Early Alzheimer's disease feature extraction method based on ResNet model

Abstract

The invention relates to an early Alzheimer's disease feature extraction method based on a ResNet model. The method comprises the following steps: preprocessing an image data sample image; a ResNet-50 feature extraction model is constructed on the basis of the ResNet-50 feature extraction network; training a feature extraction model by using the preprocessed sample image; and inputting AD image data to be processed into the trained feature extraction model, and outputting a feature extraction result. The M-Max-Avg pooling layer is used for capturing features of different levels, so that AD image feature extraction is more flexible, and the accuracy of model prediction is improved; according to the multi-scale grouping convolution layer, the grouping convolution method is applied to the ResNet-50 network, the parameter quantity during convolution is reduced, the prediction speed is increased, and the multi-scale convolution method is used in each group, so that the model can comprehensively learn the features of input data, extract rich feature information and improve the accuracy of the model on AD prediction.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of virtual surgery, and particularly relates to an early Alzheimer's disease feature extraction method based on a ResNet model. BACKGROUND

[0002] In deep learning, the key to predicting the development of Alzheimer's disease by using early (still not ill or mild cognitive impairment stage) neuroimaging data (such as MRI, PET, fMRI, etc.) lies in extracting subtle biomarker features related to disease progression from seemingly normal images. The following are key features that may exist in the feature map of an early AD patient: whether the hippocampus volume is reduced, whether the cortical thickness of the temporal lobe, parietal lobe and other regions is thinned, whether the ventricle volume is increased, etc. These features are analyzed early to predict the probability of the patient developing AD later.

[0003] Residual neural network (ResNet) is a neural network model proposed by researchers such as He Kai-ming and Zhang Xiang-yu of Microsoft Research in 2015. The main contribution of the model lies in effectively solving the problem of deep neural network training, but the model is not comprehensive enough for feature extraction, and the calculation speed of the model is not fast enough, so when building an optimized ResNet model, how to design the model to be lightweight and better extract global features of images is still a research hotspot in the field of artificial intelligence.

[0004] In the late 1980s, with the development of convolutional neural networks, the idea of pooling operation began to be introduced and applied to convolutional neural networks by early researchers to downsample the output of the convolutional layer. Pooling operation can effectively reduce the data volume and computational complexity. Traditional pooling operations include max pooling and average pooling. For regions with obvious features, key information can be extracted through max pooling, but other useful information may be ignored. For regions with no obvious features and more noise, average pooling can reduce the influence of noise, but important features may be overwhelmed by noise, so a single pooling operation cannot well flexibly process image features, and traditional pooling layers cannot well capture features at different levels. SUMMARY

[0005] The application provides an early Alzheimer's disease feature extraction method based on a ResNet model, including: acquiring a sample image and performing a preprocessing operation on the sample image; selecting ResNet-50 as a backbone feature extraction network, adjusting the convolution structure of the ResNet-50 network by using a multi-scale grouping convolution structure, adjusting the pooling layer of the ResNet-50 network by using an M-Max-Avg pooling layer, and concatenating a channel attention module after the M-Max-Avg pooling layer to generate a ResNet-50 feature extraction model; training the ResNet-50 feature extraction model by using the preprocessed sample image; and inputting AD image data to be processed into the trained ResNet-50 feature extraction model to output a feature extraction result.

[0006] In some embodiments, the preprocessing of the sample image includes: completing missing values in the sample image, deleting and replacing abnormal values, and unifying the format of the sample image; performing normalization processing on the sample image in a unified format; and performing data augmentation on the normalized sample image.

[0007] In some embodiments, the adjusting of the convolution structure of the ResNet-50 network by using the multi-scale grouping convolution structure includes: dividing the input channels of the ResNet-50 network into k groups on average; dividing the convolution kernel of each residual unit in the ResNet-50 network into t groups on average, and using convolution kernels of different sizes within each group of convolution kernels.

[0008] In some embodiments, the adjusting of the pooling layer of the ResNet-50 network by using the M-Max-Avg pooling layer includes: defining a plurality of different sizes of pooling windows with a step length s, and connecting the pooling windows in parallel; calculating the average value of the data inside the pooling window, and calculating the standard deviation of the data inside the window according to the average value of the data inside the pooling window;

[0009] mapping the standard deviation to a set interval by using a Sigmoid function to generate a parameter M; adjusting the weights of the maximum pooling and the average pooling according to the parameter M; extracting the features of the sample image in each window according to the weights by using the maximum pooling and the average pooling; and splicing the features of the sample image extracted by the plurality of windows to generate a target AD feature map.

[0010] In some embodiments, the ResNet-50 feature extraction model is trained using the preprocessed sample image, including: after the sample image is subjected to convolution, standardization processing and nonlinear transformation, the sample image is input into an M-Max-Avg pooling layer to generate a target AD feature map; a channel attention module is used to generate channel attention weights according to the target AD feature map, and the target AD feature map is fused according to the channel attention weights to generate a fused feature; the fused feature is input into a residual unit for multi-scale grouped convolution operation to generate multi-scale AD features; the multi-scale AD features are sequentially input into an M-Max-Avg pooling layer and a channel attention module to generate a target output feature; the target output feature is input into a fully connected layer for linear integration and then input into a Softmax layer to generate a probability that the sample image belongs to a set category, and the category with the highest probability and the corresponding feature are output as the feature extraction result; wherein the probabilities of all categories sum to 1; a loss value is calculated using a cross-entropy loss function, and the training is completed when the loss value is less than a set threshold.

[0011] In some embodiments, the channel attention module is used to generate channel attention weights according to the target AD feature map, including: performing global average pooling operation on the AD feature map to generate the average value of all pixels in each channel of the AD feature map; performing global maximum pooling operation on the AD feature map to generate the maximum value of all pixels in each channel of the AD feature map; inputting the average value and the maximum value into a multilayer perceptron for nonlinear transformation to generate and fuse high-level feature representations; and using a Sigmoid function to map the fused high-level feature representations to a set interval to generate the channel attention weights.

[0012] In some embodiments, the fused feature is input into a residual unit for multi-scale grouped convolution operation to generate multi-scale AD features, including: in the fused feature, all input channels in a wth input channel group are respectively convolved with all convolution kernels in an ath convolution kernel group, and the output features are fused; until all input channels in k input channel groups are convolved with all convolution kernels in t convolution kernel groups to generate k*t fused output features; the k*t fused output features are connected according to the channel dimension to generate multi-scale AD features; wherein w, k, a, and t are positive integers, w is less than or equal to k, and a is less than or equal to t.

[0013] In some embodiments, the set categories include labeled normal sample images, mild cognitive impairment sample images and suspected Alzheimer's disease sample images.

[0014] 3. Beneficial effects

[0015] The present application adopts the above method and has the following beneficial effects:

[0016] 1. The M-Max-Avg pooling layer is proposed, which combines the advantages of maximum pooling and average pooling, allowing the model to capture features at different levels, making AD image feature extraction more flexible, avoiding the limitations of traditional single pooling methods, and improving the accuracy of model prediction.

[0017] 2. A multi-scale group convolution layer is proposed, and the group convolution method is applied to ResNet-50 for lightweight processing, reducing the number of parameters during convolution, speeding up the prediction of AD, and allowing patients to receive treatment as soon as possible; and the multi-scale convolution method is used within each group, so that the model can comprehensively learn the characteristics of the input data, extract rich feature information, and improve the accuracy of the model's AD prediction.

[0018] 3. The channel attention module is placed after the multi-scale group convolution and M-Max-Avg pooling layers to alleviate the problem of limited cross-group feature communication, allowing the model to pay more attention to the feature information of AD images that is more valuable to the task. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of a method for extracting features of early Alzheimer's disease based on a ResNet model in one embodiment;

[0020] Figure 2 A network structure diagram of a ResNet prediction model in one embodiment;

[0021] Figure 3 Schematic diagram of multi-scale group convolution in one embodiment. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0023] like Figure 1 and Figure 2 As shown, this application provides a method for extracting early Alzheimer's disease features based on the ResNet model, including:

[0024] S101: Acquire a sample image and perform a preprocessing operation on the sample image.

[0025] S201: Select ResNet-50 as the backbone feature extraction network, use the multi-scale group convolution structure to adjust the convolution structure of the ResNet-50 network, use the M-Max-Avg pooling layer to adjust the pooling layer of the ResNet-50 network, and connect the channel attention module in series after the pooling layer to generate the ResNet-50 feature extraction model.

[0026] S301, training a ResNet-50 feature extraction model using the preprocessed sample images.

[0027] S401, inputting the AD image data to be processed into the trained ResNet-50 feature extraction model, and outputting a feature extraction result.

[0028] The present application proposes an M-Max-Avg pooling layer, which combines the advantages of max pooling and average pooling, allowing the model to capture features at different levels, making image feature extraction for AD more flexible, avoiding the limitations of traditional single pooling methods, and improving the performance of the model.

[0029] The present application also proposes a multi-scale grouping convolution layer, which applies the grouping convolution method to ResNet-50 for lightweight processing, reduces the number of parameters during convolution, speeds up the prediction of AD, and allows patients to receive treatment as soon as possible; and uses multi-scale convolution method in each group to make the model learn the features of the input data comprehensively, extract rich feature information, and improve the accuracy of AD prediction.

[0030] In addition, the present application proposes placing a channel attention module after the multi-scale grouping convolution and M-Max-Avg pooling layer to alleviate the problem of limited cross-group feature exchange, allowing the model to pay more attention to the feature information of AD images that are more valuable for the task.

[0031] In some embodiments, the sample images are preprocessed, including: completing the missing values in the sample images, deleting and replacing the abnormal values, and unifying the formats of the sample images; normalizing the sample images of the unified format; and performing data augmentation on the normalized sample images.

[0032] Completing the missing values in the sample images, deleting and replacing the abnormal values, and unifying the image formats make it easier to store and share data, reducing uncertainty in the training process. Normalizing the data in the data set standardizes the data, mapping all features to the same numerical range, allowing direct comparison between different features, improving the accuracy and stability of the model. Then, the data is rotated, cropped, scaled, and other data augmentation operations are performed to increase the diversity of the data and improve the generalization ability of the model.

[0033] Through the above operations, a feature map is obtained, which usually has lower dimensions and more abstract representations than the original image. It compresses and refines the information in the image, removing unnecessary details and highlighting important features related to a specific task.

[0034] In some embodiments, the convolution structure of the ResNet-50 network is adjusted using a multi-scale grouped convolution structure, including: dividing the input channels of the ResNet-50 network into k groups on average; dividing the convolution kernel of each residual unit in the ResNet-50 network into t groups on average, and using different sizes of convolution kernel within each group of convolution kernel.

[0035] Specifically, first, ResNet-50 is selected as the backbone extraction network model, and the ResNet-50 network model has 4 groups of residual blocks. Each group of residual blocks contains 3, 4, 6, and 3 residual units, respectively, and each residual unit contains multiple convolution kernels. All the channel numbers of the input feature maps in the ResNet-50 network model are divided into k groups on average, each group containing the same number of channels, and all the convolution kernels in a residual unit are also divided into t groups on average, each group containing the same number of convolution kernels. The number of channels and the number of convolution kernels in each group are calculated as follows:

[0036] Channel=m / k

[0037] Convolution=n / t

[0038] Where Channel is the number of channels contained in each group, m is the total number of channels of the input feature map, Convolution is the number of convolution kernels contained in each group, and n is the total number of convolution kernels.

[0039] Then, all the traditional convolution structures in the residual units of the ResNet-50 network model are replaced with grouped convolution structures, which can effectively reduce the calculation amount of model parameters and improve the stability of the model.

[0040] For each group of grouped convolution kernels, different sizes of convolution kernels are applied within the group. The present application uses four different sizes of convolution kernels, which are 1x1, 3x3, 5x5, and 7x7, with a step size of 2. These four types of convolution kernels are connected in parallel to comprehensively capture the features of the input data at different spatial levels. Among them, the convolution kernel with a size of 7x7 has the largest receptive field, which is used to extract the global features of the input features in a large range, the convolution kernel with a size of 5x5 starts to involve larger range feature extraction, which is used to capture the global features of the input features in a medium range, the convolution kernel with a size of 1x1 is mainly used to extract small range detail features of the input features and focus on local areas, and the convolution kernel with a size of 3x3 increases the size of the receptive field based on the 1x1 convolution kernel, which can capture more rich local features.

[0041] In some embodiments, the pooling layer of the ResNet-50 network is adjusted by using the M-Max-Avg pooling layer, including: defining a plurality of different sizes of pooling windows with a step size s (s is an arbitrary value), and connecting the pooling windows in parallel; calculating the mean value in the pooling window, and calculating the standard deviation according to the mean value; using the Sigmoid function to map the standard deviation to the target interval (0, 1) after linear scaling, to generate a parameter M; adjusting the weights of the maximum pooling and the average pooling according to the parameter M; in each window, extracting the features of the sample image according to the weights by using the maximum pooling and the average pooling; and splicing the features of the sample image extracted by the plurality of windows to generate a target AD feature map.

[0042] Specifically, in one M-Max-Avg pooling layer, three different sizes of pooling windows, 2x2, 3x3, and 4x4, with a step size of 2 are defined, and they are connected in parallel. Among them, the 2x2 pooling window can extract relatively small scale local features, focusing on detailed information; the 3x3 pooling window can capture local features while fusing more context information, suitable for medium scale feature extraction; and the 4x4 pooling window mainly extracts large scale global features, and has a better grasp of the overall structure and layout of the image. Fusing the pooling results of the three different scales can make the model obtain multi-scale feature information of the image from details to the whole, thereby improving the recognition ability of the model to various complex scenes and different scale objects, and improving the overall efficiency and generalization ability of the model.

[0043] The mean value and the standard deviation in the pooling window are calculated, and finally the Sigmoid function is used to map the standard deviation to the target interval (0, 1) after linear scaling, to generate a parameter M. The proportion of the maximum pooling and the average pooling is adjusted by the dynamic weight M. When the standard deviation σ is large, M tends to 1, and it indicates that the difference between the elements in the window is large, so more weight is allocated to the average pooling operation to reflect the overall trend of the elements in the window. When the standard deviation σ is small, M tends to 0, indicating that the difference between the elements in the window is small, so more weight is allocated to the maximum pooling operation to capture possible significant features. In this way, the weight of the pooling operation can be dynamically adjusted according to the difference (standard deviation σ) between the elements in the window. This enables the model to adaptively select a more suitable pooling strategy for the current data features, thereby improving the flexibility and accuracy of feature extraction.

[0044] In some embodiments, the ResNet-50 feature extraction model is trained using the preprocessed sample image, including: after the sample image is subjected to convolution, standardization processing and nonlinear transformation, the sample image is input into an M-Max-Avg pooling layer to generate a target AD feature map; a channel attention module is used to generate channel attention weights according to the target AD feature map, and the target AD feature map is fused according to the channel attention weights to generate a fused feature; the fused feature is input into a residual unit for multi-scale grouped convolution operation to generate multi-scale AD features; the multi-scale AD features are sequentially input into the M-Max-Avg pooling layer and the channel attention module to generate a target output feature; the target output feature is input into a fully connected layer for linear integration and then input into a Softmax layer to generate a probability that the sample image belongs to a set category, and the category with the highest probability and the corresponding feature are output as the feature extraction result; wherein the probabilities of all categories sum to 1; a loss value is calculated using a cross-entropy loss function, and the training is completed when the loss value is less than a set threshold.

[0045] In some embodiments, the channel attention module is used to generate channel attention weights according to the target AD feature map, including: performing global average pooling operation on the AD feature map to generate the average value of all pixels in each channel of the AD feature map; performing global maximum pooling operation on the AD feature map to generate the maximum value of all pixels in each channel of the AD feature map; inputting the average value and the maximum value into a multilayer perceptron for nonlinear transformation to generate and fuse high-level feature representations; and mapping the fused high-level feature representations to a set interval using a Sigmoid function to generate the channel attention weights.

[0046] In some embodiments, the fused feature is input into a residual unit for multi-scale grouped convolution operation to generate multi-scale AD features, including: in the fused feature, all input channels in a wth input channel group are respectively convolved with all convolution kernels in an ath convolution kernel group, and the output features are fused; until all input channels in k input channel groups are convolved with all convolution kernels in t convolution kernel groups to generate k*t fused output features; the k*t fused output features are connected according to the channel dimension to generate multi-scale AD features; wherein w, k, a, and t are positive integers, w is less than or equal to k, and a is less than or equal to t.

[0047] In some embodiments, the set categories include labeled normal sample images, mild cognitive impairment sample images and suspected Alzheimer's disease sample images.

[0048] Specifically, referring to Figure 2First, the pre-processed sample image is convoluted by a convolution kernel with a size of 7x7 and a step of 2, and after the convolution operation is completed, the batch normalization technology is used for standardization processing, and then the ReLU(x) activation function is used for nonlinear transformation of the data:

[0049]

[0050] Wherein, x is the input data after standardization processing.

[0051] After the activation function operation is completed, the sample image is further processed by an M-Max-Avg pooling layer. The sample image is simultaneously processed by a pooling window with a size of 2x2, 3x3, and 4x4 and a step of 2. When the pooling window with a size of 2x2 and a step of 2 is used, the average value μ of the data in the pooling window is first calculated, and then the standard deviation σ of the data in the window is calculated by using the average value μ of the data in the pooling window. The calculation formula of the average value μ of the data in the window and the standard deviation σ of the data in the window is as follows:

[0052]

[0053] Wherein, x i is the i-th element data in the pooling window.

[0054] Then the standard deviation σ of the data in the window is mapped to M in the interval (0, 1) by using the Sigmoid function, and the calculation formula of M is as follows:

[0055]

[0056] Wherein, c is the slope of the Sigmoid function; σ0 controls the center point of the Sigmoid function, which is the average value of the standard deviation σ in the window.

[0057] The proportion of the maximum pooling and the average pooling is adjusted by using M to dynamically weight. If σ is large, it means that the difference between the elements in the window is large, and more weight is allocated to the average pooling operation. If σ is small, it means that the difference between the elements in the window is small, and more weight is allocated to the maximum pooling operation. The formula for dynamically allocating the pooling operation by using M is as follows:

[0058] Output=M·AveragePooling+(1-M)·MaxPooling

[0059] Wherein, Output is the output feature map; AveragePooling is the output of the average pooling; MaxPooling is the output of the maximum pooling.

[0060] Similarly, the AD image data is processed in the same way when it passes through pooling windows of size 3×3, 4×4 and stride 2. Finally, the three pooling output results are added and connected to obtain a new AD image feature map.

[0061] The AD image feature map is then passed through the channel attention module, where two different pooling operations are performed: global average pooling and global max pooling. The former calculates the average value of all pixels in each channel of the feature map to obtain a value representing the overall information of that channel, while the latter takes the maximum value of all pixels in each channel to obtain a value representing the most significant feature of that channel. Both pooling results are then fed into a shared multi-layer perceptron (MLP). This perceptron consists of multiple fully connected layers, each containing multiple neurons. It performs nonlinear transformations on the input to extract higher-level feature representations. After MLP processing, the two results are added together and the fused result is finally fed into a sigmoid function, which maps the result to the interval (0, 1) to obtain the channel attention weight Wc(F). A weight closer to 1 indicates a more important channel, while a weight closer to 0 indicates a less important channel. The channel attention module automatically learns and adjusts the importance of each channel, enabling the model to better understand and distinguish the importance of different channels in the input feature map. The calculation formula of channel attention weight Wc(F) is as follows:

[0062] Wc(F)=Sigmoid(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0063] Among them, F is the input feature map; MLP is a shared multi-layer perceptron; AvgPool and MaxPool are global average pooling operations and global maximum pooling operations respectively.

[0064] like Figure 3 As shown in the figure, the AD image feature map is input into the residual unit. First, all the internal channels of the first input channel group in the AD image feature map are convolved with all the convolution kernels of different sizes in the first convolution kernel group. After all the convolutions within the group are completed, the results are added and fused. Then, all the internal channels of the first input channel group are convolved with all the convolution kernels of different sizes in the second convolution kernel group, and the results are added and fused. Repeat the above similar operations until the first input channel group and all the convolution kernel groups complete the convolution operation and output. And so on, until the k input channel groups are convolved with t convolution kernel groups respectively. Finally, all the grouped output results are connected according to the channel dimension to obtain the final output feature.

[0065] After passing through all the residual units, the AD image feature map is obtained, and then the output feature map is input into an M-Max-Avg pooling layer and a channel attention module. Finally, the feature map is linearly integrated through a fully connected layer, and the output value of the fully connected layer is converted into a probability distribution through a Softmax layer. The classification labels are normal sample images, mild cognitive impairment sample images and suspected Alzheimer's disease sample images, each of which is assigned a probability value, and the sum is 1. These probability values represent the likelihood of the input image belonging to each category. According to the probability distribution output by the Softmax layer, the class with the maximum probability is selected as the predicted class of the input image, and the prediction result and the corresponding features of each image are recorded as the feature extraction result, so as to be evaluated subsequently, such as calculating the accuracy, recall rate and other indicators. The loss value is calculated by the cross-entropy loss function CEloss (Cross-entropy loss function), so that the predicted probability distribution of the model is continuously close to the true probability distribution, and the training is completed when the loss value is less than the set threshold, so that the model can classify more accurately.

[0066] In the present application, in order to ensure the training and testing effect of the model, the sample images are divided into training set, validation set and test set according to 8:1:1. The training set is used to train the ResNet-50 feature extraction model, the validation set is used to evaluate the generalization ability of the model, and the test set is used to determine the accuracy of the model in AD diagnosis or prediction.

[0067] The technical features of the above embodiments can be combined in any way. In order to make the description simple, not all possible combinations of technical features in the above embodiments are described, but as long as the combination of technical features does not exist contradictory, it should be considered as the scope of the present application.

[0068] The above-described embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. The early Alzheimer's disease feature extraction method based on the ResNet model is characterized by: include: Acquire a sample image and perform preprocessing operations on the sample image; We selected ResNet-50 as the backbone feature extraction network, adjusted the convolution structure of the ResNet-50 network using the multi-scale group convolution structure, adjusted the pooling layer of the ResNet-50 network using the M-Max-Avg pooling layer, and connected the channel attention module in series after the pooling layer to generate the ResNet-50 feature extraction model. Use the preprocessed sample images to train the ResNet-50 feature extraction model; Input the image data to be processed into the trained ResNet-50 feature extraction model and output the feature extraction results.

2. The method according to claim 1, characterized in that Preprocess the sample image, including: Fill in missing values ​​in sample images, delete and replace outliers, and unify the format of sample images; Normalize the sample images in a unified format; Perform data expansion on the normalized sample image.

3. The method according to claim 1, characterized in that The convolution structure of the ResNet-50 network is adjusted using a multi-scale grouped convolutional structure, including: Divide the input channels of the ResNet-50 network into k groups evenly; The convolution kernels of each residual unit in the ResNet-50 network are evenly divided into t groups, and convolution kernels of different sizes are used in each convolution kernel group.

4. The method according to claim 3, characterized in that Use the M-Max-Avg pooling layer to adjust the pooling layer of the ResNet-50 network, including: Define multiple pooling windows of different sizes and step size s, and connect the pooling windows in parallel; Calculate the mean within the pooling window and calculate the standard deviation based on the mean; Use the Sigmoid function to linearly scale the standard deviation and map it to the target interval (0,1) to generate the parameter M; Adjust the weights of maximum pooling and average pooling according to the parameter M; In each window, the features of the sample image are extracted using maximum pooling and average pooling according to the weights; The features of the sample images extracted by multiple windows are spliced ​​together to generate the target AD feature map.

5. The method according to claim 4, characterized in that The ResNet-50 feature extraction model is trained using the preprocessed sample images, including: After the sample image undergoes convolution, normalization, and nonlinear transformation, it is input into the M-Max-Avg pooling layer to generate the target AD feature map; Use the channel attention module to generate channel attention weights according to the target AD feature map, fuse the target AD feature map according to the channel attention weights to generate fused features; The fused features are input into the residual unit for multi-scale group convolution operation to generate multi-scale AD features; The multi-scale AD features are sequentially input into the M-Max-Avg pooling layer and the channel attention module to generate the target output features; The target output features are input into the fully connected layer for linear integration and then into the Softmax layer to generate the probability that the sample image belongs to the set category. The category with the highest probability and the corresponding feature are output as the feature extraction result; the sum of the probabilities of all categories is 1; During the training process, the cross entropy loss function is used to calculate the loss value, and the loss curves on the training set and the validation set are monitored. The training is terminated when the evaluation index of the model on the validation set reaches the expected threshold.

6. The method according to claim 5, characterized in that The channel attention module is used to generate channel attention weights based on the target AD feature map, including: Perform global average pooling on the AD feature map to generate the average value of all pixels in each channel in the AD feature map; Perform global maximum pooling operation on the AD feature map to generate the maximum value of all pixels in each channel in the AD feature map; The average value and maximum value are input into the multilayer perceptron for nonlinear transformation, generating high-level feature representation and fusion; The sigmoid function is used to map the fused high-level feature representation to the set interval to generate channel attention weights.

7. The method according to claim 5, characterized in that The fused features are input into the residual unit for multi-scale group convolution operation to generate multi-scale AD features, including: In the fusion feature, all input channels in the w-th input channel group are convolved with all convolution kernels in the a-th convolution kernel group and the output features are fused; Until all input channels in the k input channel groups and all convolution kernels in the t convolution kernel groups are convolved to generate k*t fused output features; Connect k*t fused output features according to the channel dimension to generate multi-scale AD features; where w, k, a, and t are all positive integers, w is less than or equal to k, and a is less than or equal to t.

8. The method according to claim 5, characterized in that The set categories include labeled normal sample images, mild cognitive impairment sample images, and suspected Alzheimer's disease sample images.

Images