A multi-scale spatial information network classification method based on structural magnetic resonance imaging

Abstract

A multi-scale spatial information network classification method based on structural magnetic resonance imaging relates to the field of medical image processing. It solves the problem that the existing Alzheimer's disease classification method based on sMRI does not sufficiently mine the detailed information of lesions and has weak generalization ability. The present invention provides a multi-scale spatial information network classification method based on sMRI. The method uses sMRI as input data, extracts multi-scale frequency information through a multi-frequency perception fusion module based on three-dimensional sym2 wavelet transform, and uses a graph hybrid attention module to flexibly model the relationship between channels and spatial position relationships, thereby improving the model's ability to mine detailed information of sMRI lesions and generalization performance. The present invention is also applicable to the field of Alzheimer's disease image processing and multi-scale spatial information network classification equipment based on structural magnetic resonance imaging.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a multi-scale spatial information network classification method based on structural magnetic resonance imaging. Background Art

[0002] Alzheimer's disease (AD) is a neurodegenerative disorder characterized by progressive cognitive impairment. As the most common form of dementia, AD patients experience progressively worsening cognitive impairment and impairment in daily living abilities as the disease progresses. Currently, there is no effective cure for AD, but early intervention and comprehensive treatment strategies can slow disease progression, significantly improve patients' quality of life, and reduce long-term care costs.

[0003] sMRI is a non-invasive imaging technique based on a strong magnetic field that can provide high-resolution, high-contrast detailed information on the internal anatomical structure and morphology of brain tissue. Compared to other neuroimaging techniques, sMRI can clearly demonstrate the three-dimensional spatial distribution of AD-related brain regions, such as the hippocampus, medial temporal lobe, and frontal lobe.

[0004] At present, the deep learning classification method based on sMRI is still insufficient in mining detailed information of Alzheimer's lesions, and its generalization ability on other datasets is weak. Summary of the Invention

[0005] The present invention addresses the problems in existing sMRI-based Alzheimer's disease classification methods, which are insufficient in mining lesion details and have weak generalization capabilities. To solve the above technical problems, the present invention is implemented through the following technical solutions:

[0006] The present invention proposes a multi-scale spatial information network classification method based on structural magnetic resonance imaging. The multi-scale spatial information network classification method is used in the Alzheimer's disease image processing process. The method includes the following steps:

[0007] Step 1: Obtain and preprocess sMRI images from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Australian Alzheimer's Neuroimaging, Biomarkers and Lifestyle Study (AIBL), and the Open Access Alzheimer's Disease Imaging Dataset (OASIS). Construct the whole-brain image dataset required for the classification model based on the preprocessed sMRI images.

[0008] Step 2: Divide the whole-brain image dataset constructed in step 1 into a test set and a validation set;

[0009] Step 3: Based on the test set and validation set divided in step 2, a multi-frequency perception fusion module (MFAFM) based on three-dimensional sym2 wavelet transform is constructed to perform multi-scale spatial information processing in the frequency domain to capture features at different levels.

[0010] Step 4: Based on the features of different levels captured in step 3, a graph hybrid attention mechanism module GHA is constructed. The graph hybrid attention mechanism module GHA is implemented through a dual attention coordination mechanism that combines the graph channel attention mechanism and the spatial attention mechanism.

[0011] Step 5: The inverted residual module GHA_MB in the graph hybrid attention mechanism module GHA constructed in step 4 is used as the backbone module of the multi-scale spatial information network model. The convolutional layer of the inverted residual module GHA_MB is implemented using an inverted residual structure.

[0012] Step 6: Build a multi-scale spatial information network model, combine the multi-frequency perception fusion module MFAFM built in step 3, the graph hybrid attention mechanism module GHA built in step 4, and the inverted residual module GHA_MB built in step 5, and juxtapose the classifier to complete the final classification task;

[0013] Step 7: The multi-scale spatial information network model is trained and validated on the Alzheimer's Disease Neuroimaging Initiative dataset (ADNI), and its generalization is verified on the AIBL biomarker and lifestyle study dataset and the OASIS open access series Alzheimer's disease imaging dataset, ultimately obtaining an Alzheimer's disease classification model.

[0014] Furthermore, a preferred embodiment is provided, wherein the preprocessing in step 1 includes AC-PC fusion and skull stripping, and spatial normalization to the template MIN152 operation.

[0015] Furthermore, a preferred embodiment is provided, in step 2, the method for dividing the whole-brain image dataset constructed in step 1 into a test set and a validation set is:

[0016] A five-fold cross-validation was performed on the Alzheimer's Disease Neuroimaging Project dataset (ADNI), with four copies used as test sets and one as validation set.

[0017] Furthermore, a preferred embodiment is provided, wherein the preprocessing in step 1 includes AC-PC fusion and skull stripping, and spatial normalization to the template MIN152 operation.

[0018] Furthermore, a preferred embodiment is provided, in which the method for implementing the hybrid attention mechanism module GHA in step 4 through the dual attention coordination mechanism is as follows:

[0019] Graph channel attention mechanism, which is used to model inter-channel relationships through graph structures to emphasize information-rich features and suppress useless features;

[0020] The spatial attention mechanism is used to quantify the importance of features at different spatial locations through a learnable spatial weight matrix and locate the specific spatial location of the lesion;

[0021] The dual-attention collaborative mechanism is used to enable the multi-scale spatial information network model to simultaneously capture key information of the channel dimension and the spatial dimension, and the two work together to achieve adaptive enhancement of multi-scale features.

[0022] Furthermore, a preferred embodiment is provided, wherein step 5 also includes the steps of expanding, processing and compressing the input features.

[0023] Furthermore, a preferred embodiment is provided, wherein the size of the whole brain image dataset required for constructing the classification model using the preprocessed sMRI images in step 1 is 128 128 128-pixel whole-brain image.

[0024] Solution 2: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method described in any one of Solution 1 is implemented.

[0025] Solution three: A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of Solution one.

[0026] The present invention is beneficial in that:

[0027] The multi-frequency perception fusion module based on three-dimensional sym2 wavelet transform proposed in the present invention effectively extracts multi-scale feature information based on sMRI images. Its low-frequency components can reflect global pathological changes such as overall brain atrophy, while its high-frequency components can capture detailed information in the image, including edges, textures and tiny local changes. Finally, these multi-scale frequency information are fused to provide rich information for subsequent structures.

[0028] The proposed graph hybrid attention mechanism module leverages graph structures to flexibly model inter-channel relationships and combines them with spatial attention mechanisms to pinpoint the specific spatial location of lesions. Responding to the complexity of the spatial structure of 3D MRI images, the graph hybrid attention mechanism module effectively and adaptively enhances multi-scale features.

[0029] This paper proposes a multi-scale spatial information network model based on sMRI, which achieves high-accuracy diagnosis of Alzheimer's disease on the ADNI dataset, and verifies its generalization on the AIBL and OASIS datasets, overcoming the problem of weak generalization of current sMRI-based Alzheimer's disease classification methods.

[0030] The present invention is also applicable to the field of Alzheimer's disease image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flowchart of the multi-scale spatial information network classification method based on structural magnetic resonance imaging described in embodiment 1.

[0032] Figure 2 Schematic diagram of preprocessing of sMRI brain images according to the first embodiment.

[0033] Figure 3 Schematic diagram of the multi-scale spatial information network model constructed for implementation method one. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the implementation methods of this application clearer, the technical solutions in the implementation methods of this application will be clearly and completely described below in combination with the drawings in the implementation methods of this application. Obviously, the described implementation methods are only part of the implementation methods of this application, not all of the implementation methods.

[0035] Embodiment 1: This embodiment provides a multi-scale spatial information network classification method based on structural magnetic resonance imaging. The multi-scale spatial information network classification method is used in the Alzheimer's disease image processing process. The method includes the following steps:

[0036] Step 1: Obtain and preprocess sMRI images from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Australian Alzheimer's Neuroimaging, Biomarkers and Lifestyle Study (AIBL), and the Open Access Alzheimer's Disease Imaging Dataset (OASIS). Construct the whole-brain image dataset required for the classification model based on the preprocessed sMRI images.

[0037] Step 2: Divide the whole-brain image dataset constructed in step 1 into a test set and a validation set;

[0038] Step 3: Based on the test set and validation set divided in step 2, a multi-frequency perception fusion module (MFAFM) based on the three-dimensional sym2 wavelet transform is constructed to perform multi-scale spatial information processing in the frequency domain to capture features at different levels.

[0039] Step 4: Based on the features of different levels captured in step 3, a graph hybrid attention mechanism module GHA is constructed. The graph hybrid attention mechanism module GHA is implemented through a dual attention coordination mechanism, which is implemented based on a graph channel attention mechanism and a spatial attention mechanism.

[0040] Step 5: Use the inverted residual module GHA_MB in the graph hybrid attention mechanism module GHA constructed in step 4 as the backbone module of the multi-scale spatial information network model, and the backbone module is used to extract discriminative features;

[0041] Step 6: Build a multi-scale spatial information network model, combine the multi-frequency perception fusion module MFAFM built in step 3, the graph hybrid attention mechanism module GHA built in step 4, and the inverted residual module GHA_MB built in step 5, and juxtapose the classifier to complete the final classification task;

[0042] Step 7: The multi-scale spatial information network model is trained and validated on the Alzheimer's Disease Neuroimaging Initiative dataset (ADNI), and its generalization is verified on the AIBL Biomarker and Lifestyle Study dataset and the OASIS Open Access Series Alzheimer's Disease Imaging Dataset, ultimately obtaining an Alzheimer's disease classification model.

[0043] Implementation 2: This implementation further limits the multi-scale spatial information network classification method based on structural magnetic resonance imaging described in Implementation 1. The preprocessing described in step 1 includes AC-PC fusion and skull stripping, and spatial normalization to the template MIN152 operation.

[0044] Implementation method 3: This implementation method further limits the multi-scale spatial information network classification method based on structural magnetic resonance imaging described in implementation method 1. In step 2, the method for dividing the whole-brain image dataset constructed in step 1 into a test set and a validation set is:

[0045] A five-fold cross-validation was performed on the Alzheimer's Disease Neuroimaging Project dataset (ADNI), with four copies used as test sets and one as validation set.

[0046] Implementation 4: This implementation further limits the multi-scale spatial information network classification method based on structural magnetic resonance imaging described in Implementation 1. In step 4, the hybrid attention mechanism module GHA is implemented by the dual attention coordination mechanism as follows:

[0047] Graph channel attention mechanism, which is used to model inter-channel relationships through graph structures to emphasize information-rich features and suppress useless features;

[0048] The spatial attention mechanism is used to quantify the importance of features at different spatial locations through a learnable spatial weight matrix and locate the specific spatial location of the lesion;

[0049] The dual-attention collaborative mechanism is used to enable the multi-scale spatial information network model to simultaneously capture key information of the channel dimension and the spatial dimension, and the two work together to achieve adaptive enhancement of multi-scale features.

[0050] Implementation 5: This implementation further limits the multi-scale spatial information network classification method based on structural magnetic resonance imaging described in Implementation 1. Step 5 also includes the steps of expanding, processing and compressing the input features.

[0051] Implementation 6: This implementation is a further limitation of the multi-scale spatial information network classification method based on structural magnetic resonance imaging described in Implementation 1. In step 1, the size of the whole brain image dataset required to construct the classification model using the pre-processed sMRI images is 128 128 128-pixel whole-brain image.

[0052] Implementation method seven. This implementation method proposes a computer device including a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the multi-scale spatial information network classification method based on structural magnetic resonance imaging described in any one of implementation methods one to six.

[0053] Embodiment 8. This embodiment proposes a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the multi-scale spatial information network classification method based on structural magnetic resonance imaging as described in any one of embodiments 1 to 6 is implemented.

[0054] Embodiment 9. This embodiment provides the following examples to explain the above-mentioned embodiments 1 to 8. The specific examples are as follows:

[0055] See also Figures 1 to 3 As shown, this embodiment proposes a multi-scale spatial information network classification method based on structural magnetic resonance imaging, which specifically includes the following steps:

[0056] Step 1: sMRI images were acquired from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Australian Alzheimer's Neuroimaging, Biomarker and Lifestyle Study (AIBL), and the Open Access Alzheimer's Disease Imaging Series (OASIS) dataset and preprocessed. These preprocessing steps included AC-PC consolidation, skull stripping, spatial normalization to the MNI152 template, and resampling. This preprocessing yielded the whole-brain image dataset required for building the classification model.

[0057] Step 1.1: Download sMRI images of both AD and NC (normal) from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Australian Alzheimer's Neuroimaging, Biomarker and Lifestyle Study (AIBL), and the Open Access Alzheimer's Disease Imaging Dataset (OASIS). Then, perform a unified preprocessing operation on the acquired raw images.

[0058] Step 1.2: Preprocessing In the first part, the SPM12 toolbox was used to perform AC-PC fusion operations and perform head motion correction on the images to eliminate the influence of the subject's head movement during data acquisition.

[0059] Step 1.3: In the second part of the preprocessing operation, skull stripping is performed using the CAT12 toolbox to separate non-brain tissues such as the skull, scalp, eyeballs, and fat from the image.

[0060] Step 1.4: In the third part of the preprocessing operation, the image space is normalized to the template MNI152 using the CAT12 toolbox, unifying different brain images into a standard space.

[0061] Step 1.5, the fourth part of the preprocessing operation, uses the SPM12 toolbox to resample the image to a size of 128 128 128-pixel whole-brain image.

[0062] Step 2: Divide the data based on the preprocessed data from step 1. A five-fold cross-validation was performed on the Alzheimer's Disease Neuroimaging Project dataset, with four folds used as test sets and one as validation sets. To verify the model's generalization ability, a cross-dataset validation approach was employed, using the Alzheimer's Disease Neuroimaging Project dataset as the training set and the Australian Alzheimer's Neuroimaging, Biomarker and Lifestyle Study dataset and the Open Access Series Alzheimer's Imaging Dataset (OASIS) as independent validation sets.

[0063] Step 3: Based on the data set divided in step 2, a multi-frequency perception fusion module based on the three-dimensional sym2 wavelet transform is constructed to perform multi-scale processing in the frequency domain to capture features at different levels. The construction process of this module includes the following steps:

[0064] Step 3.1: Import subunits to build the multi-frequency perception fusion module. Import the libraries needed, including torch, torch.nn, and torch.nn.functional.

[0065] Step 3.2: Define a function for converting wavelet transform coefficients to build the multi-frequency perception fusion module and extract wavelet coefficients: Define a function called _to_wavelet_coefs that accepts a wavelet type parameter (wavelet) that supports three types: string, tensor, and Wavelet object, and converts it to wavelet filter coefficients in PyTorch tensor format. By default, the sym2 wavelet is used for the transformation to maintain smoothness and signal compactness.

[0066] Step 3.3. Build a learnable scaling module for constructing a multi-frequency perception fusion module for feature scaling: define a class named _ScaleModule, initialize the __init__() method, accept the shape shape and the initial scaling factor init_scale as parameters; call super(_ScaleModule, self).__init__() to inherit the parent class nn.Module; use nn.Parameter to create a learnable scaling parameter scale; use the forward propagation method forward to scale the input x.

[0067] Step 3.4. Construct a multi-frequency perception fusion module to build a multi-scale spatial information network model and perform multi-scale processing in the frequency domain to capture features at different levels: define a class named MFAFM, initialize the __init__() method, accept seven parameters: the number of input channels in_channels, the number of output channels out_channels, the convolution kernel size kernel_size, the step size stride, the bias bias, and the wavelet type wt_type; call super(MFAFM, self).__init__(); use the assert function to ensure that the number of input and output channels is consistent; use the sym2 wavelet type and call _to_wavelet_coefs('sym2') to obtain the wavelet transform filter self.wt_filter and the inverse transform filter self.iwt_fi lter, used for forward wavelet transform self.dwt and inverse wavelet transform self.idwt; create a base convolution layer self.base_conv for preliminary feature extraction, and combine it with _ScaleModule for scaling to enhance feature expression capabilities; use _ScaleModule to create the wavelet_scale function to scale the features after wavelet transform to enhance stability; define three convolutions with a convolution kernel size of 3 and expansion rates of 1, 2, and 3, respectively, named self.lowfreq_conv_1, self.lowfreq_conv_2, and self.lowfreq_conv_3, for multi-scale processing of low-frequency information; define the self.wavelet_conv function, and set in_channels to in_channels 7. out_channels is set to out_channels 7, to adapt to the high-frequency information after wavelet transform; use the forward propagation method forward to perform wavelet transform on the input x through self.dwt to extract low-frequency and high-frequency features. The high-frequency features are processed by wavelet_convs, while the low-frequency features are processed by lowfreq_conv_1, lowfreq_conv_2, and lowfreq_conv_3, and mean fusion is adopted. After that, the low-frequency and high-frequency features are concatenated and uniformly scaled by wavelet_scale, and the original shape is restored by inverse wavelet transform idwt, and x is put into base_conv to extract basic features, and scaled by _ScaleModule. Finally, the feature x_t obtained by inverse wavelet transform is fused with the basic convolution feature x to obtain the final output.

[0068] Step 4: Based on the dataset divided in step 2, build a graph hybrid attention mechanism module to model the relationship between channels and spatial positions in a flexible way. The construction process of this module is as follows:

[0069] Step 4.1. Import subunits to build the graph hybrid attention mechanism module. Import the required libraries, including torch, torch.nn, and torch.nn.functional.

[0070] Step 4.2. Define a class named GHA to build a multi-scale spatial information network model, which models the relationship between channels and spatial positions in a flexible way: initialize the __init__() method, receive two parameters: the number of input channels in_channel and the channel compression ratio ratio; call super(GHA, self).__init__() to inherit the parent class nn.Module; define the number of hidden channels hide_channel and set it to in_channel / / ratio to reduce computational complexity; define a global pooling layer self.avg_pool for global pooling; define a convolution layer self.conv1 with a convolution kernel size of 1 for dimensionality reduction; define self.softmax for calculating attention weights; to build a graph structure, define the unit matrix self.A0, A0 is used to represent the feature vertex itself, and define a learnable adjacency matrix self.A2 with a smaller initial value, A2 can more flexibly customize the connection relationship between any vertices; define a convolution layer self.conv2 with a convolution kernel size of 1 for self-attention transformation; define self.Relu for nonlinear activation; define a convolution kernel size The convolution layer self.conv3 with a size of 1 is used to calculate channel attention; self.Sigmoid normalization weight is defined; a convolution layer self.conv4 with a convolution kernel size of 1 is defined to restore the number of channels; a convolution layer self.conv5 with a convolution kernel size of 1 is defined to calculate spatial attention; the forward propagation method forward is used to compress the spatial information of the input x through avg_pool, and then it is reduced in dimension through conv1, and then the attention A1 is calculated through conv2 and the channel attention weight is obtained using the Softmax activation function. 0, A1 and A2 form the attention matrix D, where D represents the connection relationship between different feature vertices. Then D is used to update the feature vertices, and conv3 and Relu activation functions are used for nonlinear transformation. Conv4 is used to restore the original number of channels and Sigmoid activation function is used to generate the final channel attention weight, which is multiplied by x to obtain the recalibrated channel-level response x1. Then avg_pool is used to compress the channel dimension of x1 to obtain spatial salient information. Finally, conv5 and Sigmoid are used to learn the weight distribution of the spatial position, which is multiplied by x1 to achieve spatial-level response and obtain the final output x2.

[0071] Step 5: Based on the graph hybrid attention mechanism module constructed in step 4, an inverted residual module based on the graph hybrid attention mechanism is constructed. This module can further capture complex channel and spatial modeling relationships while balancing model capacity and computational efficiency. The construction process of this module is as follows:

[0072] Step 5.1. Import subunits to build the libraries needed for the inverted residual module based on the graph hybrid attention mechanism, including torch, torch.nn, and torch.nn.functional.

[0073] Step 5.2. Define a convolution module class named ConvBNActivation to build an inverted residual module. As a basic functional module, it can be reused in the backbone network or feature fusion paths at all levels: the initialization function __init__() receives six parameters: the number of input channels in_planes, the number of output channels out_planes, the convolution kernel size kernel_size (default is 3), the stride (default is 1), the number of convolution groups groups (default is 1), the optional normalization function norm_layer and the optional activation function activation_layer; use padding=(kernel_size-1) / / 2 to ensure that the output and input sizes are consistent, and realize automatic compensation for the boundaries; if the normalization function and activation function are not passed in, nn.BatchNorm3d is used as the normalization layer by default, and nn.SiLU (Swish function) is used as the default. ) As an activation function, the Swish activation function has a smoother gradient characteristic, which helps to improve the nonlinear expression ability and training stability of the model; through super().__init__(), it inherits the parent class nn.Sequential and encapsulates the following three sub-modules in sequence: the first layer is nn.Conv3d, which is used for feature extraction in three-dimensional space, the input channel is in_planes, the output channel is out_planes, the convolution kernel size is kernel_size, the step size is stride, the number of groups is groups, and the bias item (bias=False) is turned off to reduce parameter redundancy; the second layer is BatchNorm3d, which performs batch normalization on the convolution output to improve training efficiency; the third layer is the SiLU activation function to introduce nonlinear features; as a basic functional module, this module can be reused in the backbone network or feature fusion paths at all levels, providing unified feature mapping capabilities and flexible structural configuration.

[0074] Step 5.3. Define a class named InvertedResidualConfig to construct the structural configuration parameters of the inverted residual module InvertedResidual: The initialization function __init__() receives the following ten parameters: convolution kernel size kernel, number of input channels input_c, number of output channels out_c, channel expansion ratio expanded_ratio, stride, whether to use the GHA module use_GHA, residual path random drop rate drop_rate, module index, whether to enable the multi-frequency perception fusion module use_MFAFM; define self.input, self.kernel, self.out_c, self.use_GHA, self.drop_rate, self.index, self.use_MFAFM in sequence; in addition, define self.expanded_c=self.input_c*expanded_ratio to determine whether channel expansion convolution is required.

[0075] Step 5.4. Define a class named InvertedResidual to build an inverted residual module based on the graph hybrid attention mechanism to further capture the complex channel and spatial modeling relationship: the initialization function __init__() receives two parameters: the configuration object cnf class of InvertedResidualConfig, which contains all the structural settings of the current module (including the number of channels, step size, convolution type, attention mechanism, etc.), and the normalization layer constructor norm_layer; inherit the parent class nn.Module through super(InvertedResidual_3d, self).__init__(); set the residual connection strategy judgment, enable the residual connection when stride=1 and the input and output channels are consistent, otherwise not enabled; create an ordered dictionary named layers to store submodules in order; the activation function activation_layer is set to nn.SiLU by default: set the multi-frequency perception fusion module strategy judgment Determine whether to enable the multi-frequency perception fusion module to replace the ConvBNActivation convolution layer, and further determine whether to enable channel expansion. If expanded_c≠input_c, perform convolution with a convolution kernel size of 1 for channel expansion; set to determine whether to enable the GHA module; set the convolution with a convolution kernel size of 1 to compress the expanded channel expanded_c into the output channel out_c; define a self.block to pass the previously defined layers as parameters to nn.Sequential and assign it to self.block; set a self.dropout to randomly discard certain elements of the input to prevent overfitting; use the forward propagation method forward to pass the input x through the constructed self.block in sequence, and then if dropout is enabled in the configuration, the main branch output is discarded. If the current module enables residual connection, the main branch output is added to the input x to form the final output.

[0076] Step 6. Based on the graph hybrid attention mechanism module constructed in steps 4 and 5 respectively and the inverted residual module based on the graph hybrid attention mechanism, a multi-scale spatial information network model is constructed, and important modules such as MFAFM, GHA, and GHA_MB are combined, and a classifier is set to complete the final classification task.

[0077] Step 6.1. Import subunits for building a multi-scale spatial information network model. Import the required libraries, including torch, torch.nn, torch.nn.functional, GHA, ConvBNActivation, and InvertedResidua.

[0078] Step 6.2, define a class named MSIN for building a multi-scale spatial information network model: the initialization function __init__() receives multiple parameters, the number of categories num_classes (default is 2), the dropout rate dropout_rate (default is 0.2), the connection dropout rate drop_connect_rate (default is 0.2), the optional basic module block (default is InvertedResidual) and the optional normalization function norm_layer (default is nn.BatchNorm3d); inherit the parent class nn.M through super(EfficientNet_b0_3d,self).__init__() odule; then use default_cnf to configure the module, including the convolution kernel size, the number of input and output channels, the expansion ratio, the stride, whether to enable the graph hybrid attention mechanism, the connection drop rate, whether to enable the multi-frequency perception fusion module, and the number of module repetitions; then create a layer structure to combine multiple set modules together; define a self.avgpool for global average pooling to compress all features to a spatial size of 1x1; define a classifier self.classifier, use nn.Linear for classification, and use nn.dropout to prevent overfitting; use the forward propagation method forward to process the input x layer by layer through the convolution blocks of each stage, and finally return the classification result.

[0079] Step 7: Based on the multi-scale spatial information network model in step 6, the model performance is trained and verified. The multi-scale spatial information network model is trained and verified on the ANDI dataset, and the generalization is verified on the AIBL and OASIS datasets, and finally an Alzheimer's disease classification model is obtained.

[0080] Step 7.1. Define a class named NiiDataset to encapsulate medical image data and obtain the data required for training and validating the model. This class inherits from the Dataset class in PyTorch. Its main function is to read image data in NIfTI format under the specified root directory. The __init__ method receives the data root directory path root_dir as a parameter, obtains the complete path of all samples by traversing the subdirectories (such as AD and CN), and stores them in the self.file_list list. Construct a category-to-index mapping dictionary class_to_idx so that digital labels can be used to represent categories during training. In the __getitem__ method, the corresponding image path is obtained by index, and the nibabel library is used to read the .nii file, convert it into a NumPy array, normalize it, and then convert it into the PyTorch tensor format. Finally, the image tensor, the corresponding category label, and the image path are returned to facilitate subsequent training and classification analysis.

[0081] Step 7.2. In order to evaluate the performance of the model in the Alzheimer's disease recognition task, this study uses a five-fold cross-validation method to effectively evaluate the results of model training and verification. In this process, the ANDI dataset is divided using the StratifiedFold function to ensure that the category ratio in each fold remains consistent, thereby avoiding the interference of category imbalance on the model performance evaluation; in each fold, the training set and validation set are respectively encapsulated as datasets of Subset type, and the corresponding data loader is created with the help of torch.utils.data.DataLoader. Set shuffle=True in the training loader train_loader to disrupt the sample order and improve the generalization ability of the model; set shuffle=False in the verification loader test_loader to ensure the stability of the evaluation process;

[0082] Step 7.3: Set the experimental parameters for model training and validation. Define the training model function MSIN, set the number of classes to 2, the loss function to CrossEntropyLoss, and the optimizer to SGD (learning rate 0.001). Run the model training for a total of 100 epochs. Validation is performed after each round of training, recording the training loss and accuracy of the current round, as well as the prediction results and accuracy of the validation set. During the validation phase, the model is in evaluation mode, with gradient updates disabled using torch.no_grad() to save computing resources. The validation loss and accuracy of each round are recorded. If the validation accuracy of the current round is higher than the historical best, the "best validation accuracy" for that fold and its corresponding round are updated. Finally, the average accuracy (ACC) and area under the curve (AUC) for the five-fold validation are calculated and output to assess the stability and reliability of the overall model performance.

[0083] In Step 7.4, we used a cross-dataset validation method to verify the model's generalization ability. Generalization experiments were conducted using ADNI as the training set and AIBL and OASIS-2 as independent validation sets. The parameters used remained consistent with the main experiment. Finally, the model's accuracy (ACC) and area under the curve (AUC) were calculated and output to assess the generalization ability of the model's overall performance.

[0084] The present invention uses sMRI as input data, extracts multi-scale frequency information through a multi-frequency perception fusion module based on three-dimensional sym2 wavelet transform, and uses a graph hybrid attention module to flexibly model the relationship between channels and spatial position, thereby improving the model's ability to mine sMRI lesion details and generalization performance.

[0085] The technical solution and technical effects of the present invention are further described below in conjunction with specific experimental data.

[0086] In this experiment, the baseline model was configured as follows: the Graph Hybrid Attention (GHA) module in the Multiscale Spatial Information Network model was replaced with a Squeeze and Excitation (SE) module, and the Multi-Frequency Aware Fusion Module (MFAFM) was replaced with a standard convolution operation. Subsequently, the MFAFM and GHA modules were gradually introduced to explore the impact of each key module on model performance. The experiment was conducted on the ADNI dataset, with a binary classification task of Alzheimer's disease (AD) versus normal controls (NC). The experimental results are shown in Table 1. From the data in Table 1, it can be seen that after adding the multi-frequency perception fusion module (MFAFM) to the baseline model, the accuracy is improved to 92.13%±1.24, and the AUC is improved to 93.82%±0.98. This shows that the MFAFM module has a strong ability in extracting multi-frequency features and can effectively enhance the model's ability to recognize the brain structural features of Alzheimer's disease. After further introducing the graph hybrid attention module (GHA), the performance of the model is further improved, reaching an accuracy of 92.91%±1.03 and an AUC value of 94.42%±0.84, which are improved by 2.67% and 2.03% respectively compared with the baseline model, and the standard deviation is smaller, indicating that the GHA module proposed in this invention has significant advantages in integrating spatial structural information.

[0087] To validate the performance advantages and excellent generalization capabilities of the proposed Multi-Scale Spatial Information Network (MSIN) model in Alzheimer's disease classification, we conducted comparative experiments on three publicly available datasets: ADNI, AIBL, and OASIS. The model was compared with widely used classic 3D classification models (including 3D ResNet50, 3D DenseNet121, 3D EfficientNetB0, and 3D ViT). As shown in Table 2, across multiple datasets, the proposed model outperformed existing technologies in terms of accuracy and AUC, demonstrating stronger classification capabilities and cross-dataset generalization. This further validates the practical value and technological advancement of the proposed model in Alzheimer's disease diagnosis.

[0088] Table 1 Ablation experiment

[0089]

[0090] Table 2 Comparison experiments with other classic classification models

[0091]

[0092] Those skilled in the art will appreciate that the features described in the various embodiments and / or claims of this disclosure may be combined or coupled in various ways, even if such combinations or couplings are not explicitly described in this disclosure. In particular, the various embodiments of this disclosure may be combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations are intended to fall within the scope of this disclosure.

[0093] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the invention. Thus, the present invention is intended to include such changes and modifications as fall within the scope of the claims and their equivalents.

Claims

1. A multi-scale spatial information network classification method based on structural magnetic resonance imaging, characterized in that: The multi-scale spatial information network classification method is used in the Alzheimer's disease image processing process, and the method includes the following steps: Step 1: Obtain and preprocess sMRI images from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Australian Alzheimer's Neuroimaging, Biomarkers and Lifestyle Study (AIBL), and the Open Access Alzheimer's Disease Imaging Dataset (OASIS). Construct the whole-brain image dataset required for the classification model based on the preprocessed sMRI images. Step 2: Divide the whole-brain image dataset constructed in step 1 into a test set and a validation set; Step 3: Based on the test set and validation set divided in step 2, a multi-frequency perception fusion module (MFAFM) based on the three-dimensional sym2 wavelet transform is constructed to perform multi-scale spatial information processing in the frequency domain to capture features at different levels. Step 4: Based on the features of different levels captured in step 3, a graph hybrid attention mechanism module GHA is constructed. The graph hybrid attention mechanism module GHA is implemented through a dual attention coordination mechanism, which is implemented based on a graph channel attention mechanism and a spatial attention mechanism. Step 5: Use the inverted residual module GHA_MB in the graph hybrid attention mechanism module GHA constructed in step 4 as the backbone module of the multi-scale spatial information network model, and the backbone module is used to extract discriminative features; Step 6: Build a multi-scale spatial information network model, combine the multi-frequency perception fusion module MFAFM built in step 3, the graph hybrid attention mechanism module GHA built in step 4, and the inverted residual module GHA_MB built in step 5, and juxtapose the classifier to complete the final classification task; Step 7: The multi-scale spatial information network model is trained and validated on the Alzheimer's Disease Neuroimaging Initiative dataset (ADNI), and its generalization is verified on the AIBL Biomarker and Lifestyle Study dataset and the OASIS Open Access Alzheimer's Disease Imaging Dataset, ultimately obtaining an Alzheimer's disease classification model. Step 3.1, import subunits to build the libraries needed for the multi-frequency perception fusion module; Step 3.2: Define a function for converting wavelet transform coefficients to build a multi-frequency perception fusion module and extract wavelet coefficients: Step 3.3: Build a learnable scaling module to construct a multi-frequency perception fusion module for feature scaling: Step 3.4: Construct a multi-frequency perception fusion module to build a multi-scale spatial information network model and perform multi-scale processing in the frequency domain to capture features at different levels.

2. The multi-scale spatial information network classification method based on structural magnetic resonance imaging according to claim 1, characterized in that: The preprocessing described in step 1 includes AC-PC commissure and skull stripping, and spatial normalization to the template MIN152 operation.

3. The multi-scale spatial information network classification method based on structural magnetic resonance imaging according to claim 1, characterized in that: In step 2, the method for dividing the whole-brain image dataset constructed in step 1 into a test set and a validation set is: A five-fold cross-validation was performed on the Alzheimer's Disease Neuroimaging Project dataset (ADNI), with four copies used as test sets and one as validation set.

4. The multi-scale spatial information network classification method based on structural magnetic resonance imaging according to claim 1, characterized in that: In step 4, the hybrid attention mechanism module GHA is implemented through the dual attention coordination mechanism as follows: Graph channel attention mechanism, which is used to model inter-channel relationships through graph structures to emphasize information-rich features and suppress useless features; The spatial attention mechanism is used to quantify the importance of features at different spatial locations through a learnable spatial weight matrix and locate the specific spatial location of the lesion; The dual-attention collaborative mechanism is used to enable the multi-scale spatial information network model to simultaneously capture key information of the channel dimension and the spatial dimension, and the two work together to achieve adaptive enhancement of multi-scale features.

5. The multi-scale spatial information network classification method based on structural magnetic resonance imaging according to claim 1, characterized in that: Step 5 also includes steps of expanding, processing and compressing the input features.

6. The multi-scale spatial information network classification method based on structural magnetic resonance imaging according to claim 1, characterized in that: The size of the whole brain image dataset required to construct the classification model using the preprocessed sMRI images in step 1 is 128 128 128-pixel whole-brain image.

7. A computer device comprising a memory and a processor, characterized in that: A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images