An alzheimer's disease diagnosis method based on contrast learning and mamba

Abstract

The application belongs to the field of artificial intelligence and medical technology, and specifically relates to an Alzheimer's disease diagnosis method based on contrast learning and Mamba, comprising the following steps: after feature fusion of images in three dimensions, slice by dimension and utilize a pre-trained convolutional neural network for feature extraction. The extracted features are processed through nonlinear mapping, and each slice is represented as a feature vector. Subsequently, the feature vectors fused with multiple dimension information are input into the selective state space model Mamba, which has advantages in long distance dependence capture and context memory modeling, and the complex relationship between slices is modeled and analyzed. Finally, an integrated KAN classifier is used to classify the input data and predict whether the sample has Alzheimer's disease. The method effectively improves the diagnosis accuracy, reduces the dependence on computing resources, and has good clinical application value.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence and medical technology, and specifically relates to an Alzheimer's disease diagnosis method based on contrastive learning and Mamba. Background Art

[0002] Alzheimer's disease (AD) is an irreversible neurodegenerative disease and one of the most common cognitive impairments in the elderly population, especially in today's aging society. It is estimated that there are more than 55 million people with dementia worldwide, and this number is expected to soar to 139 million by 2050. The exact cause of AD is still unknown, but once the disease occurs, it is incurable. In the past, diagnosis mainly relied on the rich clinical experience of doctors, which was very time-consuming and laborious. Brain scans, such as structural magnetic resonance imaging (sMRI), provide a non-invasive way to capture the pathological pattern of the disease. Currently, structural magnetic resonance imaging (sMRI) has become an important tool for detecting neurodegenerative diseases in clinical practice, providing valuable insights into the dynamic morphological characteristics associated with Alzheimer's disease (AD).

[0003] The key to dementia diagnosis based on structural magnetic resonance imaging (sMRI) lies in learning discriminative representations. Brain atrophy caused by Alzheimer's disease (AD) is subtle and occurs only in a few localized regions, making it challenging to extract discriminative feature representations from sMRI for accurate AD diagnosis. Inspired by advances in deep learning, extensive research has focused on using powerful deep neural networks (DNNs) as sMRI feature extractors to learn discriminative disease representations, including two-dimensional convolutional neural networks (2D CNNs), three-dimensional convolutional neural networks (3D CNNs), and Transformers. 2D-CNN methods typically transfer networks pretrained on ImageNet (e.g., ResNet) to classify sMRI slices and then aggregate all slice-level predictions to produce subject-level predictions. 3D-CNN methods apply 3D convolutions to whole-brain sMRI or to a few empirically predefined anatomical regions and directly make subject-level predictions. Due to the increased kernel dimensionality, these methods tend to adopt shallower architectures to avoid a large number of training parameters.

[0004] In recent years, with the popularity of visual Transformers, some studies have explored re-adapting the Transformer architecture to adapt sMRI slices or whole-brain sMRI for diagnostic tasks related to Alzheimer's disease. Although these studies have reported impressive diagnostic accuracy, further improvement has been hindered by the neglect of dementia-related regions. This is because the feature representations at the higher levels of deep neural networks tend to respond to the global semantics of the entire image. On the other hand, Transformers also consume considerable computing resources, which limits their deployment in clinical applications, given that medical institutions are unlikely to be equipped with very expensive computing equipment. Summary of the Invention

[0005] To address the technical issues of traditional Alzheimer's disease diagnostic methods, this paper proposes an Alzheimer's disease diagnostic method based on contrastive learning and Mamba. This method focuses on enhancing the model's ability to diagnose Alzheimer's disease using sMRI while reducing computational resource consumption. The model design aligns with the diagnostic mechanisms of Alzheimer's disease and achieves a 94% accuracy rate for distinguishing between AD and NC using only sMRI data.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A method for diagnosing Alzheimer's disease based on contrastive learning and Mamba, comprising the following steps:

[0008] S1. Dataset selection and establishment: Select and collect T1 structural MRI data from different public datasets to construct a dataset for diagnosing Alzheimer's disease;

[0009] S2. Data preprocessing: The data preprocessing process includes: format conversion, noise removal, bias field correction, skull stripping, image registration, resampling, intensity normalization and secondary format conversion;

[0010] S3. Model Construction: This model consists of two parts. The first part is the contrastive learning framework, which aims to help the model learn the underlying features of the image. This module generates the positive sample pairs required for contrastive learning by randomly scrambling the order of sMRI slices in a certain dimension. The second part is the MMK model. The MMK model consists of four parts: feature fusion module, multi-view multi-plane feature extraction module, Mamba module, and classifier module. The model is trained using the dataset to obtain a trained model.

[0011] S4. Use the AdamW algorithm to train and optimize the model and set the corresponding model training parameters;

[0012] S5. Training model: Use the training set, validation set, and test set to train, validate, and test the model. Use the cross-entropy loss function and save the best model in the validation process based on the evaluation indicators. Use the test set to experimentally test the effectiveness of the proposed model. All data are divided according to the subjects to ensure that there is no data leakage.

[0013] The method for selecting and collecting T1 structural magnetic resonance imaging (sMRI) data from different public datasets in S1 is as follows: selecting and collecting T1 structural magnetic resonance imaging (sMRI) data from ADNI, AIBL and OASIS databases, deleting unavailable data contained therein, collecting T1 structural magnetic resonance imaging (sMRI) data from the ADNI database to form an ADNI dataset, which is split into a training set, a validation set and a test set. The division here is based on the subjects, so that the data of the same subject will not appear in the training set and the test set or the validation set at the same time, avoiding data leakage problems; the T1 structural magnetic resonance imaging (sMRI) data collected from the AIBL dataset and the OASIS dataset are used as the test set to test and verify the generalization and robustness of the model.

[0014] The initial processing of the input data in S2 is as follows: first, the raw data is converted from DCM format to NIfTI format; then, the image is processed using the N4 bias field correction algorithm to remove the influence of low-frequency bias fields caused by magnetic field inhomogeneities; then, the sMRI data is skull-stripped using the HD-BET neural network to remove non-brain tissue; next, the processed sMRI is registered to the MNI152 standard template to ensure spatial consistency; then, the image is resampled to adjust the voxel spacing; finally, the image intensity of all voxels is normalized using zero-mean unit variance normalization; finally, the normalized NIfTI image is converted to the .npy format suitable for model reading. The data size after preprocessing is H = W = D.

[0015] In the S3, a contrastive learning framework is used to pre-train the model so that the model can initially capture the underlying features of sMRI, and the selective state space model Mamba and KAN are used to process the vectorized three-dimensional information.

[0016] The method for training the model is:

[0017] S3.1. Use a contrastive learning framework to train on the generated positive and negative sample pairs. Positive samples are generated by randomly selecting one of the three dimensions of the structural magnetic resonance imaging (sMRI) image: the sagittal, axial, and coronal planes, and scrambling the image slice order of one of the dimensions.

[0018] S3.2. The MMK model is pre-trained using positive and negative samples to obtain the weights of a pre-trained model. All subsequent work will be performed on this model.

[0019] S3.3. Construct an MMK model using pre-trained weights to process structural magnetic resonance imaging (sMRI) data.

[0020] The method for constructing the MMK model for processing structural magnetic resonance imaging (sMRI) data using pre-trained weights in S3.3 is:

[0021] S3.3.1. Feature Encoding Module: This module is based on a three-dimensional convolutional neural network (3D-CNN). The initial convolutional layer uses a convolution kernel of size 5x5x5, a stride of 1, and a padding of 2 to ensure that the input size remains unchanged after the convolution operation. Subsequently, batch normalization is implemented through BatchNorm3d to accelerate model training and improve network stability. The activation function uses GELU, introducing nonlinear transformations to enhance the model's expressiveness. Next, the module includes a second three-dimensional convolutional layer, which still uses a 5x5x5 convolution kernel, a stride of 1, and a padding of 2, and is normalized with BatchNorm3d. The entire module further enhances its ability to capture and express input image features by increasing the number of output channels layer by layer while keeping the input size constant.

[0022] S3.3.2 Multi-view multi-plane feature embedding module: This module includes two stages: slicing each dimension and extracting features from the slices. In the slicing stage, the positions of the axial, sagittal, coronal and channel dimensions are swapped in turn to achieve multi-dimensional slicing of structural magnetic resonance imaging (sMRI). Next, the sliced ​​data is merged in the second dimension, that is This generates a sequence of slices containing three dimensions. After entering the feature extraction phase, the slices are processed using a pre-trained ResNet34 model, which maps each slice from different dimensions into a vector representation. This is followed by a nonlinear mapping module: This module, consisting of two multi-layer perceptrons (MLPs) and a RELU activation function, performs nonlinear mapping on the feature vector output by the previous module, adjusting its shape to the target dimension to meet the needs of subsequent processing.

[0023] S3.3.3. Construction of Mamba module:

[0024] This module consists of a selective state space model Mamba, which is used to perform long-distance modeling and capture global context information of the feature vector sequence output by the above modules. Its calculation formula is as follows:

[0025]

[0026] in and The calculation formula is as follows:

[0027]

[0028] , the shape of the data will not change after passing through this module, which is convenient for subsequent processing.

[0029] S3.3.4. Classifier module: This module consists of an adaptive average pooling layer, a flattening layer, a dropout layer, and a KAN layer. The adaptive average pooling layer is used to reduce the dimensionality of the input features and to generate a feature map with a fixed output size by adaptively adjusting the size of the pooling window, thereby improving the model's adaptability to inputs of different sizes. Subsequently, the multi-dimensional feature map is converted into a one-dimensional vector through the flattening layer to prepare for the subsequent processing of the fully connected layer. Next, a dropout layer is added to randomly block some neurons during the training process, thereby effectively preventing overfitting and enhancing the generalization ability of the model. Finally, the KAN layer is used to model high-dimensional features and complete classification through its powerful nonlinear mapping capability.

[0030] In S4, the AdamW algorithm is used to train and optimize the model, and β=0.9, β2=0.999 are set, 50 epochs are trained, the learning rate is 1e-5, and the batch size is 3.

[0031] The model training process in S5 is as follows:

[0032] After preprocessing all structural magnetic resonance imaging (sMRI) images, the contrastive learning component randomly shuffles the order of slices along a certain dimension of the sMRI to generate the positive sample pairs required for contrastive learning. The remaining data in a batch serves as negative samples and is input into the MMK model. Using a class-consistent contrastive loss function, the parameters of each layer in the network are updated through backpropagation of the loss and stochastic gradient descent. This generates a set of pretrained weights.

[0033] In the second part of MMK training, the data in the ADNI dataset is divided into training set, validation set and test set; the training set is input into the MMK model using the pre-trained weights generated by contrastive learning for training, and the cross-entropy loss function is used to calculate the error between the predicted value and label output by the model, and the parameters in the model are updated through backpropagation and gradient descent algorithms; at the same time, the validation set is used for model selection at the end of each epoch.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention fully exploits the rich feature information contained in 3D medical images. After fusing the features of the images in three dimensions (coronal, sagittal and axial), the images are sliced ​​dimension by dimension and feature extracted using a pre-trained convolutional neural network. The extracted features are processed by nonlinear mapping to represent each slice as a feature vector. Subsequently, the feature vector that fuses information from multiple dimensions is input into the selective state space model Mamba, which uses its advantages in capturing long-distance dependencies and modeling contextual memory to model and analyze the complex relationships between slices. Finally, a classifier integrating KAN (Kolmogorov–Arnold Network) is used to classify the input data and predict whether the sample has Alzheimer's disease. This method effectively improves diagnostic accuracy, reduces dependence on computing resources, and has good clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can, without inventive effort, derive other implementation drawings based on the provided drawings.

[0037] The structures, proportions, sizes, etc. illustrated in this specification are intended solely to complement the contents disclosed herein and to facilitate understanding and reading by persons skilled in the art. They are not intended to limit the conditions under which the present invention may be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportions, or adjustments in sizes, without affecting the efficacy and objectives of the present invention, shall remain within the scope of the technical contents disclosed herein.

[0038] Figure 1 is a flow chart of the method of the present invention;

[0039] Figure 2 Schematic diagram of the contrastive learning framework proposed in the present invention;

[0040] Figure 3 This is a schematic diagram of the structure of the model proposed in the present invention;

[0041] Figure 4 It is the experimental result curve graph of the present invention;

[0042] Figure 5 It is a bar graph of the experimental results of the present invention. DETAILED DESCRIPTION

[0043] 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. Obviously, the described embodiments are only part of the embodiments of this application, not all the embodiments. These descriptions are only to further illustrate the features and advantages of the present invention, rather than to limit the claims of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0044] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0045] like Figures 1 to 5 As shown, the present invention proposes a method for diagnosing Alzheimer's disease based on contrastive learning and Mamba, comprising the following steps:

[0046] Step 1. Dataset selection and establishment: Select and collect the required T1 structural MRI data from ADNI1, ADNIGO, ADNI2, ADNI3, OASIS, and AIBL from the IDA official website to construct a dataset for diagnosing Alzheimer's disease. The specific steps are as follows:

[0047] Step 1.1: Collect 4133 T1 structural MRI images from the ADNI database, including 1.5T and 3T images. Divide these data into training, validation, and test sets according to the subjects, with the ratio of each part being 8:1:1.

[0048] In step 1.2, we collected 335 T1 structural MRI images from the OASIS database and 997 T1 structural MRI images from the AIBL database. These two datasets served as test sets to verify the generalization and robustness of the trained model.

[0049] Step 2: Data preprocessing: Data preprocessing mainly includes: format conversion; noise filtering; bias field correction; skull stripping; image registration; resampling; intensity normalization; format conversion again; and dividing the processed data into training set, validation set, and test set according to the subjects (to ensure no data leakage). The specific steps are as follows:

[0050] Step 2.1: Convert the raw data from DICOM format to NIfTI format.

[0051] Step 2.2: Use the N4 bias field correction algorithm to process the image to eliminate the influence of the low-frequency bias field caused by magnetic field inhomogeneity.

[0052] Step 2.3: Use the HD-BET neural network to perform skull stripping on the structural magnetic resonance imaging (sMRI) to remove non-brain tissue.

[0053] Step 2.4: Register the processed sMRI images to the MNI152 standard template to ensure spatial consistency.

[0054] Step 2.5: Resample the image to standardize the voxel spacing, adjust the voxel size to (1.75 mm × 1.75 mm × 1.75 mm), and unify the spatial resolution of the volume data to (128 × 128 × 128) voxels.

[0055] Step 2.6: Normalize the image intensity of all voxels using zero-mean unit variance normalization.

[0056]

[0057] Among them, x is the value of the original data; x ′ is the normalized value; μ is the mean of the original data, and σ is the standard deviation of the original data.

[0058] Step 2.7: Convert the normalized NIfTI image to .npy format suitable for model reading.

[0059] Step 3: Model Construction: The entire model consists of two parts: the contrastive learning framework and the MMK model. In the contrastive learning module, a dimension of the sMRI image is randomly selected, and the slice order along this dimension is randomly shuffled. This generated data sample forms a positive pair with the original sample, while other sMRI data in the same batch serve as negative samples. The MMK model consists of four parts: a feature encoding module, a multi-view and multi-plane feature embedding module, a Mamba module, and a classifier module.

[0060] The specific steps are as follows:

[0061] Step 3.1: Use the contrastive learning framework to train on the generated positive and negative sample pairs. The positive samples are generated by randomly selecting one dimension from the sagittal, axial, and coronal planes of the sMRI image and disrupting the order of the image slices in this dimension. These samples are used to pre-train the model and finally generate a set of pre-trained parameters, which are considered to be a set of parameters that have learned the underlying features of the data. All subsequent tasks will be performed on this set of pre-trained parameters.

[0062] The class-consistent contrast loss function used in step 3.2, contrastive learning, is as follows:

[0063]

[0064] Where q is the query sample, k + is the generated positive sample, k - is a negative sample, and the temperature parameter τ ensures that the similarity distribution is smoother. Numerator: exp(q·k + / τ) represents the query q and the positive sample k + Denominator: represents the sum of similarities between query q and all samples (including positive and negative samples).

[0065] Step 3.3, MMK's feature encoding module: use a 5x5x5 convolution kernel with a step size of 1 and padding = 2; use BatchNorm3d; use the GELU activation function; 3D convolutional network, use a 5x5x5 convolution kernel with a step size of 1 and padding = 2; use BatchNorm3d; use the GELU activation function; this module increases the output channels while keeping the input size unchanged to enhance the feature representation of the input image.

[0066] Step 3.4, MMK multi-view multi-plane feature embedding module: This module consists of two parts: slicing operation and slicing feature extraction. In the slicing operation part, the structural magnetic resonance imaging (sMRI) data is sliced ​​by sequentially exchanging the positions of the axial, sagittal, coronal and channel dimensions, that is, Then, the slice data are fused along the second dimension to generate a slice sequence containing information of three dimensions at the same time, i.e. Next, in the feature extraction part, a pre-trained two-dimensional convolutional neural network ResNet-34 is used to extract features from the slice sequence. The final output feature shape is batch × sequence_length × 512. The nonlinear mapping module consists of two multi-layer perceptrons (MLPs) and a RELU activation function. The shapes of the two MLPs are (512 × 256) and (256 × 128), respectively. RELU is used to increase the nonlinear characteristics of the model. Its specific formula is as follows:

[0067]

[0068] This module maps the feature vector generated by the previous module to the desired 128-dimensional vector to further refine the feature information, thereby enhancing the feature expression ability and the discriminative performance of the model.

[0069] Step 3.5, Mamba module:

[0070] This module consists of a selective state space model Mamba, which is used to perform long-distance modeling and capture global context information of the feature vector sequence output by the above modules. Its calculation formula is as follows:

[0071]

[0072] in and The calculation formula is as follows:

[0073]

[0074] Mamba processes these features through its Selective Scan Mechanism (S6), effectively capturing the contextual relationships and complex long-range dependencies between slices. At the same time, the output of the Mamba module maintains the same shape as the input.

[0075] Step 3.6, Classifier Module: The Adaptive Average Pooling layer is used to reduce the dimensionality of the input features. Its value is selected as 32. By adaptively adjusting the size of the pooling window, a feature map with a fixed output size is generated, thereby improving the model's adaptability to inputs of different sizes. Subsequently, the multidimensional feature map is converted into a one-dimensional vector through the Flatten Layer to prepare for the subsequent processing of the fully connected layer. Next, a Dropout layer is added, in which the dropout value is selected as 0.8, which is used to randomly block some neurons during the training process, thereby effectively preventing overfitting and enhancing the generalization ability of the model. Finally, the Kolmogorov-Arnold Networks (KAN) layer is used, and the specific details are as follows:

[0076] KAN(Z)=(Φ K-1 °Φ K-2 °…°Φ1°Φ0)Z, Z is the input feature vector,

[0077] Through its nonlinear mapping capability, high-dimensional features are modeled and classification is completed.

[0078] Step 4: Use the AdamW algorithm to optimize model training and set the corresponding model training parameters: β1 = 0.9, β2 = 0.999, train for 50 epochs, use a learning rate of 1e-5, a batch size of 3, and a weight decay rate of 1e-4. Use a custom learning rate change strategy, keeping the learning rate constant at the beginning of training and then gradually reducing it as training progresses. This improves the stability and efficiency of model training.

[0079] Step 5: Training of MMK model:

[0080] After preprocessing all structural MRI images, the contrastive learning phase randomly perturbs the slice order of one dimension of the sMRI to generate the positive sample pairs required for contrastive learning. The remaining data in a batch serves as negative samples and is input into the MMK model. Using a class-consistent contrastive loss function, the parameters of each layer in the network are updated through backpropagation of the loss and stochastic gradient descent. This generates a set of pretrained weights.

[0081] In the second phase of MMK training, the ADNI dataset was divided into training, validation, and test sets. The training set was fed into the MMK model using the pre-trained weights generated through contrastive learning. The cross-entropy loss function was used to calculate the error between the model's output predictions and labels. The model parameters were updated using backpropagation and gradient descent. The validation set was used for model selection at the end of each epoch.

[0082] Step 6: Model testing and evaluation:

[0083] The trained MMK model was tested on the test set, and the test results on the ADNI, AIBL and OASIS datasets were obtained respectively.

[0084] This paper uses four indicators to evaluate the performance of the classification model: accuracy (ACC), area under the curve (AUC), sensitivity (SEN), and specificity (SPE). These indicators are very commonly used in medical image analysis and binary classification tasks. Their definitions and formulas are as follows:

[0085] Step 6.1, Accuracy (ACC), the calculation formula is as follows:

[0086] Step 6.2, Sensitivity (SEN): also known as recall, is the model's ability to identify positive samples. The calculation formula is:

[0087] Step 6.3: Specificity (SPE) is the model's ability to identify negative samples. The calculation formula is as follows:

[0088] Where: TP, TN, FP and FN represent true positive, true negative, false positive and false negative respectively. TN: the number of samples correctly classified as negative, FP: the number of negative samples incorrectly classified as positive, TP: the number of sample data correctly classified as positive, and FN: the number of positive samples incorrectly classified as negative.

[0089] Step 6.4. AUC (Area Under the Curve) is an important indicator for evaluating the performance of a classification model. It refers to the area under the ROC curve. The horizontal axis of the ROC curve is the false positive rate, and the vertical axis is the true positive rate. The larger the area, the better the model performance.

[0090] The method of the present invention was compared with advanced methods at home and abroad, and the comparison results are shown in Tables 1, 2, and 3. As can be seen from the results, compared with other methods, the Alzheimer's disease diagnosis method based on contrastive learning and Mamba proposed in the present invention has the advantages of high accuracy and high robustness.

[0091] Table 1 Comparison of diagnostic results of different methods on structural MRI collected from ADNI dataset

[0092] method ACC AUC SEN SPE 3D ResNet152 0.8344 0.8692 0.7504 0.8292 3D ViT 0.8524 0.8135 0.7367 0.8489 MRNet 0.8796 0.9316 0.8204 0.8974 MedicalNet 0.8889 0.8880 0.8162 0.8739 M3T 0.8980 0.9104 0.8367 0.9563 The present invention 0.9402 0.9453 0.8381 0.9856

[0093] Table 2 Comparison of diagnostic results of different methods on structural MRI collected from AIBL dataset

[0094]

[0095]

[0096] Table 3 Comparison of diagnostic results of different methods on structural MRI collected from OASIS dataset

[0097] method ACC AUC SEN SPE 3D ResNet152 0.7134 0.7287 0.6639 0.8533 3D ViT 0.7569 0.7713 0.7109 0.8441 MRNet 0.7077 0.8197 0.7001 0.8527 MedicalNet 0.7385 0.7272 0.6938 0.8482 M3T 0.8047 0.8167 0.7895 0.9109 The present invention 0.8574 0.8459 0.7914 0.9218

[0098] As shown in Tables 1, 2, and 3, the proposed Alzheimer's disease diagnosis method based on contrastive learning and Mamba achieves optimal performance across all metrics on the ANDI, AIBL, and OASIS public datasets, outperforming 3D ResNet152, 3D ViT, MRNet, MedicalNet, and M3T. These results demonstrate the effectiveness of the proposed method.

[0099] The above only describes in detail the preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the purpose of the present invention, and various changes should be included in the scope of protection of the present invention.

Claims

1. A method for diagnosing Alzheimer's disease based on contrastive learning and Mamba, characterized in that: The following steps are involved: S1. Dataset selection and establishment: We selected and collected T1 structural magnetic resonance imaging (sMRI) data from various public datasets to construct a dataset for diagnosing Alzheimer's disease. S2. Data preprocessing: The data preprocessing process includes: format conversion, noise removal, bias field correction, skull stripping, image registration, resampling, intensity normalization and secondary format conversion; S3. Model Construction: This model consists of two parts. The first part is the contrastive learning framework, which aims to help the model learn the underlying features of the image. This first part generates the positive sample pairs required for contrastive learning by randomly disrupting the order of slices in a certain dimension of the MRI image. The second part is the MMK model. The MMK model consists of four parts: feature fusion module, multi-view multi-plane feature extraction module, Mamba module, and classifier module. The model is trained using the dataset to obtain a trained model. S4. Use the AdamW algorithm to train and optimize the model and set the corresponding model training parameters; S5. Training model: Use the training set, validation set, and test set to train, validate, and test the model. Use the cross-entropy loss function and save the best model in the validation process based on the evaluation indicators. Use the test set to experimentally test the effectiveness of the proposed model. All data are divided according to the subjects to ensure that there is no data leakage.

2. The Alzheimer's disease diagnosis method based on contrastive learning and Mamba according to claim 1, characterized in that: The method for selecting and collecting T1 structural magnetic resonance imaging sMRI data from different public datasets in S1 is: selecting and collecting T1 structural magnetic resonance imaging sMRI data from ADNI, AIBL and OASIS databases, deleting unavailable data contained therein, collecting T1 structural magnetic resonance imaging sMRI data from the ADNI database to form an ADNI dataset, which is split into a training set, a validation set and a test set, and the division here is based on the division conducted by the subjects; the T1 structural magnetic resonance imaging sMRI data collected from the AIBL dataset and the OASIS dataset are used as the test set to test and verify the generalization and robustness of the model.

3. The Alzheimer's disease diagnosis method based on contrastive learning and Mamba according to claim 1, characterized in that: The method for preliminary processing of input data in S2 is as follows: first, the raw data is converted from DICOM format to NIfTI format; then, the image is processed using the N4 bias field correction algorithm to eliminate the influence of low-frequency bias fields caused by magnetic field inhomogeneity; then, the sMRI data is skull stripped using the HD-BET neural network to remove non-brain tissue; next, the processed sMRI is aligned to the MNI152 standard template to ensure spatial consistency; then, the image is resampled and the voxel spacing is adjusted; then, the image intensity of all voxels is normalized using the zero-mean unit variance normalization method; finally, the normalized NIfTI image is converted into a .npy format suitable for model reading. The data size after preprocessing is H=W=D.

4. The Alzheimer's disease diagnosis method based on contrastive learning and Mamba according to claim 1, characterized in that: In the S3, a contrastive learning framework is used to pre-train the model so that the model can initially capture the underlying features of sMRI, and the selective state space model Mamba and KAN are used to process the vectorized three-dimensional information.

5. The Alzheimer's disease diagnosis method based on contrastive learning and Mamba according to claim 4, characterized in that: The method for training the model is: S3.

1. Use a contrastive learning framework to train on the generated positive and negative sample pairs. The positive samples are generated by randomly selecting one of the three dimensions of the sMRI image: the sagittal, axial, and coronal planes, and scrambling the image slice order of one of the dimensions. S3.

2. The MMK model is pre-trained using positive and negative samples to obtain the weights of a pre-trained model. All subsequent work will be performed on this model. S3.

3. Construct an MMK model using pre-trained weights to process structural magnetic resonance imaging (sMRI) data.

6. The Alzheimer's disease diagnosis method based on contrastive learning and Mamba according to claim 5, characterized in that: The method for constructing the MMK model for processing sMRI data using pre-trained weights in S3.3 is: S3.3.

1. Feature Encoding Module: This module is based on a 3D-CNN. The initial convolutional layer uses a 5x5x5 kernel with a stride of 1 and a padding of 2 to ensure that the input size remains unchanged after the convolution operation. Subsequently, batch normalization is implemented through BatchNorm3d to accelerate model training and improve network stability; The activation function uses GELU, introducing nonlinear transformations to improve the model's expressiveness. Next, the module includes a second three-dimensional convolutional layer, still using a 5x5x5 kernel, a stride of 1, and a padding of 2. BatchNorm3d is also used for normalization. The entire module increases the number of output channels layer by layer while maintaining a constant input size, further improving its ability to capture and express input image features. S3.3.2 Multi-view multi-plane feature embedding module: This module includes two stages: slicing each dimension and extracting features from the slices. In the slicing stage, the positions of the axial plane, sagittal plane, coronal plane and channel dimension are swapped in turn to achieve multi-dimensional slicing of sMRI data. ; Then, merge the sliced ​​data in the second dimension, that is , forming a sequence of slices containing three dimensions; After entering the feature extraction stage, the slices are processed using a pre-trained ResNet34 model, which maps each slice from different dimensions into a vector representation. This model then passes through a nonlinear mapping module: This module consists of two layers of multi-layer perceptrons (MLPs) and a RELU activation function, which performs nonlinear mapping on the feature vector output by the previous module to adjust the shape of the feature vector to the target dimension to meet the needs of subsequent processing. S3.3.

3. Construction of Mamba module: This module consists of a selective state space model Mamba, which is used to perform long-distance modeling and capture global context information of the feature vector sequence output by the above modules. Its calculation formula is as follows: in and The calculation formula is as follows: The shape of the data will not change after passing through this module; S3.3.4, Classifier module: This module consists of an adaptive average pooling layer, a flattening layer, a dropout layer, and a KAN layer; The Adaptive Average Pooling layer is used to reduce the dimensionality of the input features. By adaptively adjusting the size of the pooling window, a feature map with a fixed output size is generated, thereby improving the model's adaptability to inputs of different sizes. Subsequently, the multidimensional feature map is converted into a one-dimensional vector through the Flatten Layer to prepare for the subsequent processing of the fully connected layer. Next, a Dropout layer is added to randomly block some neurons during training, thereby effectively preventing overfitting and enhancing the generalization ability of the model. Finally, the KAN layer is used to model high-dimensional features and complete classification through its powerful nonlinear mapping capabilities.

7. The method for diagnosing Alzheimer's disease based on contrastive learning and Mamba according to claim 1, characterized in that: In S4, the AdamW algorithm is used to train and optimize the model, and the = 0.9, = 0.999, train for 50 epochs, with a learning rate of 1e-5 and a batch size of 3.

8. The method for diagnosing Alzheimer's disease based on contrastive learning and Mamba according to claim 1, characterized in that: The model training process in S5 is as follows: After preprocessing all structural MRI images, the contrastive learning phase randomly shuffles the order of slices in a certain dimension of the sMRI to generate the positive sample pairs required for contrastive learning. The remaining data in a batch is used as negative samples and fed into the MMK model. Using a class-consistent contrastive loss function, the parameters of each layer in the network are updated through backpropagation of the loss and stochastic gradient descent. This generates a set of pretrained weights. In the second part of MMK training, the data in the ADNI dataset is divided into training set, validation set and test set; the training set is input into the MMK model using the pre-trained weights generated by contrastive learning for training, and the cross-entropy loss function is used to calculate the error between the predicted value and label output by the model, and the parameters in the model are updated through backpropagation and gradient descent algorithms; at the same time, the validation set is used for model selection at the end of each epoch.

Images