An alzheimer's disease classification method and system based on dynamic cross-flow fusion double-flow neural network, a computer device and a storage medium

Abstract

The application discloses a double-flow neural network Alzheimer's disease classification method based on dynamic cross-flow fusion and relates to the technical field of medical image processing, and comprises the following steps: acquiring a brain medical image, inputting the brain medical image into a pre-trained Alzheimer's disease classification model for classification and outputting an Alzheimer's disease classification result; the application constructs two parallel global feature flow branches and local feature flow branches, the global feature flow branches are used for extracting global middle-layer features and global deep-layer features of the overall brain structure, and the local feature flow branches are used for extracting local middle-layer features and local deep-layer features of brain regions such as hippocampus and temporal lobe; the dynamic cross-flow interaction module is used for realizing feature interaction and supplement of the global features and the local features at two levels of the middle layer and the deep layer, completing deep fusion of the global structure features and the local lesion features of the brain, and significantly improving the classification precision of the Alzheimer's disease classification model.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method, system, computer device, and storage medium for classifying Alzheimer's disease using a dual-stream neural network based on dynamic cross-stream fusion. Background Technology

[0002] Alzheimer's disease is a progressive neurodegenerative disease that poses a significant threat to the health of middle-aged and elderly people and is difficult to diagnose in its early stages. With the rapid development of artificial intelligence technology, deep learning-based medical image analysis technology has provided a new approach to the diagnosis of Alzheimer's disease. By extracting features of brain atrophy and lesions in specific brain regions and performing automated analysis to derive a diagnosis of Alzheimer's disease, this technology can provide important support for clinical diagnosis and treatment intervention. Currently, deep learning models such as convolutional neural networks have shown great application potential in the diagnosis of Alzheimer's disease.

[0003] However, existing Alzheimer's disease classification models still have some shortcomings in practical applications. For example, conventional single-stream network models such as ResNet mostly only extract global brain structural features from brain medical images, making it difficult to accurately capture local lesion features in key brain regions such as the hippocampus and temporal lobe. Single-stream network models that focus on local regions often lack contextual information on global brain structural features and the degree of atrophy, resulting in low classification accuracy.

[0004] To improve classification accuracy, some studies have proposed dual-branch networks to extract global and local features from brain medical images. However, the fusion mechanisms used are mostly simple feature splicing or static fusion methods with equal weights, which cannot dynamically adjust feature weights according to the lesion distribution of different samples. This results in a lack of feature interaction between global and local features, which can easily lead to the loss of key diagnostic features. It is difficult to give full play to the advantages of dual-branch networks and has limited effect on improving the classification accuracy of Alzheimer's disease. Summary of the Invention

[0005] To address the shortcomings of existing network models in Alzheimer's disease classification with low accuracy, this invention provides a method, system, computer device, and storage medium for Alzheimer's disease classification based on a dual-stream neural network using dynamic cross-stream fusion.

[0006] To achieve the above objectives, the specific solution adopted by this invention is as follows: a dual-stream neural network Alzheimer's disease classification method based on dynamic cross-stream fusion, comprising the following steps: Acquiring medical images of the brain; The Alzheimer's disease classification model inputs brain medical images into a pre-trained model for classification and outputs the classification results. The model includes a two-stream feature extraction module, a dynamic cross-stream interaction module, a weighted fusion module, and a classification module. The two-stream feature extraction module comprises two parallel global feature stream branches and a local feature stream branch. The global feature stream branch includes a global backbone network based on ResNet, and the local feature stream branch includes a local backbone network based on EfficientNet. The dynamic cross-stream interaction module includes a mid-level cross-stream attention submodule and a deep dynamic cross-attention submodule. The Alzheimer's disease classification model classifies brain medical images using the following methods: Global backbone networks are used to extract global brain structural features from medical brain images to obtain global mid-level features and global deep features. Local backbone networks are used to extract local brain structural features from medical brain images to obtain local mid-level features and local deep features. The intermediate layer cross-flow attention submodule is used to fuse global intermediate features and local intermediate features to obtain intermediate fused features. The intermediate fused features are then enhanced across layers to global deep features and local deep features to obtain global enhanced deep features and local enhanced deep features, respectively. The deep dynamic cross-attention submodule is used to update the global enhanced deep features and the aggregated local enhanced deep features to obtain the global updated features and the local updated features; The weighted fusion module is used to perform weighted fusion of global update features and local update features to obtain the final fused features; The classification module outputs Alzheimer's disease classification results based on the final fused features.

[0007] As an optimization of the aforementioned Alzheimer's disease classification method based on dynamic cross-stream fusion dual-stream neural networks, the method for extracting global brain structural features using a global backbone network to obtain global mid-level features and global deep features includes: The preprocessed brain medical images were used as global input data for the global backbone network; The first three layers of the global backbone network are used to extract features from the global input data to obtain global mid-level features. The global mid-level features are then propagated forward to the fourth layer of the global backbone network to obtain global preliminary deep features. The global preliminary deep features are then weighted to obtain global deep features.

[0008] As another optimization scheme for the above-mentioned dual-stream neural network Alzheimer's disease classification method based on dynamic cross-stream fusion: the output end of the global backbone network is provided with a global attention submodule, which includes a channel attention unit and a spatial attention unit. The channel attention unit is used to perform channel weighting on the channels of the global preliminary deep features, and the spatial attention unit is used to perform spatial weighting on the channel-weighted global preliminary deep features to obtain global deep features.

[0009] As an alternative optimization of the aforementioned Alzheimer's disease classification method based on dynamic cross-stream fusion dual-stream neural networks, the method for extracting local brain structural features using local backbone networks to obtain local mid-level and local deep features includes: Brain medical images are cropped according to preset cropping rules to obtain central region image blocks, left temporal lobe region image blocks and right temporal lobe region image blocks. The central region image blocks, left temporal lobe region image blocks and right temporal lobe region image blocks are merged in the batch dimension and used as local input data for the local backbone network. The first five or six stages of the local backbone network are used to extract features from the local input data to obtain local mid-level features. The local mid-level features are then propagated forward to the ninth stage of the local backbone network to obtain local preliminary deep features. The local preliminary deep features are then weighted to obtain local deep features.

[0010] As another optimization scheme for the above-mentioned dual-stream neural network Alzheimer's disease classification method based on dynamic cross-stream fusion: the output end of the local backbone network is equipped with a coordinate attention submodule, which is used to spatially weight the local preliminary deep features to obtain local deep features.

[0011] As an alternative optimization of the aforementioned Alzheimer's disease classification method based on dynamic cross-flow fusion of two-stream neural networks, the method for obtaining the mid-level fused features by weighted fusion of global mid-level features and local mid-level features using the intermediate layer cross-flow attention submodule includes: The intermediate layer cross-stream attention submodule includes an attention generation layer and an intermediate weighted fusion layer. The attention generation layer includes a multilayer perceptron and a Softmax activation function set sequentially. After mapping the global mid-level features and local mid-level features to the same dimension, the feature residual information is extracted. The global mid-level features and local mid-level features after extracting the feature residual information are then concatenated to obtain intermediate data. The intermediate data is processed by the attention generation layer to obtain the global initial attention weight and the local initial attention weight. The global initial attention weight and the local initial attention weight are normalized to obtain the global final attention weight and the local final attention weight respectively. The intermediate weighted fusion layer is used to perform weighted fusion of the global final attention weight and the local final attention weight with the global intermediate features and the local intermediate features, respectively, to obtain the intermediate fused features.

[0012] As another optimization scheme for the above-mentioned Alzheimer's disease classification method based on dynamic cross-stream fusion dual-stream neural network: the method of using a deep dynamic cross-attention submodule to update the globally enhanced deep features and the aggregated locally enhanced deep features to obtain the globally updated features and locally updated features includes: The local feature flow branch includes an image patch aggregation submodule for aggregating local enhanced deep features. The image patch aggregation submodule includes a local multi-head self-attention unit and an image patch gating network. The local enhanced deep features are input into the local multi-head self-attention unit for fusion to obtain local fused deep features. A role embedding vector is introduced into the local fused deep features. The role embedding vector is added to the local fused deep features and then input into the image patch gating network to obtain the dynamic weight score of each local enhanced deep feature. The local enhanced deep features and the corresponding dynamic weight scores are weighted and summed to obtain the local aggregated deep features. The deep dynamic cross-attention submodule includes a first multi-head attention submodule and a second multi-head attention submodule. The first multi-head attention submodule takes locally aggregated deep features as the query vector and globally enhanced deep features as the key vector and value vector, and inputs them to obtain locally updated features. The second multi-head attention submodule takes globally enhanced deep features as the query vector and locally updated features as the key vector and value vector, and inputs them to obtain globally updated features.

[0013] A dual-stream neural network Alzheimer's disease classification system based on dynamic cross-stream fusion includes: A data processing device is used to classify brain medical images using a pre-trained Alzheimer's disease classification model and output Alzheimer's disease classification results; The data visualization module is used to visually represent the classification results of Alzheimer's disease.

[0014] Computer equipment, including: Memory, used to store computer programs; The processor is used to read and execute computer programs to implement the aforementioned Alzheimer's disease classification method based on a dynamic cross-stream fusion dual-stream neural network.

[0015] A storage medium for storing a computer program that, when executed, implements the aforementioned two-stream neural network Alzheimer's disease classification method based on dynamic cross-stream fusion.

[0016] Compared with existing technologies, the present invention has the following beneficial effects: The present invention constructs two parallel global feature flow branches and local feature flow branches. The global feature flow branch is used to extract global mid-level features and global deep features of the overall brain structure, while the local feature flow branch is used to extract local mid-level features and local deep features of brain regions such as the hippocampus and temporal lobe. Through a dynamic cross-flow interaction module, global features and local features can interact and complement each other at the mid-level and deep-level, achieving deep fusion of global brain structural features and local lesion features, which significantly improves the classification accuracy of the Alzheimer's disease classification model. Attached Figure Description

[0017] Figure 1 This is a flowchart of the Alzheimer's disease classification method of the present invention; Figure 2 This is a schematic diagram of the Alzheimer's disease classification model of the present invention; Figure 3 This is a schematic diagram of the cropping of brain medical images according to the present invention; Figure 4 This is a schematic diagram of the structure of the local feature flow branch of the present invention; Figure 5 This is a schematic diagram of the structure of the dynamic cross-stream interaction module of the present invention; Figure 6 This is a flowchart of the three-stage training process of this invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. Parts not described or disclosed in detail in the following embodiments of the present invention should be understood as prior art known or should be known by those skilled in the art.

[0019] like Figures 1 to 6 As shown, a dual-stream neural network-based Alzheimer's disease classification method based on dynamic cross-stream fusion includes S1 to S2.

[0020] S1. Obtain brain medical images. In this embodiment, the brain medical images are grayscale images obtained using magnetic resonance imaging.

[0021] S2. Brain medical images are input into a pre-trained Alzheimer's disease classification model for classification, and the model outputs Alzheimer's disease classification results. These results include three categories: Alzheimer's disease, mild cognitive impairment, and normal cognition. The Alzheimer's disease classification model includes a two-stream feature extraction module, a dynamic cross-stream interaction module, a weighted fusion module, and a classification module. The two-stream feature extraction module includes two parallel global feature stream branches and a local feature stream branch. The global feature stream branch includes a global backbone network based on the ResNet network, specifically using the ResNet101 network, which includes 5 layers. The local feature stream branch includes a local backbone network based on the EfficientNet network, specifically using the EfficientNet-B0 network, which includes 9 stage blocks. The dynamic cross-stream interaction module includes a mid-level cross-stream attention submodule and a deep dynamic cross-attention submodule. The methods used by the Alzheimer's disease classification model to classify brain medical images include S21 to S25.

[0022] S21. A global backbone network is used to extract global brain structural features from medical brain images to obtain global mid-level and global deep features. Specifically: the number of input channels in the first convolutional layer of the global backbone network is modified to 1 to adapt to the input of grayscale magnetic resonance imaging images. The preprocessed medical brain image is used as the global input data of the global backbone network. The first three layers of the global backbone network are used to extract features from the global input data to obtain global mid-level features. The global mid-level features are then propagated forward to the fourth layer of the global backbone network to obtain preliminary global deep features. The preliminary global deep features are weighted using a global attention submodule set at the output of the global backbone network to obtain global deep features.

[0023] More specifically, the preprocessing of brain medical images includes uniform size adjustment and pixel normalization. The global attention submodule includes channel attention units and spatial attention units. The channel attention unit is used to perform channel weighting on the preliminary global deep features, and the spatial attention unit is used to perform spatial weighting on the channel-weighted preliminary global deep features to obtain the global deep features. The channel attention unit performs global average pooling of the preliminary global deep features in the spatial dimension and compresses them to obtain channel context vectors. Then, it is processed through two convolutional layers with a kernel size of 1×1 and a modified linear unit activation function. Finally, the weight coefficients of each channel are generated by the sigmoid activation function. The preliminary global deep features are multiplied by the weight coefficients channel by channel to complete the channel weighting of the preliminary global deep features. The spatial attention unit performs average pooling and max pooling on the channel-weighted global preliminary deep features. The two pooled features are concatenated along the channel dimension and then spatially convolved using a 7×7 large convolution kernel. The spatial weight coefficients are then output via Sigmoid and applied to the channel-weighted global preliminary deep features to obtain the global deep features.

[0024] A local backbone network is used to extract local brain structural features from medical brain images, yielding local mid-level and local deep features. Specifically, the number of input channels in the first convolutional layer of the local backbone network is modified to 1 to accommodate the input of grayscale MRI images. The medical brain image is then cropped according to a preset cropping rule to obtain central region, left temporal lobe, and right temporal lobe image blocks. These blocks are then merged in batches as local input data for the local backbone network. The first five or six stages of the local backbone network are used to extract features from the local input data, yielding local mid-level features. These mid-level features are then propagated forward to the ninth stage of the local backbone network to obtain preliminary local deep features. Finally, a coordinate attention submodule at the output of the local backbone network is used to weight these preliminary deep features to obtain the final local deep features.

[0025] More specifically, such as Figure 3As shown, the preset cropping rules are as follows: A baseline size is determined, which is one-third of the smaller of the width and height in the brain medical image. A lower limit threshold for pixels is also set; in this embodiment, the lower limit threshold is 32 pixels. Cropping of the central region image block: Using the geometric center of the brain medical image as the reference point, a square image block with both height and width of the baseline size is cropped. The central region image block is used to retain information about central structures such as the ventricles. Cropping of left and right temporal lobe image patches: To fully cover the hippocampus region, the left and right temporal lobe image patches are rectangular in size with a height greater than their width. The width is the baseline size, and the height is a preset multiple of the baseline size. In this embodiment, the preset multiple is 1.4 times. The cropping centers of the left and right temporal lobe image patches are offset downwards in the height direction from the center of the central region image patch by 12% of the height of the brain medical image. The cropping center of the left temporal lobe image patch is located at 1 / 4 or 1 / 5 of the width of the brain medical image, and the cropping center of the right temporal lobe image patch is located at 3 / 4 or 4 / 5 of the width of the brain medical image. The cropped central region image patch, left temporal lobe image patch, and right temporal lobe image patch are each subjected to pixel normalization processing, and then merged in the batch dimension to obtain local input data. The process of obtaining local deep features by weighting the preliminary local deep features using a coordinate attention submodule located at the output of the local backbone network is as follows: Global average pooling is performed on the preliminary local deep features along both the height and width directions. The pooled features from both directions are then concatenated, dimensionality reduced, and activated before being separated back into the height and width directions. A sigmoid activation function is then used to generate horizontal and vertical attention weights. The trained relative coordinate parameters are mapped to the spatial index of the preliminary local deep features. The horizontal and vertical attention weights corresponding to the spatial index are then weighted and enhanced. Finally, the enhanced horizontal and vertical attention weights are multiplied element-wise with the preliminary local deep features to obtain the local deep features.

[0026] S22. The intermediate layer cross-flow attention submodule is used to fuse global and local intermediate layer features to obtain intermediate layer fused features. Specifically: Figure 5As shown, the intermediate layer cross-stream attention submodule includes an attention generation layer and an intermediate weighted fusion layer. The attention generation layer includes a multilayer perceptron and a Softmax activation function set sequentially. After mapping the global intermediate features and local intermediate features to the same dimension, feature residual information is extracted. The global intermediate features and local intermediate features after extracting feature residual information are concatenated to obtain intermediate data. The attention generation layer processes the intermediate data to obtain global initial attention weights and local initial attention weights. Both global initial attention weights and local initial attention weights are L2 normalized and truncated according to a fixed range determined during the training phase to obtain global final attention weights and local final attention weights. The intermediate weighted fusion layer uses the global final attention weights and local final attention weights to perform weighted fusion with the global intermediate features and local intermediate features respectively to obtain intermediate fused features.

[0027] The mid-level fusion features are enhanced across layers to obtain global and local deep features, respectively, resulting in globally enhanced deep features and locally enhanced deep features. Specifically: For global deep features, the mid-level fusion features are projected to the same dimension as the global deep features. The projected mid-level fusion features are then concatenated with the global deep features, and the concatenated features are linearly fused and dimensionality reduced to obtain the globally enhanced deep features. For local deep features, the mid-level fusion features are projected to the same dimension as the local deep features. The projected mid-level fusion features are then added to the local deep features using residual summation to obtain locally enhanced deep features.

[0028] S23. The deep dynamic cross-attention submodule is used to update the global enhanced deep features and the aggregated local enhanced deep features to obtain the global updated features and the local updated features.

[0029] like Figure 4 As shown, the aggregation method for local enhanced deep features is as follows: the local feature flow branch includes an image patch aggregation submodule for aggregating local enhanced deep features. The image patch aggregation submodule includes a local multi-head self-attention unit and an image patch gating network. The local enhanced deep features are input into the local multi-head self-attention unit for fusion to obtain local fused deep features. A role embedding vector is introduced into the local fused deep features. The role embedding vector is added to the local fused deep features and then input into the image patch gating network to obtain the dynamic weight score of each local enhanced deep feature. The local enhanced deep features and the corresponding dynamic weight scores are weighted and summed to obtain the local aggregated deep features.

[0030] The deep dynamic cross-attention submodule includes a first multi-head attention submodule and a second multi-head attention submodule. The first multi-head attention submodule takes locally aggregated deep features as the query vector and globally enhanced deep features as the key vector and value vector, and inputs them to obtain locally updated features. The second multi-head attention submodule takes globally enhanced deep features as the query vector and locally updated features as the key vector and value vector, and inputs them to obtain globally updated features.

[0031] S24. The weighted fusion module is used to perform weighted fusion of the global and local updated features to obtain the final fused feature. The fusion weights determined during the training phase are used to perform weighted summation of the global and local updated features to obtain the final fused feature.

[0032] S25. The classification module outputs Alzheimer's disease classification results based on the final fused features. The classification module includes a classifier composed of fully connected layers and a Softmax activation function. The final fused features are input into the classifier, which outputs the probability distribution of brain medical images belonging to Alzheimer's disease, mild cognitive impairment, and normal cognition. The category with the highest probability is taken as the Alzheimer's disease classification result.

[0033] It should be noted that the specific structures and pre-training processes of the ResNet101 and EfficientNet-B0 networks are conventional existing technologies in this field and will not be elaborated further.

[0034] During the training of the global backbone network, in order to utilize large-scale pre-trained parameters, the weights of the first convolutional layer of the global backbone network are initialized to the mean of the RGB three-channel weights of the pre-trained model. To preserve the basic edge and texture features of brain medical images and reduce computational overhead, the pre-convolutional layers, max pooling layers, and the first layer group of the global backbone network are frozen.

[0035] During the training of the local backbone network, in order to utilize large-scale pre-trained parameters, the weights of the first convolutional layer of the local backbone network are initialized to the mean of the RGB three-channel weights of the pre-trained model. In order to preserve the general low-level visual features, the first four stage blocks of the local backbone network are frozen.

[0036] During the training of the coordinate attention submodule, a localization mechanism is introduced. If the input brain medical image has undergone random horizontal flipping data augmentation, the horizontal component of the relative coordinate parameters needs to be inverted accordingly, while the vertical component remains unchanged. If the input brain medical image has not been flipped, the relative coordinate parameters remain unchanged.

[0037] like Figure 6As shown, during the training process of the weighted fusion module, a three-stage dynamic weight strategy is adopted to dynamically calculate the fusion weights of global and local updated features based on the training round progress of the Alzheimer's disease classification model.

[0038] The first stage, the initial training phase, employs equal fusion weights and introduces learnable, small dynamic offset parameters. The global feature flow branch primarily learns macroscopic brain structural features such as the ventricles and cortex, while the local feature flow branch primarily learns preliminary features. The second stage, the mid-training phase, introduces a task-guided gating mechanism based on class differences and feature uncertainty estimation. By evaluating the distribution of global and local features, the fusion weights are dynamically adjusted, guiding the Alzheimer's classification model's attention smoothly from the macroscopic whole to local, subtle lesions. The third stage, the late training phase, calculates fusion weights based on the cosine similarity of global and local features and the attention intensity of local image patches. A dynamic noise enhancement mechanism is introduced in the fusion weight calculation to expand the exploration scope.

[0039] During model training, a joint loss function is used for parameter updates. The joint loss function includes cross-entropy loss, which measures the classification prediction error; cross-flow consistency loss, which constrains the consistency between global and local features; and position guidance loss, which guides the alignment of the spatial location of pathological features of regions with strong local feature responses with those of lesions.

[0040] To further verify the effectiveness of the Alzheimer's disease classification method based on dynamic cross-stream fusion of two-stream neural networks provided by this invention, this embodiment conducted ablation experiments and comparison experiments with mainstream models on a publicly available Alzheimer's disease magnetic resonance imaging dataset. The evaluation metrics included accuracy, precision, recall, and F1 score.

[0041] To verify the specific contribution of each module in this invention to the overall performance of the Alzheimer's disease classification model, the following experimental groups were set up using the controlled variable method: 1. Global feature flow branch only: Only the global feature flow branch was used to extract features for classification; 2. Local feature flow branch only: Only the local feature flow branch was used to extract features for classification; 3. Traditional dual-flow fusion: A parallel structure of global and local dual flows was adopted, but no cross-flow interaction module was introduced, and the end was fused using the traditional static splicing method; 4. Cross-flow interaction + traditional static splicing: The dynamic cross-flow interaction module of this invention was introduced, but the end was fused using the traditional static splicing method; 5. Complete Alzheimer's disease classification model: A complete Alzheimer's disease classification model containing all core modules. The ablation experiment results are shown in Table 1.

[0042] Table 1 Ablation Experiment of Core Module As shown in Table 1, the complete Alzheimer's disease classification model of this invention has significantly improved accuracy, precision, recall and F1 score compared with other experimental groups.

[0043] To verify the leading performance of the model of this invention in the Alzheimer's disease classification task, a comparative experiment was conducted with the current mainstream deep learning image classification models, and the results are shown in Table 2.

[0044] Table 2 Model Comparison As shown in Table 2, the accuracy, precision, recall, and F1 score of the complete Alzheimer's disease classification model of this invention all reach the optimal level. The overall performance of the Alzheimer's disease classification model is better than the current mainstream deep learning image classification models.

[0045] This invention provides a dual-stream neural network Alzheimer's disease classification system based on dynamic cross-stream fusion, including a data processing module and a data display module.

[0046] The data processing module is used to classify brain medical images using a pre-trained Alzheimer's disease classification model and output Alzheimer's disease classification results. The data visualization module is used to visually represent the classification results of Alzheimer's disease.

[0047] The present invention also provides a computer device, including a memory and a processor.

[0048] Memory is used to store computer programs.

[0049] A processor is used to read and execute the computer program to implement the above-described two-stream neural network Alzheimer's disease classification method based on dynamic cross-stream fusion.

[0050] Finally, the present invention provides a storage medium for storing a computer program that, when executed, implements the above-described method for classifying Alzheimer's disease using a dual-stream neural network based on dynamic cross-stream fusion.

[0051] Memory, as a carrier of resources, can be read-only memory, random access memory, disk, or optical disk, etc. The resources stored on it can include operating systems, computer programs, etc., and the storage method can be temporary or permanent storage. The processor can be a general-purpose processor product based on architectures such as x86, IA64, RISC, MIPS, and ARM.

[0052] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A two-stream neural network-based Alzheimer's disease classification method based on dynamic cross-stream fusion, characterized in that: Includes the following steps: Acquiring medical images of the brain; The Alzheimer's disease classification model inputs brain medical images into a pre-trained model for classification and outputs the classification results. The model includes a two-stream feature extraction module, a dynamic cross-stream interaction module, a weighted fusion module, and a classification module. The two-stream feature extraction module comprises two parallel global feature stream branches and a local feature stream branch. The global feature stream branch includes a global backbone network based on ResNet, and the local feature stream branch includes a local backbone network based on EfficientNet. The dynamic cross-stream interaction module includes a mid-level cross-stream attention submodule and a deep dynamic cross-attention submodule. The Alzheimer's disease classification model classifies brain medical images using the following methods: Global backbone networks are used to extract global brain structural features from medical brain images to obtain global mid-level features and global deep features. Local backbone networks are used to extract local brain structural features from medical brain images to obtain local mid-level features and local deep features. The intermediate layer cross-flow attention submodule is used to fuse global intermediate layer features and local intermediate layer features to obtain intermediate layer fused features. The intermediate layer fused features are then enhanced across layers to global deep features and local deep features to obtain global enhanced deep features and local enhanced deep features, respectively. The deep dynamic cross-attention submodule is used to update the global enhanced deep features and the aggregated local enhanced deep features to obtain the global updated features and the local updated features; The weighted fusion module is used to perform weighted fusion of global update features and local update features to obtain the final fused features; The classification module outputs Alzheimer's disease classification results based on the final fused features.

2. The Alzheimer's disease classification method based on dynamic cross-stream fusion dual-stream neural network according to claim 1, characterized in that: The method for extracting global brain structural features using a global backbone network to obtain global mid-level features and global deep features includes: The preprocessed brain medical images were used as global input data for the global backbone network; The first three layers of the global backbone network are used to extract features from the global input data to obtain global mid-level features. The global mid-level features are then propagated forward to the fourth layer of the global backbone network to obtain global preliminary deep features. The global preliminary deep features are then weighted to obtain global deep features.

3. The Alzheimer's disease classification method based on dynamic cross-stream fusion of a two-stream neural network according to claim 2, characterized in that: The output of the global backbone network is equipped with a global attention submodule, which includes a channel attention unit and a spatial attention unit. The channel attention unit is used to perform channel weighting on the channels of the preliminary global deep features, and the spatial attention unit is used to perform spatial weighting on the channel-weighted preliminary global deep features to obtain the global deep features.

4. The Alzheimer's disease classification method based on a two-stream neural network with dynamic cross-stream fusion as described in claim 1, characterized in that: The method for extracting local brain structural features using local backbone networks to obtain local mid-level features and local deep features includes: Brain medical images are cropped according to preset cropping rules to obtain central region image blocks, left temporal lobe region image blocks and right temporal lobe region image blocks. The central region image blocks, left temporal lobe region image blocks and right temporal lobe region image blocks are merged in the batch dimension and used as local input data for the local backbone network. The first five or six stages of the local backbone network are used to extract features from the local input data to obtain local mid-level features. The local mid-level features are then propagated forward to the ninth stage of the local backbone network to obtain local preliminary deep features. The local preliminary deep features are then weighted to obtain local deep features.

5. The Alzheimer's disease classification method based on a two-stream neural network with dynamic cross-stream fusion according to claim 4, characterized in that: The output of the local backbone network is equipped with a coordinate attention submodule, which is used to spatially weight the local preliminary deep features to obtain local deep features.

6. The Alzheimer's disease classification method based on dynamic cross-stream fusion of a two-stream neural network according to claim 1, characterized in that: The method for obtaining mid-level fused features by weighted fusion of global and local mid-level features using the intermediate layer cross-current attention submodule includes: The intermediate layer cross-stream attention submodule includes an attention generation layer and an intermediate weighted fusion layer. The attention generation layer includes a multilayer perceptron and a Softmax activation function set sequentially. After mapping the global mid-level features and local mid-level features to the same dimension, the feature residual information is extracted. The global mid-level features and local mid-level features after extracting the feature residual information are then concatenated to obtain intermediate data. The intermediate data is processed by the attention generation layer to obtain the global initial attention weight and the local initial attention weight. The global initial attention weight and the local initial attention weight are normalized to obtain the global final attention weight and the local final attention weight respectively. The intermediate weighted fusion layer is used to perform weighted fusion of the global final attention weight and the local final attention weight with the global intermediate features and the local intermediate features, respectively, to obtain the intermediate fused features.

7. The Alzheimer's disease classification method based on dynamic cross-stream fusion of a two-stream neural network according to claim 1, characterized in that: The method for updating global enhanced deep features and local enhanced deep features using a deep dynamic cross-attention submodule to obtain global updated features and local updated features includes: The local feature flow branch includes an image patch aggregation submodule for aggregating local enhanced deep features. The image patch aggregation submodule includes a local multi-head self-attention unit and an image patch gating network. The local enhanced deep features are input into the local multi-head self-attention unit for fusion to obtain local fused deep features. A role embedding vector is introduced into the local fused deep features. The role embedding vector is added to the local fused deep features and then input into the image patch gating network to obtain the dynamic weight score of each local enhanced deep feature. The local enhanced deep features and the corresponding dynamic weight scores are weighted and summed to obtain the local aggregated deep features. The deep dynamic cross-attention submodule includes a first multi-head attention submodule and a second multi-head attention submodule. The first multi-head attention submodule takes locally aggregated deep features as the query vector and globally enhanced deep features as the key vector and value vector, and inputs them to obtain locally updated features. The second multi-head attention submodule takes globally enhanced deep features as the query vector and locally updated features as the key vector and value vector, and inputs them to obtain globally updated features.

8. An Alzheimer's disease classification system based on a two-stream neural network with dynamic cross-stream fusion, characterized in that: include: The data processing module is used to classify brain medical images using a pre-trained Alzheimer's disease classification model and output Alzheimer's disease classification results. The data visualization module is used to visually represent the classification results of Alzheimer's disease.

9. A computer device, characterized in that: include: Memory, used to store computer programs; A processor for reading and executing computer programs to implement the Alzheimer's disease classification method based on a dynamic cross-stream fusion dual-stream neural network as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: Used to store a computer program, which, when executed, implements the Alzheimer's disease classification method based on a dynamic cross-stream fusion dual-stream neural network as described in any one of claims 1 to 7.

Images