI images are effectively extracted and fused in the convolutional neural network, enhancing the model's ability to capture local and global information of the images.

[0068] (3) The present invention adopts a dynamic threshold adjustment strategy, sets different thresholds for each category, optimizes the performance of the classifier, enables the classifier to adaptively adjust the decision boundary according to the difficulty of each category, and further improves the classification accuracy.

[0069] (4) The ED-Mamba network of the present invention combines the MWT, MB, DM, and MCC modules to form an efficient and robust classification network. This network structure can not only effectively extract features from T1-MRI images but also perform accurate classification, providing strong support for early AD diagnosis.

[0070] (5) This method is based on T1-MRI image data for diagnosis, can adapt to different situations of different patients, has strong versatility and clinical application value. Through efficient training and prediction, it can provide accurate diagnostic information for doctors, assist doctors in early intervention, and delay the progression of AD.

[0071] (6) The present invention uses the Adam optimizer and the method of dynamically adjusting the learning rate to ensure the efficiency of the training process and the stability of the model, enabling a high accuracy to be achieved within a fewer number of training epochs, reducing the training cost and time. Description of the Drawings

[0072] Other features, objectives, and advantages of the present invention will become more apparent by reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0073] Figure 1 It is a flowchart of an early Alzheimer's disease diagnosis method based on deep learning. Specific embodiments

[0074] The present invention will be described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several changes and improvements can still be made. These all belong to the protection scope of the present invention.

[0075] Embodiment

[0076] As Figure 1 , the present invention provides an early AD diagnosis method based on deep learning, including:

[0077] Step S1: Obtain the T1-weighted magnetic resonance imaging (T1-MRI) image data of Alzheimer's disease (AD) and the corresponding classification labels, denoted as and respectively. Divide into a training set and a test set in a ratio of 8:2, where is used for model training, and is used for the performance evaluation and verification of the model;

[0078] Step S2: Design and construct a discrete wavelet transform module (MWT) to decompose the MRI image into low-frequency and high-frequency subbands at different scales. Then, all subband images are concatenated through a Concat operation to obtain a feature map ;

[0079] Step S3: Construct a feature extraction module to extract multi-scale features from the feature map . The feature extraction module includes a Mamba module (MB) and a Dense module (DM). The MB module extracts global information through global modeling, and the DM module uses a densely connected convolutional structure to enhance the local feature representation ability;

[0080] Step S4: To improve the classification accuracy, design a multi-class classifier (MCC) to classify the feature map;

[0081] Step S5: Based on the above constructed MWT, MB, DM, and MCC modules, design and build a classification network ED-Mamba for processing the classification task of T1-MRI images;

[0082] Step S6: ED-Mamba is trained on until the loss function stabilizes, and the final model parameters are saved. Finally, ED-Mamba is tested on . The accuracy of the proposed deep learning-based early AD diagnosis method of the present invention is not less than 0.9.

[0083] The specific steps of step S1 include:

[0084] The obtained T1-MRI images are screened to exclude low-quality images (such as excessive noise, blurring, incomplete images, etc.) and data desensitization to ensure that all images meet the predetermined quality standards; the screened T1-MRI images are resampled in size using image processing techniques, and the size of each image is adjusted to 224×224×16; they are divided into a training set and a test set in a ratio of 8:2;

[0085] The specific steps of step S2 include:

[0086] Step S2.1: Construct the MWT module. First, establish the scaling function and the wavelet function as follows:

[0087]

[0088]

[0089] where j is the scaling parameter and k is the translation parameter;

[0090] Step S2.2: For each sample , where are the length, width, and number of channels respectively; first, perform a one-dimensional wavelet transform operation on along the W direction. The low-frequency coefficient and the high-frequency coefficient in the W direction can be calculated by the following formula:

[0091]

[0092]

[0093] Step S2.3: Similarly, perform a one-dimensional wavelet transform in the H direction to also obtain the low-frequency coefficient and the high-frequency coefficient. Finally, 4 subbands can be obtained, ; subsequently, each subband passes through 1 1. The convolution is used for filtering, and the wavelet features that fuse global and local information of multiple sizes are obtained through the Concat operation. , which can be obtained by the following formula:

[0094]

[0095] Where is the convolutional layer;

[0096] Step S3 specifically includes:

[0097] Step S3.1: Construct the MB module. The MB module includes a dual-branch VSS layer and a patch merging layer. The dual-branch VSS layer contains a convolutional branch and a spatial perception module (SSM) branch to extract global and local features. The patch merging layer halves the size of the feature map and doubles the number of channels. Define the input of the th MB module as . In the dual-branch VSS layer, perform a channel division operation on to obtain and ; then, perform a series of operations on and to obtain the output of the dual-branch VSS layer. can be obtained by the following formula:

[0098]

[0099] Where represents the channel aggregation operation to restore the number of channels of the feature map to the original number of channels, represents the channel discretization operation to randomly scatter the feature map at the channel level and preserve the information within the channels, is obtained by the convolutional branch of the feature map and can be obtained by the following formula:

[0100]

[0101] Where and represent the ReLU activation function and the batch normalization operation; is the channel aggregation operation used to restore the number of channels of the feature map to the original number of channels;

[0102] is obtained by the SSM branch of the feature map and can be obtained by the following formula:

[0103]

[0104] represents the element-wise addition operation of the matrix represents a linear transformation operation, and the feature map and can be obtained by the following formula:

[0105]

[0106]

[0107] where and SiLU activation function and layer normalization operation, the SSM operation enables the feature map to contain information of the global receptive field, is the depthwise separable convolution.

[0108] Step S3.2: Construct a DM module. Each DM module contains 6 layers of dense convolutions, and each layer of dense convolution contains convolution operations, batch normalization operations, and ReLU activation functions. Define the input of the DM module as , and the output of the DM module can be obtained by the following formula:

[0109]

[0110] where and represent the ReLU activation function and batch normalization operation.

[0111] Step S4 specifically includes:

[0112] Construct a multi-class classifier. When there are k classes, construct k binary classifiers , for learning the decision boundary between the k-th class and other classes; construct a dynamic adjustment threshold strategy to determine the threshold for each class according to the difficulty of distinguishing each class. The difficulty of distinguishing each class can be determined by calculating the predicted value of the classifier for the sample. Define the difficulty of distinguishing each class as , which can be calculated by the following formula:

[0113]

[0114] where is the total number of samples, represents the predicted value of the sample at the t-th training round for the sample The larger is, the easier it is to distinguish the samples of the -th class; then normalize to [0, 1], and use the normalized To scale the predefined threshold , the formula is as follows:

[0115]

[0116] Where, is the dynamic threshold of the class at the t-th training round, is the maximum value of the difficulty levels of all classes at the t-th training round.

[0117] Step S5 specifically includes:

[0118] Step S5.1: Build the modules proposed in Steps S2 to S4 to obtain ED-Mamba. This network contains 1 MWT module, 4 MB modules, 4 DM modules, and 1 MCC module; First, send the image into the MWT module to obtain the output feature map ;

[0119] Step S5.2: Send into 1 MB module and 1 DM module to obtain the output feature map ;

[0120] Step S5.3: Repeat Step S5.2 three times to obtain the feature map ;

[0121] Step S5.4: Send into the MCC module to obtain the final output of the model, where is the class of

[0122] Step S6 specifically includes:

[0123] Step S6.1: Train ED-Mamba using the Adam optimizer with a learning rate of . After 400 iterations, the network converges, and save the model parameters at this time;

[0124] Step S6.2: Load the test set into the trained model parameters for prediction to obtain the classification result and calculate the accuracy rate;

[0125] Step S6.3: If the accuracy rate is lower than 0.9, repeat Steps S6.1 and S6.2.

[0126] The present invention also provides an early AD diagnosis system based on deep learning, including:

[0127] A data acquisition module for acquiring T1-weighted magnetic resonance imaging (T1-MRI) image data of a patient and corresponding classification labels;

[0128] A preprocessing module for performing data screening on the T1-MRI images, resampling the size to 224×224×16, and dividing them into a training set and a test set according to a ratio;

[0129] A discrete wavelet transform module MWT for decomposing the T1-MRI images into multi-scale low-frequency subbands and high-frequency subbands, and generating multi-scale feature maps through convolutional fusion;

[0130] A feature extraction module including a Mamba module MB and a dense convolution module DM for extracting global and local features from the multi-scale feature maps;

[0131] A multi-class classifier MCC that adopts a dynamic threshold adjustment strategy to adaptively adjust the decision boundary based on the classification difficulty;

[0132] A classification network ED-Mamba that integrates the MWT module, MB module, DM module, and MCC module to output the final diagnosis result.

[0133] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps of the AD early diagnosis method based on deep learning are implemented.

[0134] The present invention also provides an electronic device including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the computer program is executed by the processor, the steps of the AD early diagnosis method based on deep learning are implemented.

[0135] Those skilled in the art know that in addition to implementing the systems, devices, and their respective modules provided by the present invention in the form of pure computer-readable program codes, the method steps can be logically programmed to enable the systems, devices, and their respective modules provided by the present invention to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers, etc. to achieve the same program. Therefore, the systems, devices, and their respective modules provided by the present invention can be regarded as a kind of hardware component, and the modules included therein for implementing various programs can also be regarded as the structures within the hardware component; the modules for implementing various functions can also be regarded as both software programs for implementing the method and the structures within the hardware component.

[0136] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An early diagnosis method for AD based on deep learning, characterized in that, Including: Step S1: Obtain the T1-weighted magnetic resonance imaging (T1-MRI) image data of Alzheimer's disease (AD) and the corresponding classification labels, denoted as and respectively. Divide into a training set and a test set according to a preset ratio, which are used for model training and performance evaluation and verification of the model respectively; Step S2: Construct a discrete wavelet transform module MWT to decompose the MRI image into low-frequency subbands and high-frequency subbands of different scales, and obtain a feature map by performing a Concat operation on all subband images ; Step S3: Construct a feature extraction module to extract multi-scale features from the feature map wherein the feature extraction module includes a Mamba module MB and a Dense module DM. The MB module extracts global information through global modeling, and the DM module uses a densely connected convolutional structure to enhance the local feature representation ability; Step S4: Design a multi-class classifier MCC to classify the feature map; Step S5: Based on the constructed MWT, MB, DM, and MCC modules, design and build a classification network ED-Mamba for processing the classification task of T1-MRI images; Step S6: ED-Mamba is trained on until the loss function tends to be stable, save the final model parameters, and finally test ED-Mamba on the test set ; The said Step S5 includes: Step S5.1: Build the modules proposed in Steps S2 to S4 to obtain ED-Mamba. This network includes 1 MWT module, 4 MB modules, 4 DM modules, and 1 MCC module; send the image into the MWT module to obtain the output feature map ; Step S5.2: Feed into one MB module and one DM module to obtain the output feature map ; Step S5.3: Repeat step S5.2 three times to obtain the feature map ; Step S5.4: Send to the MCC module to obtain the final output of the model , which is the category of.

2. The early AD diagnosis method based on deep learning according to claim 1, characterized in that, The step S1 includes: screening the acquired T1-MRI images, excluding the preset low-quality images and performing data desensitization to ensure that all images meet the predetermined quality standards; using image processing techniques to perform size resampling on the screened T1-MRI images, adjusting the size of each image to 224×224×16; dividing them into a training set and a test set in a ratio of 8:2 and a test set .

3. The early AD diagnosis method based on deep learning according to claim 1, characterized in that, The said Step S2 includes: Step S2.1: Establish a scaling function and a wavelet function , as follows: where j is the scale parameter and k is the translation parameter; Step S2.2: For each sample , where are the length, width, and number of channels respectively; perform a one-dimensional wavelet transform operation on along the W direction. The low-frequency coefficient and high-frequency coefficient are obtained by the following formula: Step S2.3: Perform a one-dimensional wavelet transform in the H direction to obtain low-frequency coefficients and high-frequency coefficients, and finally obtain 4 sub-bands , ; Each sub-band is filtered through a 1 ×1 convolution and the wavelet features that fuse multi-scale global and local information are obtained through a Concat operation , which is obtained by the following formula: Among them, is a convolutional layer.

4. The early AD diagnosis method based on deep learning according to claim 3, characterized in that, The said Step S3 includes: Step S3.1: Construct an MB module. The MB module includes a dual-branch VSS layer and a patch merging layer. The dual-branch VSS layer contains a convolutional branch and a spatial perception module (SSM) branch to extract global and local features. The patch merging layer halves the size of the feature map and doubles the number of channels. Define the input of the th MB module as . In the dual-branch VSS layer, perform a channel partitioning operation on to obtain and . Then, perform a series of operations on and to obtain the output of the dual-branch VSS layer. is obtained by the following formula: Among them, represents that the channel aggregation operation restores the number of channels of the feature map to the original number of channels, represents that the channel discretization operation randomly shuffles the feature map at the channel level and preserves the information within the channels, is obtained from the convolutional branch of the feature map and is obtained by the following formula: Among them, and represent the ReLU activation function and the batch normalization operation; is the channel aggregation operation, which is used to restore the number of channels of the feature map to the original number of channels; is obtained from the feature map through the SSM branch and is obtained by the following formula: Represents the addition operation of corresponding elements of the matrix, Indicates a linear transformation operation, and the feature map and Is obtained by the following formula: Among them, and With the SiLU activation function and layer normalization operation, the SSM operation enables the feature map to contain information about the global receptive field, is the depthwise separable convolution; Step S3.2: Construct the DM module. Each DM module contains 6 layers of dense convolutions. Each layer of dense convolution contains convolution operations, batch normalization operations, and ReLU activation functions. Define the input of the DM module as , and the output of the DM module is obtained by the following formula: Among them, and represent the ReLU activation function and the batch normalization operation.

5. The AD early diagnosis method based on deep learning according to claim 4, characterized in that The said Step S4 includes: Construct a multi-class classifier. When there are k classes, construct k binary classifiers , for learning the decision boundary between the k-th class and other classes; construct a dynamic threshold adjustment strategy to determine the threshold for each class according to the difficulty of distinguishing each class. The difficulty of distinguishing each class is determined by calculating the predicted value of the classifier for the sample, and the difficulty of distinguishing each class is defined as , the larger it is, the easier it is to distinguish the samples of the -th class; then normalize to [0, 1], and use the normalized to scale the predefined threshold , and the formula is as follows: Among them, is the dynamic threshold of the category at the t-th training round, is the maximum value of the difficulty levels of all categories at the t-th training round.

6. The early AD diagnosis method based on deep learning according to claim 1, characterized in that The said Step S6 includes: Step S6.1: Train ED-Mamba in using the Adam optimizer with a learning rate of . After 400 iterations, the network converges, and save the model parameters at this time; Step S6.2: Load the test set into the trained model parameters for prediction to obtain the classification result and calculate the accuracy rate; Step S6.3: If the accuracy rate is lower than 0.9, then repeat Step S6.1 and Step S6.

2.

7. An early AD diagnosis system based on deep learning, characterized in that, Adopt the deep learning-based early AD diagnosis method described in Claim 1, including: A data acquisition module for acquiring T1-weighted magnetic resonance imaging (T1-MRI) image data of patients and corresponding classification labels; A preprocessing module for screening the T1-MRI images, resampling the size to 224×224×16, and dividing them into a training set and a test set according to a ratio; A discrete wavelet transform module MWT for decomposing the T1-MRI images into multi-scale low-frequency sub-bands and high-frequency sub-bands, and generating multi-scale feature maps through convolutional fusion; A feature extraction module, including a Mamba module MB and a dense convolution module DM, for extracting global and local features from the multi-scale feature maps; A multi-class classifier MCC, adopting a dynamic threshold adjustment strategy to adaptively adjust the decision boundary based on the classification difficulty; A classification network ED-Mamba, integrating the MWT module, MB module, DM module, and MCC module, and outputting the final diagnosis result.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the deep learning-based early AD diagnosis method described in any one of Claims 1 to 6.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the deep learning-based early AD diagnosis method described in any one of Claims 1 to 6.

Images