Alzheimer prediction method based on Faster SNN neural network

Abstract

The invention discloses an Alzheimer prediction method based on a Faster SNN neural network. The Alzheimer prediction method based on the Faster SNN neural network comprises the following steps: constructing a Faster SNN neural network model; the Faster SNN neural network model comprises a multi-scale feature fusion module and a plurality of feature extraction modules used for extracting features of different sizes; the feature extraction module comprises an LIF model for sparse features and a pulse weighted attention model; after the input of the feature extraction module is subjected to sparse features of the LIF model, weighted adjustment is carried out through a pulse channel attention path and a pulse space attention path of the pulse weighted attention model, and a mixed feature map is output; the feature extraction modules are connected in sequence, and the output of the previous feature extraction module is used as the input of the next feature extraction module; the pulse weighted attention model also comprises an LIF model; and the multi-scale feature fusion module is used for fusing the mixed feature maps output by the feature extraction modules to generate a fused feature map. According to the invention, the recessive pathological features of the AD stage can be accurately captured, and the diagnosis accuracy of the MCI stage is improved.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision deep learning and medical image processing, and in particular to an Alzheimer's disease prediction method based on a FasterSNN neural network. Background Art

[0002] Alzheimer's disease (AD) is an irreversible condition characterized by persistent cognitive decline, behavioral abnormalities, and neurodegeneration. AD typically progresses from mild cognitive impairment (MCI) to severe dementia, ultimately leading to a terminal stage with complete loss of self-care. Prompt identification and intervention at the MCI stage can potentially slow disease progression and even reverse some cognitive impairment. Therefore, accurate early diagnosis of AD at the MCI stage is crucial for disease management.

[0003] Although clinical attention has been paid to the necessity of early diagnosis, there are still huge challenges in the accurate identification of AD, especially the MCI stage, in actual diagnosis and treatment. Among traditional medical diagnostic methods, the Alzheimer's Disease Assessment Scale-Cognitive Scale (ADAS-Cog) currently widely used in clinical practice can effectively evaluate multiple cognitive dimensions such as memory, language, and executive function, but its test items are limited, its sensitivity is insufficient, and it is highly dependent on subjective judgment, making it difficult to accurately distinguish the continuous pathological spectrum from MCI to AD. Advanced imaging technologies such as structural MRI, PET-CT, and PET / MR can effectively detect brain atrophy and molecular pathological changes, but their high cost and expert dependence greatly limit the breadth of clinical application. Compared with the high-cost PET, MRI, which is lower in cost and has stable image quality, has gradually become the mainstream choice.

[0004] In recent years, artificial intelligence technology, especially deep learning methods, has shown great promise in the field of AD-assisted diagnosis, and can efficiently and objectively capture key imaging features such as hippocampal atrophy and prefrontal lobe abnormalities. However, existing artificial neural network (ANN) models generally have problems with high power consumption and high computing resource requirements, which seriously limit their widespread application in primary healthcare and resource-constrained areas. In contrast, the bionic spiking neural network (SNN) simulates the pulse discharge mechanism of biological neurons and only triggers pulses when the membrane potential exceeds the threshold, thereby significantly reducing redundant calculations. However, the existing SNN model has deficiencies in expression ability and training stability, which limits its performance in complex medical tasks. Summary of the Invention

[0005] In view of the above-mentioned defects of the prior art, the present invention provides an Alzheimer's prediction method based on the FasterSNN neural network, which integrates the leaky integrate-and-fire neuron (LIF), the FasterNet model and the multi-scale feature fusion attention mechanism based on the pyramid structure. It can efficiently extract and integrate multi-dimensional feature information from MRI images, accurately capture the latent pathological characteristics of different AD stages, and thus significantly improve the diagnostic accuracy of AD, especially the MCI stage.

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

[0007] A method for predicting Alzheimer's disease based on a FasterSNN neural network comprises the following steps:

[0008] Step S1, constructing a FasterSNN neural network model;

[0009] The FasterSNN neural network model includes a multi-scale feature fusion module and a plurality of feature extraction modules for extracting features of different sizes;

[0010] The feature extraction module includes a LIF model and a pulse weighted attention model for sparse features;

[0011] The input of the feature extraction module is subjected to the sparse features of the LIF model and then weighted adjustment by the pulse channel attention path and the pulse spatial attention path of the pulse weighted attention model to output a mixed feature map;

[0012] The feature extraction modules are connected sequentially, with the output of the previous feature extraction module serving as the input of the next feature extraction module;

[0013] The pulse weighted attention model also includes the LIF model;

[0014] The multi-scale feature fusion module is used to fuse the mixed feature maps output by each of the feature extraction modules to generate a fused feature map for image classification;

[0015] Step S2: establishing a prediction data set, wherein the prediction data set includes a training set and a test set;

[0016] Step S3: using the training set to train the FasterSNN neural network model and perform parameter tuning;

[0017] Step S4: Use the test set to test the trained FasterSNN neural network model.

[0018] Preferably, in step S1, the feature extraction module performs a 3D convolution on the input feature map to obtain a single convolution feature map; then, the single convolution feature map is distinguished between the central area and the edge area through a spatial mask mechanism, and then the central area and the edge area are extracted using different convolution methods and then spliced ​​to generate a spliced ​​feature map; the spliced ​​feature map is then normalized to obtain a dual convolution feature map.

[0019] Preferably, in step S1, the dual convolution feature map is subjected to feature sparseness by the LIF model to generate a sparse dual convolution feature map; then the single convolution feature map is added as a residual to the sparse dual convolution feature map to form a preliminary processing feature map; the preliminary processing feature map is subjected to a GELU activation function and the pulse weighted attention model to generate a weighted attention feature map.

[0020] Preferably, in step S1, the feature extraction module uses standard 3×3×3 convolution to extract features for the central area and uses depthwise separable convolution to extract features for the edge area.

[0021] Preferably, in step S1, the pulse weighted attention model uses learnable parameters to control the weights of the pulse channel attention path and the pulse space attention path in the weighted attention feature map.

[0022] Preferably, there are four feature extraction modules, and the pulse weighted attention model only exists in the first three feature extraction modules.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. The main innovations of the present invention are:

[0025] A hybrid neural network architecture that combines the accuracy of ANN models with the energy efficiency of SNN models is proposed to overcome the shortcomings of existing methods in actual clinical applications. In the FasterSNN neural network model of this invention, a multi-scale feature fusion module and a feature extraction module work together with the LIF model and the pulse weighted attention model. This not only retains the ANN's ability to extract corresponding features, but also avoids the problems of large parameter counts, long training time, and high computing resource requirements. It also has the low energy consumption and low parameter count characteristics of SNN, avoiding the problems of poor training stability and convergence of SNN models and the need for complex training strategies and techniques. At the same time, it effectively compensates for the huge challenges faced by the deployment of computationally intensive ANN models in diagnosis, especially in primary care settings. At the same time, it provides a way to improve the reliability and ease of use of SNN models in actual clinical applications. The model first constructs a basic computing unit using LIF neurons and adopts a membrane potential update mechanism to achieve biologically interpretable sparse computation. After the input 3D MRI data passes through the initial convolutional layer, it enters four feature extraction blocks composed of the Faster SNN Block as the backbone network layer. This module uses a region-adaptive convolution strategy, applying different convolution methods to different regions of the input feature map, and finally concatenating the features extracted by the different convolutions. A SWA model is added to each Faster SNN Block, dynamically weighting the trainable parameters corresponding to the dual pathways of pulse channel attention and pulse spatial attention. The entire network uses the MSF module to construct a four-level pyramid architecture, achieving feature reuse through cross-layer connections and using learnable parameters to perform a weighted summation of each layer's output to achieve feature aggregation. Finally, the fused feature map is temporally averaged and input into the classifier for accurate classification.

[0026] 2. The main innovations of the present invention are:

[0027] The Alzheimer's prediction method (FasterSNN) proposed in the present invention only needs to input MRI data to realize automatic and intelligent Alzheimer's prediction function. The present invention makes it possible to establish an Alzheimer's prediction model and assist doctors in diagnosis and treatment, which is of great significance in practical applications. In practical applications, although the PET modality has high specificity, its examination cost is high, which significantly increases the economic burden on patients and families, and the diagnostic process still requires the participation of experienced experts; the cognitive scale is economical and practical, and simple to operate, but the patient's educational background, psychological state and regional differences (such as high illiteracy rates in remote areas) will greatly affect the consistency and reliability of the scale assessment. Therefore, the present invention uses relatively moderate MRI imaging technology, which not only maintains the objectivity of medical images, but also has obvious advantages in cost and equipment maintenance. At the same time, the present invention avoids subjective interference and waste of resources in multimodal fusion, and is committed to developing a more accurate, economical and easy-to-promote diagnostic model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is the overall structure diagram of the FasterSNN neural network model of Example 1 of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are merely descriptive and not restrictive, and the scope of protection of the present invention cannot be limited by them.

[0030] Example 1:

[0031] like Figure 1 The Alzheimer's prediction method based on the FasterSNN neural network shown in the embodiment includes the following steps:

[0032] Step S1: Construct a FasterSNN neural network model for Alzheimer's disease prediction. The FasterSNN neural network model includes: a feature extraction module (Faster SNN Block, hereinafter referred to as "Block") and a multi-scale feature fusion module (Multi-Scale Feature Fusion, hereinafter referred to as "MSF module").

[0033] There are multiple feature extraction modules, and there are 4 feature extraction modules in this embodiment. Among them, the 1st to 3rd blocks contain the Leaky Integrate-and-Fire model (hereinafter referred to as the "LIF model") and the Spiking Weighted Attention model (hereinafter referred to as the "SWA model"). The 4th block only contains the LIF model. The feature extraction module is the backbone network of the entire model, which is used to extract features of different sizes of the original MR image. Each block is connected in sequence, and the output of the previous block serves as the input of the next block. That is, the input of the first block is the original MR image, and the inputs of the second, third, and fourth blocks are the feature maps output by the previous block respectively.

[0034] The output of the block is a mixed feature map. This module adopts a regional adaptive convolution strategy, that is, a 3D convolution is performed on the original MR image or the feature map output by the previous block to obtain a single convolution feature map, and then different convolution strategies are used for different regions of the single convolution feature map, and then normalized to obtain a double convolution feature map, and further passed through the LIF model to obtain a sparse double convolution feature map. The single convolution feature map is then added as a residual to the sparse double convolution feature map to form a preliminary processing feature map, and then the preliminary processing feature map is passed through the GELU activation function and the SWA model to obtain a weighted attention feature map. Figure 1 As shown in the upper left corner, the block divides a single convolutional feature map into a central region and an edge region, and processes each region using different convolution strategies: a standard 3×3×3 convolution is used in the central region to extract complete feature information, while a depthwise separable convolution is used in the edge region to reduce computational overhead. This effectively balances expressive power and computational overhead while maintaining the receptive field.

[0035] In addition, the feature extraction module introduces a spatial masking mechanism, using the mask matrix M to distinguish between the center and edge regions to ensure the continuity and consistency of the spliced ​​feature map in the spatial dimension. Thanks to this design, the feature extraction module can significantly improve computational efficiency while maintaining high accuracy, providing high-quality feature representation for subsequent multi-scale fusion and classification. The specific formula is as follows:

[0036] F center =Conv 3×3×3 (X[:,:,h s :h e ,w s :w e ,d s :d e ])

[0037] F edge =Conv 1×1×1 (Conv depthwise (X)

[0038]

[0039] F out =M⊙F center +M⊙F edge

[0040] Y=X+LIF(BN(F out ))

[0041] Among them, h s :h e ,w s :w e ,d s :de Respectively represent the start index and end index of h, w, d, F center Represents the central feature. Conv depthwise represents depth-wise separable convolution, F edge Represents edge features. M represents the mask matrix, ⊙ represents element-by-element multiplication, F out Represents the preliminary processing of the input feature map.

[0042] The LIF model is a common method of the SNN model, and it is also the key to the low-energy performance of the model of the present invention. However, the LIF model loses a large amount of feature information. In order to reduce the loss of features, the present invention introduces the Faster SNN Block as the backbone network, and uses the MSF module to fuse features of different scales together. At the same time, the SWA model is used to capture the sparse output features obtained by the LIF model. The LIF model is also provided in the SWA model. The LIF model is used to output the dense features of the input as sparse features to reduce high-energy floating-point operations. In the main path of the Faster SNN Block, its input is a dual convolution feature map.

[0043] The FasterSNN neural network model is constructed by LIF neurons as the basic computing unit, and uses the membrane potential update mechanism to achieve biologically explainable sparse computing. The specific addition locations are as follows: Figure 1 As shown in . At each moment, LIF neurons update the membrane potential by attenuating the membrane potential of the previous moment by the attenuation coefficient and superimposing it with the current input. When the membrane potential exceeds the set threshold, the neuron triggers a pulse and resets the membrane potential. This mechanism not only makes FasterSNN closer to the processing method of the biological nervous system, but also improves the expression ability and computational efficiency of complex 3D MRI. In addition, the sparsity of LIF neurons also helps the model focus on key features in the data, reduces the risk of overfitting, and provides stable support for subsequent feature extraction and classification modules. The membrane potential update and pulse emission mechanism of LIF neurons can be expressed as follows:

[0044]

[0045] Where V t represents the membrane potential at time t, T is the time step (the default value is 2), λ controls the decay rate of the membrane potential in the previous time step (the initial value is 0.9), V th is the pulse emission threshold (initial value is 1), S t It is in pulse emission state.

[0046] The SWA model is used to extract sparse features from the spatial and channel dimensions. Its input is the activation-processed feature map after GELU activation in each block, and its output is a weighted attention feature map that combines the spatial and channel dimensions. Dynamic weighting is achieved through the trainable parameters corresponding to the dual pathways of Spike Channel Attention (CA) and Spike Spatial Attention (SA). Combining the temporal sparsity of spiking neurons with the dynamic modeling capabilities of the attention mechanism, the activation-processed feature map is finely modeled and weighted through the parallel Spike Channel Attention and Spike Spatial Attention structures, enhancing the network's perception of key lesion areas. The SWA model only exists in the first three blocks. The feature maps obtained from the first three blocks have already undergone multiple sparse and weighted operations. If the fourth Faster SNN block performs attention weighting, the features will become too sparse, which is not conducive to final classification.

[0047] The LIF model in the SWA model uses a convolutional half-space feature map and a global average pooling and convolutional half-channel feature map as input in the spatial and channel dimensions, respectively. Specifically, in the spiking channel attention branch, the SWA model first performs global average pooling (GAP) on the activation feature map to generate a feature representation of size (B, C, 1, 1, 1) to aggregate spatial information. Subsequently, a 1×1×1 convolution is used to compress the number of channels to 1 / 4 of the original to reduce computational cost. The LIF model is then introduced for spike discretization to enhance information sparsity. Next, a 1×1×1 convolution is performed to restore the number of channels to their original dimensions. A sigmoid activation function is then applied to output channel attention weights, which are then element-wise multiplied with the activation feature map to produce the channel attention map. Similarly, the spiking spatial attention branch performs similar operations on the activation feature map to produce the spatial attention map. In order to dynamically balance the effects of the two attention mechanisms, the SWA model introduces two learnable parameters to control the proportion of the two attention mechanisms in the final fusion, thereby achieving adaptive adjustment of the dual-pathway fusion. The specific formula is as follows:

[0048] W c =σ(LIF(Cov(GAP(x))))

[0049] W c =σ(LIF(Cov(Cov(x))))

[0050] X out =X⊙(αW c +βW s )

[0051] Among them, σ represents the Sigmoid activation function, W c represents the channel attention weight, W sRepresents the spatial attention weight, Cov represents the convolution operation, α and β represent the channel attention weight and the spatial attention weight, the initial value is 0.5, and ⊙ represents element-by-element multiplication.

[0052] The MSF module is used to fuse the mixed feature maps output by each block. Its input is the mixed feature map output by each block, and the output result is the final fused feature map that finally fuses multiple mixed feature maps of different scales.

[0053] The MSF module constructs a four-level pyramid architecture, enabling feature reuse through cross-layer connections and performing a weighted summation of each layer's output using learnable parameters to achieve feature aggregation. Finally, the fused features are temporally averaged and fed into a classifier for accurate classification. Specifically, the MSF module constructs a four-level pyramid structure, expanding from 64 channels to 128, 256, and finally 512 channels. Each layer's output is resized to 64 channels via 1×1 convolution to produce a fused feature map. The fused feature maps from the first three layers are then downsampled to the same size using MaxPooling and then added to the feature maps from the fourth layer for spatially aligned fusion. To further enhance cross-layer fusion capabilities, MSF integrates multi-scale features using a weighted summation approach. Specifically, after convolution and pooling, the fused feature maps of each of the first three layers are multiplied by learnable weights to dynamically adjust the contribution between layers. These weights are automatically optimized during training to fully leverage the discriminative power of different layers. Finally, the model performs a temporal average operation on the fused feature maps of each time step to obtain the final fused feature map, and completes the classification of the image through a flattening layer and a linear layer activated by softmax. The specific formula is as follows:

[0054]

[0055] Among them, F fusion represents the generated feature image, F final represents the final generated feature map, T represents the time step set by the present invention, the default value is 4, w i Represents the feature map weight of the i-th layer, with an initial value of 1.0.

[0056] Step S2: Use MRI image data to establish a multi-stage Alzheimer's prediction dataset, which is divided into a training set and a test set. The prediction dataset has two datasets: the ADNI dataset and the AIBL dataset.

[0057] The ADNI dataset combines two Alzheimer's Disease Neuroimaging Initiative (ADNI) data sources to construct a comprehensive neuroimaging dataset encompassing three categories: Alzheimer's disease (AD), normal cognition (CN), and mild cognitive impairment (MCI). A portion of the ADNI dataset was derived from the finaldata dataset publicly available on the Kaggle platform. This dataset has been standardized from DICOM to NIfTI format, with a spatial resolution of 256×256×166. It contains 426 3D images, including 168 AD patients and 254 CN controls. The other portion of the ADIN dataset was obtained directly from the ADNI platform by our research team, focusing on the original DICOM files of MCI cases. An automated processing pipeline based on the pydicom and nibabel libraries was developed to achieve precise patient ID matching and instance number sequence sorting. The 3D image data were normalized to NIfTI format with a spatial resolution of 256×256×92, ultimately selecting 283 high-quality images for inclusion in the study. The entire dataset was stratified using an 8:2 ratio for training and testing. The complete data organization structure is shown in Table 1.

[0058] Table ADNI dataset distribution

[0059]

[0060] The AIBL dataset, obtained directly from the Australian Imaging, Biomarker and Lifestyle (AIBL) dataset by our research team, includes three comprehensive neuroimaging datasets: Alzheimer's disease (AD), normal cognition (CN), and mild cognitive impairment (MCI). This dataset has been standardized from DICOM to NIfTI format, with a spatial resolution of 160×240×256. It contains 682 3D images, including 219 AD patients, 200 CN controls, and 263 MCI controls. The dataset was stratified and sampled using a 7:3 ratio to construct training and test sets. The complete data organization is shown in Table 2.

[0061] Table 2 Distribution of AIBL dataset

[0062]

[0063] Step S3, use the training set to train the FasterSNN neural network model, and optimize the prediction classification effect by tuning the parameters of the FasterSNN neural network model. Specifically, the FasterSNN neural network model is trained using the training set, and the deep learning framework PyTorch2.4.1 is uniformly adopted, and is carried out on a hardware platform equipped with an NVIDIA GeForceRTX 4090 graphics card and CUDA 11.8. In order to optimize training efficiency, the present invention particularly adopts a mixed precision training strategy. In terms of data processing, the batch size (batch size) of the training set and the test set is set to 16. In the training phase, a dynamic random shuffling strategy is adopted to enhance the generalization ability of the model, and a 4-thread parallel loader is configured to accelerate data preprocessing efficiency. In terms of model optimization, the present invention selects the Adam algorithm as the optimizer, and sets the initial learning rate to 1e-3. At the same time, in order to achieve effective model regularization, a weight decay coefficient of 1e-3 is configured. In addition, the present invention adopts a dynamic learning rate decay strategy, ReduceLROnPlateau. When the validation set accuracy does not improve within three consecutive epochs, the learning rate is halved to avoid falling into a local optimum. A global gradient clipping strategy is also used to ensure the stability of the training process. For the LIF model, the present invention sets the initial time step to 2, the spike threshold to 1, and the decay coefficient to 0.9 to simulate the dynamic characteristics of biological neurons.

[0064] Step S4, use the test set to test the network model obtained by evaluation, and finally realize the automatic and intelligent Alzheimer's prediction function. Specifically, the network model obtained by testing the test set is used to finally realize the automatic and intelligent Alzheimer's prediction function. The present invention adopts a multidimensional evaluation system to quantitatively analyze the model from the two dimensions of classification performance and computational efficiency, and complementarily evaluates the overall performance of the model. For classification performance, this study adopts commonly used classification indicators such as accuracy, precision, recall and F1 score (imbalanced data set) as well as Kappa coefficient and average AUC value as classification performance indicators.

[0065]

[0066] Among them, TP refers to the number of samples correctly predicted as positive, TN refers to the number of samples correctly predicted as negative, FP refers to the number of samples incorrectly predicted as positive, and FN refers to the number of samples incorrectly predicted as negative. o is the prediction accuracy, P e Indicates accidental consistency; n i+is the total number (row sum) of the true labels for the i-th category, n +i is the total number (column sum) of the predicted labels for the i-th class. AUC i represents the AUC value of the i-th category.

[0067] Regarding computational efficiency, this study introduced a multi-dimensional evaluation system including energy consumption, parameter scale complexity, and training time cost to comprehensively measure the resource utilization efficiency of the model.

[0068] Energy consumption:

[0069]

[0070] e spike =77×10 -15 J

[0071] e MAC =12.5×10 -12 J

[0072] MAC s =C in ×C out ×k 3 ×W out ×H out ×D out

[0073] Energy=(MACs×e mac ×(1-Sparsity))+Spikes×e spike

[0074] Among them, e spike and e MAC They represent the energy consumed by a single pulse and a single multiplication-addition operation respectively.

[0075] Parameter calculation:

[0076] The total number of model parameters is obtained by traversing the trainable parameters (including weights and biases) of all layers of the network and accumulating them, and finally converting the result into millions of units (M) output. The calculation formula is:

[0077]

[0078] Where L represents the total number of network layers, W l shape is the dimension product of the weight parameters of the lth layer, is the dimension product of the bias parameters of the lth layer.

[0079] Time calculation:

[0080] In order to measure the training time of each model, the present invention adds a timestamp as the training starting point before starting a training process, and obtains the timestamp again after the training is completed. The difference between the two is the training time of a single epoch.

[0081] Time train =Time end -Time start

[0082] Among them, Time start The timestamp of the training cycle start, Time end The timestamp of when training is completed.

[0083] In terms of model architecture, traditional ANN models demonstrate high accuracy in experiments but still have significant limitations in practical applications. First, the use of two-dimensional slices results in a significant loss of spatial information, limiting the model's ability to comprehensively analyze complex brain structures in real-world clinical settings. Second, while models using full three-dimensional data offer some improvements in accuracy, they are often accompanied by issues such as a large number of parameters, prolonged training time, and high computational resource requirements, severely limiting their application and widespread adoption in areas with limited medical resources. Deployment of such computationally intensive models is particularly challenging in primary care settings. Current SNN approaches still face numerous challenges in their application to AD diagnosis. First, SNNs have limited expressive power. When processing complex medical data such as MRI, their feature extraction and generalization capabilities struggle to match the performance of traditional CNN and Transformer models. Second, SNN models suffer from poor training stability and convergence, requiring complex training strategies and techniques, which impacts their reliability and ease of use in real-world clinical applications. Furthermore, research on SNN models for medical imaging, particularly three-dimensional MRI data, is limited, and mature and stable model architectures are still lacking. Therefore, the present invention aims to combine the energy efficiency advantages of SNNs with the expressive power of traditional neural networks, proposing a hybrid neural network architecture that combines accuracy and energy efficiency to overcome the shortcomings of existing methods in practical clinical applications. The model first constructs basic computational units using LIF neurons and employs a membrane potential update mechanism to achieve biologically interpretable sparse computation. After passing the input 3D MRI data through an initial convolutional layer, it enters four feature extraction blocks, consisting of a backbone network layer composed of Faster SNN blocks. This module employs a region-adaptive convolution strategy, applying different convolution methods to different regions of the input feature map and concatenating the features extracted from these different convolutions. A SWA model is incorporated into each Faster SNN block, dynamically weighting the features using trainable parameters corresponding to the two pathways of spike channel attention and spike spatial attention. The entire network utilizes a four-level pyramid architecture constructed using the MSF module, enabling feature reuse through cross-layer connections and weighted summation of each layer's output using learnable parameters to achieve feature aggregation. Finally, the fused features are temporally averaged and input into a classifier for accurate classification.

Claims

1. A method for predicting Alzheimer's disease based on FasterSNN neural network, characterized in that: The following steps are involved: Step S1, constructing a FasterSNN neural network model; The FasterSNN neural network model includes a multi-scale feature fusion module and a plurality of feature extraction modules for extracting features of different sizes; The feature extraction module includes a LIF model and a pulse weighted attention model for sparse features; The input of the feature extraction module is subjected to the sparse features of the LIF model and then weighted adjustment by the pulse channel attention path and the pulse spatial attention path of the pulse weighted attention model to output a mixed feature map; The feature extraction modules are connected sequentially, with the output of the previous feature extraction module serving as the input of the next feature extraction module; The pulse weighted attention model also includes the LIF model; The multi-scale feature fusion module is used to fuse the mixed feature maps output by each of the feature extraction modules to generate a fused feature map for image classification; Step S2: establishing a prediction data set, wherein the prediction data set includes a training set and a test set; Step S3: using the training set to train the FasterSNN neural network model and perform parameter tuning; Step S4: Use the test set to test the trained FasterSNN neural network model.

2. The Alzheimer's prediction method based on the FasterSNN neural network according to claim 1, characterized in that: In step S1, the feature extraction module performs a 3D convolution on the input feature map to obtain a single convolution feature map; then, the single convolution feature map is distinguished between the central area and the edge area through a spatial mask mechanism, and then the central area and the edge area are extracted using different convolution methods and then spliced ​​to generate a spliced ​​feature map; the spliced ​​feature map is then normalized to obtain a dual convolution feature map.

3. The Alzheimer's prediction method based on FasterSNN neural network according to claim 2, characterized in that: In step S1, the dual convolution feature map is subjected to feature sparseness by the LIF model to generate a sparse dual convolution feature map; then the single convolution feature map is added as a residual to the sparse dual convolution feature map to form a preliminary processing feature map; the preliminary processing feature map is subjected to a GELU activation function and the pulse weighted attention model to generate a weighted attention feature map.

4. The Alzheimer's prediction method based on FasterSNN neural network according to claim 2, characterized in that: In step S1, the feature extraction module uses standard 3×3×3 convolution to extract features for the central area and uses depthwise separable convolution to extract features for the edge area.

5. The Alzheimer's prediction method based on FasterSNN neural network according to claim 1, characterized in that: In step S1, the pulse weighted attention model uses learnable parameters to control the weights of the pulse channel attention path and the pulse space attention path in the weighted attention feature map.

6. The Alzheimer's prediction method based on FasterSNN neural network according to claim 1, characterized in that: There are four feature extraction modules, and the pulse weighted attention model only exists in the first three feature extraction modules.

Images