A method for predicting the risk of Alzheimer's disease onset based on the CTA framework

Abstract

The present invention discloses a method for predicting the onset risk of Alzheimer's disease based on the CTA framework, comprising the following steps: S1, inputting brain images into a 3D convolutional neural network for feature extraction to obtain brain image feature maps; S2, inputting the brain image feature maps into a Transformer attention module to adaptively adjust the information between different positions of the brain image feature maps to obtain final attention feature maps; S3, inputting the final attention feature maps into a fully connected regression layer after global pooling, and jointly predicting several cognitive scores in combination with the total mean square error loss to obtain the predicted probability of the onset of Alzheimer's disease. The present invention can automatically identify discriminant positions of specific participants in whole brain images, which helps to evaluate the pathological stage of Alzheimer's disease and predict disease progression. Evaluations on the ADNI baseline dataset show that the method of the present invention exhibits good performance in cognitive score estimation.

Description

Technical Field

[0001] The present invention belongs to the field of medical image and intelligent information processing, and particularly relates to a method for predicting the onset risk of Alzheimer's disease based on the CTA framework. Background Art

[0002] Alzheimer's disease (AD) is one of the most common types of dementia, characterized by a gradual decline in cognitive ability and daily function. This neurodegenerative disease usually affects the elderly and is irreversible. The onset of AD severely disrupts the lives of patients and their families, bringing a heavy burden to society and the economy. Currently, there is no effective drug or clinical treatment method to stop or reverse the progression of AD. Therefore, early diagnosis and assessment of AD are crucial for delaying its progression. However, due to the complexity of AD, it is difficult for humans to fully understand its pathogenesis, which poses many challenges to diagnosis and assessment. Nowadays, the diagnosis of AD is mainly carried out through a comprehensive assessment of cognitive behavior assessment and imaging biomarkers. Among them, cognitive scales are used to evaluate the phenotypic characteristics of the cognitive state of subjects, such as the Mini-Mental State Examination (MMSE), Clinical Dementia Rating Scale (CDRS), and Alzheimer's Disease Assessment Scale (ADAS). Due to problems such as the subjective factors of doctors and the individual differences of patients, there are also many subjective biases in the assessment of cognitive impairment.

[0003] With the progress of artificial intelligence and the accumulation of brain imaging data, accurately estimating clinical scores through brain imaging can help understand the pathological stage of AD and predict its progression. As one of the most important imaging techniques in clinical practice, sMRI is widely used in the diagnosis and assessment of AD. Therefore, with the help of AI technology, the AD assessment based on brain sMRI can avoid subjective biases caused by various factors such as doctors or subjects. Currently, many machine learning-based methods usually pre-select dementia-related regions from sMRI data for feature extraction and prediction model construction. However, due to the potential heterogeneity between different steps and the individual specificity in the progression of dementia, these methods may not be able to achieve optimization. Combining deep learning methods for feature encoding and model construction can more effectively achieve task-oriented feature learning. Summary of the Invention

[0004] Aiming at the above deficiencies in the prior art, the method for predicting the onset risk of Alzheimer's disease based on the CTA framework provided by the present invention solves the problem of inaccurate prediction of the onset of Alzheimer's disease in the prior art.

[0005] In order to achieve the above invention object, the technical solution adopted by the present invention is: a method for predicting the onset risk of Alzheimer's disease based on the CTA framework, comprising the following steps:

[0006] S1. Input the brain image into a 3D convolutional neural network for feature extraction to obtain a brain image feature map;

[0007] S2. Input the brain image feature map into a Transformer attention module to adaptively adjust the information between different positions of the brain image feature map to obtain a final attention feature map;

[0008] S3. Convolve the final attention feature map with the brain image feature map. After global pooling of the convolution result, input it into a fully connected regression layer. Through the fully connected regression layer, jointly predict several cognitive scores by combining the total mean square error loss to obtain the prediction probability of Alzheimer's disease onset.

[0009] Furthermore: In the above S1, the 3D convolutional neural network includes the first to fourth convolutional sub-networks connected in sequence;

[0010] Among them, the structures of the first to third convolutional sub-networks are the same, and each includes a first convolutional layer, a second convolutional layer, and a 2×2×2 max pooling layer connected in sequence. The convolutional kernels of the first convolutional layer and the second convolutional layer are both 3×3×3;

[0011] The fourth convolutional sub-network includes a third convolutional layer, a fourth convolutional layer, and a 1×1×1 global average pooling layer connected in sequence. The convolutional kernel of the third convolutional layer is 3×3×3, and the convolutional kernel of the fourth convolutional layer is 1×1×1;

[0012] The outputs of the first to fourth convolutional layers are all connected to a batch normalization layer and a ReLU activation function layer in sequence, which are used to extract and represent the spatial features of the brain image.

[0013] Furthermore: In the above S2, the Transformer attention module includes an encoder, a decoder, a connection layer, and a softmax activation function layer connected in sequence;

[0014] Among them, the encoder includes a first multi-head attention layer and a first forward position feedback network layer. The input end of the first forward position feedback network layer is connected to the output end and the input end of the first multi-head attention layer through normalization and residual connection;

[0015] The decoder includes a masked multi-head attention layer, a second multi-head attention layer, and a second forward position feedback network layer. The input end of the second multi-head attention layer is connected to the output end and the input end of the first forward position feedback network layer, and the output end and the input end of the masked multi-head attention layer through normalization and residual connection. The input end of the second forward position feedback network layer is connected to the output end and the input end of the second multi-head attention layer through normalization and residual connection. The input end of the connection layer is connected to the output end and the input end of the second forward position feedback network layer through normalization and residual connection.

[0016] Furthermore, the structures of the masked multi-head attention layer, the first multi-head attention layer, and the second multi-head attention layer are the same, including three parallel first convolutions, a number of scaled dot-product attention modules, and a second convolution. The input end of each scaled dot-product attention module is connected to the three parallel first convolutions, and the output end of each scaled dot-product attention module is connected to the second convolution.

[0017] Furthermore, S2 includes the following sub-steps:

[0018] S21. Embed the brain image feature map into 3D spatial position information through position encoding to obtain an input feature map;

[0019] S22. Input the input feature map into the first multi-head attention layer to obtain a first attention feature map;

[0020] S23. Input the first attention feature map and the input feature map into the first forward position feedback network layer to obtain the attention feature map of the encoder;

[0021] S24. Input the input feature map into the masked multi-head attention layer to obtain a second attention feature map;

[0022] S25. Input the second attention feature map, the input feature map, and the attention feature map of the encoder into the second multi-head attention layer to obtain a third attention feature map;

[0023] S26. Input the third attention feature map and the second attention feature map into the second forward position feedback network layer to obtain the attention feature map of the decoder;

[0024] S27. Input the attention feature map of the decoder and the third attention feature map into the softmax activation function layer through a connection layer to obtain an attention weight map;

[0025] S28. Fuse the attention weight map and the brain image feature map to obtain a final attention feature map.

[0026] Furthermore, in S21, the position encoding has the same five-dimensional shape parameters as the brain image feature map. The five-dimensional shape parameters include batch size, depth, height, width, and number of channels. The specific expression of the position encoding PE is:

[0027] PE = sin(pos × div_term)

[0028] In the formula, sin(·) is the sine function, pos is the position index tensor, and div_term is the scaling factor.

[0029] Furthermore, the method for the masked multi-head attention layer, the first multi-head attention layer, and the second multi-head attention layer to calculate the output feature map based on the input feature map is the same, specifically as follows:

[0030] SA1. Map the input feature map to queries, keys, and values through three parallel first convolutions;

[0031] SA2. Input the queries, keys, and values into all scaled dot-product attention modules to obtain the calculation results of each scaled dot-product attention module;

[0032] SA3. Concatenate the calculation results of all scaled dot-product attention modules through a second convolution to obtain the output feature map.

[0033] Furthermore, in SA2, the expression for obtaining the calculation result head i of the i-th scaled dot-product attention module is specifically:

[0034] head i = Conv3D(Attention(Q, K, V))

[0035] where Conv3D(·) is a 3D convolution operation, and Attention(Q i , K i , V i ) is the self-attention dot-product matrix, and its expression is specifically:

[0036]

[0037] where softmax(·) is the softmax activation function, Q is the query, K is the key, V is the value, T is the transpose symbol, and d k is the dimension of the key vector;

[0038] In SA3, the expression for the result MultiHead(Q, K, V) after concatenating the calculation results of all scaled dot-product attention modules is specifically as follows:

[0039] MultiHead(Q, K, V) = Conv3D(concat(head1,..., head h ))

[0040] where concat(·) is the concatenation operation, head1 is the calculation result of the first scaled dot-product attention module, and head h is the calculation result of the h-th scaled dot-product attention module.

[0041] Furthermore, in S3, the expression for the total mean squared error loss L reg is specifically:

[0042] L reg = MSE ADAS11 + MSE CDRSB + MSE MMSE

[0043] Wherein, MSE ADAS11 is the mean square error of the cognitive score with the scoring label ADAS11, and MSE CDRSB is the mean square error of the cognitive score with the scoring label CDRSB, and MSE MMSE is the mean square error of the cognitive score with the scoring label MMSE. Among them, the expression for calculating the mean square error MSE of any cognitive score is specifically:

[0044]

[0045] Wherein, N is the number of samples, and y i is the true score value of the nth sample, is the predicted score value of the nth sample.

[0046] The beneficial effects of the present invention are as follows: The present invention provides a method for predicting the risk of Alzheimer's disease based on the CTA framework, which can automatically identify the discriminant positions of specific participants in the whole brain image, seamlessly learn high-level feature representations to construct a regression model for clinical cognitive score estimation. The contributions of the present invention are specifically:

[0047] (1) Use a 3D convolutional neural network to encode the brain image;

[0048] (2) Use the Transformer attention module to allow adaptive fusion of information between different positions or features, and dynamically adjust the weights according to importance. This means that during the encoding of the brain image, the attention mechanism can adaptively fuse information from different brain regions to better represent the structural features of the whole brain.

[0049] (3) Through the fully connected regression layer, combined with the total mean square error loss, jointly predict multiple cognitive scores, and finally improve the prediction accuracy of the clinical score. Description of the Drawings

[0050] Figure 1 is a flowchart of a method for predicting the risk of Alzheimer's disease based on the CTA framework of the present invention.

[0051] Figure 2 is a deep learning architecture diagram integrating CNN and Transformer attention of the present invention.

[0052] Figure 3 is a structural diagram of the 3D convolutional neural network of the present invention.

[0053] Figure 4 This is the structural diagram of the Transformer self-attention module of the present invention.

[0054] Figure 5 This is the structural diagram of the multi-head attention layer in the Transformer self-attention module of the present invention. Specific embodiments

[0055] The following describes the specific embodiments of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0056] Example 1:

[0057] As Figure 1 shown, in an embodiment of the present invention, a method for predicting the risk of Alzheimer's disease based on the CTA framework includes the following steps:

[0058] S1. Input the brain image into a 3D convolutional neural network for feature extraction to obtain a brain image feature map;

[0059] S2. Input the brain image feature map into the Transformer attention module to adaptively adjust the information between different positions of the brain image feature map to obtain the final attention feature map;

[0060] S3. Convolve the final attention feature map with the brain image feature map, and after global pooling of the convolved result, input it into a fully connected regression layer. Through the fully connected regression layer, combined with the total mean square error loss, several cognitive scores are jointly predicted to obtain the predicted probability of the onset of Alzheimer's disease.

[0061] As Figure 2 shown, in this embodiment, the present invention designs a deep learning architecture integrating CNN and Transformer attention (CTA) for estimating multiple cognitive scores from brain images scanned by SMRI. The deep learning architecture includes a 3D convolutional neural network, a Transformer attention module, and a fully connected regression layer for estimating multiple cognitive scores.

[0062] The sampling method of brain images is divided into multiple training batches. The specific sampling strategy for each training batch is as follows: randomly select N subjects from the training dataset to form a batch of training data containing N brain images, and then send N brain images with a size of 121×145×121 into the 3D convolutional neural network for training, and then output a brain image feature map with a size of 9×12×9.

[0063] In S1, the 3D convolutional neural network includes first to fourth convolutional sub-networks connected in sequence;

[0064] Among them, the structures of the first to third convolutional sub-networks are the same, and each includes a first convolutional layer, a second convolutional layer, and a 2×2×2 max pooling layer connected in sequence. The convolutional kernels of the first convolutional layer and the second convolutional layer are both 3×3×3;

[0065] The fourth convolutional sub-network includes a third convolutional layer, a fourth convolutional layer, and a 1×1×1 global average pooling layer connected in sequence. The convolutional kernel of the third convolutional layer is 3×3×3, and the convolutional kernel of the fourth convolutional layer is 1×1×1;

[0066] The outputs of the first to fourth convolutional layers are all connected to a batch normalization layer and a ReLU activation function layer in sequence, which are used to extract and represent the spatial features of the brain images.

[0067] In this embodiment, the structure of the 3D convolutional neural network is as Figure 3 shown. A batch normalization (BatchNorm) and a ReLU activation function are used after each convolutional layer of the 3D convolutional neural network to effectively extract and represent the spatial features of the brain images. Specifically, the convolutional layer uses a 3×3×3 convolutional kernel to capture local features in three-dimensional space, and the final 1×1×1 convolutional kernel convolutional layer is used to integrate the information in the feature map. The 2×2×2 max pooling layer reduces the size of the feature map through downsampling operations while retaining important spatial information. The 1×1×1 global average pooling (GAP) layer summarizes the global information of each feature map to provide a compact feature representation; a batch normalization (BatchNorm) layer is used after each convolutional layer to stabilize the training process and accelerate convergence. The ReLU activation function is applied after each convolutional layer to introduce non-linearity, thereby enhancing the expressive power of the model. This architecture design not only ensures that the depth and complexity of the model are appropriate, but also retains sufficient feature extraction capabilities, which helps to better capture and represent the key information in the brain images.

[0068] In S2, the Transformer attention module includes an encoder, a decoder, a connection layer, and a softmax activation function layer connected in sequence;

[0069] Among them, the encoder includes a first multi-head attention layer and a first forward position feedback network layer. The input end of the first forward position feedback network layer is connected to the output end and the input end of the first multi-head attention layer through normalization and residual connection;

[0070] The decoder includes a masked multi-head attention layer, a second multi-head attention layer, and a second forward position feedback network layer. The input end of the second multi-head attention layer is connected to the output end and the input end of the first forward position feedback network layer, the output end and the input end of the masked multi-head attention layer through normalization and residual connection. The input end of the second forward position feedback network layer is connected to the output end and the input end of the second multi-head attention layer through normalization and residual connection. The input end of the connection layer is connected to the output end and the input end of the second forward position feedback network layer through normalization and residual connection.

[0071] In this embodiment, the structure of the Transformer attention module is as Figure 4 shown. In the multi-head attention layer of the encoder, the input feature map is mapped into queries, keys, and values through three parallel convolutions, and the attention-based image feature representation is obtained by calculating the attention weights of each head. To enhance the stability and representation ability of the model, these output features are subjected to batch normalization, activation function processing, and residual connection. In the decoder, by receiving the attention feature map from the encoder and the original input feature map as inputs, feature fusion and restoration are performed. In the multi-head attention layer of the decoder, the features are calculated and fused, and then mapped back to the high-dimensional space through restoration and deconvolution operations to generate a feature map with the same size as the input image. During this process, the decoder uses skip connections to fuse shallow and deep features, thereby obtaining a more discriminative feature representation. Through this multi-level feature fusion and deconvolution operation, the decoder not only realizes feature fusion and restoration, but also optimizes the feature representation and enhances the performance of the model in MRI analysis.

[0072] As Figure 5 shown, the structures of the masked multi-head attention layer, the first multi-head attention layer, and the second multi-head attention layer are the same, including three parallel first convolutions, several scaled dot-product attention modules, and a second convolution. In this embodiment, h are set, and the input end of each scaled dot-product attention module is connected to the three parallel first convolutions, and the output end of each scaled dot-product attention module is connected to the second convolution.

[0073] S2 includes the following sub-steps:

[0074] S21. Embed the brain image feature map into 3D spatial position information through position encoding to obtain the input feature map;

[0075] S22. Input the input feature map into the first multi-head attention layer to obtain the first attention feature map;

[0076] S23. Input the first attention feature map and the input feature map into the first forward position feedback network layer to obtain the attention feature map of the encoder;

[0077] S24. Input the input feature map into the masked multi-head attention layer to obtain a second attention feature map;

[0078] S25. Input the second attention feature map, the input feature map, and the attention feature map of the encoder into the second multi-head attention layer to obtain a third attention feature map;

[0079] S26. Input the third attention feature map and the second attention feature map into the second forward position feedback network layer to obtain the attention feature map of the decoder;

[0080] S27. Input the attention feature map of the decoder and the third attention feature map through a connection layer into the softmax activation function layer to obtain an attention weight map;

[0081] S28. Fuse the attention weight map and the brain image feature map to obtain the final attention feature map.

[0082] In S21, the positional encoding has the same five-dimensional shape parameters as the brain image feature map. The five-dimensional shape parameters include batch size N, depth D, height H, width W, and number of channels C. The expression of the positional encoding PE is specifically:

[0083] PE = sin(pos × div_term)

[0084] In the formula, sin(·) is the sine function, pos is the position index tensor, whose range is from 0 to C / 2 - 1, and div_term is the scaling factor, which is calculated through the exponential function.

[0085] The positional encoding PE can also be calculated through the cosine function, specifically as the following formula:

[0086] PE = cos(pos × div_term)

[0087] The methods for the masked multi-head attention layer, the first multi-head attention layer, and the second multi-head attention layer to calculate the output feature map based on the input feature map are the same, specifically:

[0088] SA1. Map the input feature map into queries, keys, and values through three parallel first convolutions;

[0089] SA2. Input the queries, keys, and values into all scaled dot-product attention modules to obtain the calculation results of each scaled dot-product attention module;

[0090] SA3. Concatenate the calculation results of all scaled dot-product attention modules through a second convolution to obtain the output feature map.

[0091] In this embodiment, the scaled dot-product attention module can capture the queries, keys, and values. After the 3D convolution operation, the queries, keys, and values are reduced to a spatial dimension of 32.

[0092] In the SA2, the calculation result head of the i-th scaled dot-product attention module is obtained. i The specific expression is:

[0093] head i = Conv3D(Attention(Q, K, V))

[0094] In the formula, Conv3D(·) is a 3D convolution operation, and Attention(Q i , K i , V i ) is a self-attention dot-product matrix, and its specific expression is:

[0095]

[0096] In the formula, softmax(·) is the softmax activation function, Q is the query, K is the key, V is the value, T is the transpose symbol, and d k is the dimension of the key vector;

[0097] In the SA3, the expression of the result MultiHead(Q, K, V) after concatenating the calculation results of all scaled dot-product attention modules is specifically the following formula:

[0098] MultiHead(Q, K, V) = Conv3D(concat(head1,..., head h ))

[0099] In the formula, concat(·) is the concatenation operation, head1 is the calculation result of the 1st scaled dot-product attention module, and head h is the calculation result of the h-th scaled dot-product attention module.

[0100] Due to the characteristics of 3D images, the Transformer attention module does not flatten the 3D image into two dimensions for calculation. Instead, it uses a 3D convolution kernel to replace the weight matrix. Each head in each scaled dot-product attention module performs multiple convolutions using different convolution kernels, and the size of each convolution kernel is set to 1. After completing the 3D convolution calculations of Q, K, and V, a deconvolution operation is performed to restore the original dimension of the image features.

[0101] In the S3, the expression of the total mean squared error loss L reg is specifically:

[0102] L reg = MSE ADAS11 + MSE CDRSB + MSE MMSE

[0103] wherein, MSE ADAS11 is the mean square error of the cognitive score with the scoring label ADAS11, and MSE CDRSB is the mean square error of the cognitive score with the scoring label CDRSB, and MSE MMSE is the mean square error of the cognitive score with the scoring label MMSE. Among them, the expression for calculating the mean square error MSE of any cognitive score is specifically:

[0104]

[0105] wherein, N is the number of samples, and y i is the true score value of the nth sample, is the predicted score value of the nth sample.

[0106] In step S4, during the process of jointly predicting each cognitive score, the final attention feature map input includes samples of N subjects, denoted as {(X n , y n )} N , where X n represents the nth sample, represents T cognitive scores. To achieve this goal, the total mean square error loss L reg is the sum of the mean square errors of the cognitive scores of three different scoring labels in the regression task. The total mean square error loss designed in the present invention will optimize the CTA network through backpropagation and integrate its effect into the Transformer attention to assist the network in learning.

[0107] Embodiment 2:

[0108] This embodiment is directed to the experimental data provided in Embodiment 1 to verify the effectiveness of the present invention in improving the prediction accuracy of clinical scores.

[0109] In this embodiment, the experiment is carried out using the PyTorch platform. The network is trained using the Adam optimizer, the batch size is 4, and the dropout rate is 0.3. In the training stage, a total of 70 epochs are carried out, and a regression task of scoring estimation is carried out using three scoring labels (ADAS11, CDRSB, MMSE).

[0110] In the ADNI database, the irregularity of cognitive impairment diseases leads to significant differences in the neuroimages of different subjects. Therefore, for score estimation, two independent sub-datasets (dataset1 and dataset2) were used for validation. Specifically, two sub-datasets were constructed from the obtained ADNI database and split into training-validation-test at a ratio of 70%, 15%, and 15%. In addition, the subjects in the two test sets did not overlap, which enhanced the evaluation of the generalization ability of this method.

[0111] The present invention evaluates the performance of the proposed method by assessing three clinical cognitive scores including CDRSB, ADAS11, and MMSE. The prediction performance is quantified using the correlation coefficient (CC) and the root mean square error (RMSE). The present invention compares six different methods, including ResNet and CNN backbone baseline networks. In addition, a spatial attention block is applied to these baseline networks respectively to construct the Sa-ResNet and Sa-CNN methods. The present invention also includes the Transformer attention block without position encoding method (CTA without PE) and the proposed method (CTA with PE). The ResNet method uses 3D ResNet10 as the CNN encoder for brain images, and then performs regression prediction through three fully connected layers. The CNN method uses a 3DCNN backbone without an attention block to extract brain image features, and then performs regression prediction through three fully connected layers. In the Sa-ResNet, Sa-CNN, and CTA methods, the attention block is applied to the last layer of the 3D CNN backbone. In the above attention-based methods, the feature map output by the backbone network is encoded by the attention block and then fused with the original feature map through dot product fusion. Through these comparative experiments, the present invention can evaluate the effectiveness of the proposed Transformer block in improving the accuracy of cognitive score estimation.

[0112] Table 1 Prediction results of the model on dataset 1

[0113]

[0114] Table 2 Prediction results of the model on dataset 2

[0115]

[0116] Table 3 Comparison results include traditional learning methods and deep learning methods based on the ADNI dataset

[0117]

[0118] The ablation experiment results of the score estimation for two sub-datasets of the ADNI dataset are shown in Table 1 and Table 2. By comparing and analyzing the experimental data obtained from six different methods, the performance of the proposed method under different configurations can be evaluated, and the following conclusions can be drawn.

[0119] First of all, the experimental results clearly show that the method of integrating the attention mechanism in the deep learning model performs better on both datasets than those methods without using the attention mechanism. This finding indicates that applying the attention module to the feature map after feature extraction can effectively enhance the model's ability to represent the spatial information of MRI images, thus contributing to the construction of a more accurate regression model. Secondly, the performance of the CTA without PE method exceeds that of the method based on spatial attention. This shows that Transformer attention can supplement local information and improve the model performance by effectively focusing on specific regions of brain images. Finally, the performance of the CTA method with spatial position encoding is better than that of the CTA without PE method. This result emphasizes the importance of spatial position encoding in enhancing the efficiency of the Transformer attention module in processing 3D MRI. Spatial position encoding not only preserves the three-dimensional spatial information of the image, enhances the model's understanding of the image's spatial structure and context relationship, but also improves the ability of feature representation. Therefore, introducing this mechanism significantly enhances the model's ability to capture and utilize the spatial information of SMRI, thereby improving the accuracy of cognitive score prediction.

[0120] In addition, in Table 3, the method of the present invention is compared with some traditional learning methods and deep learning methods based on the ADNI dataset to demonstrate the effectiveness of the proposed method. The experimental results in Table 3 represent the average results of two different test sets.

[0121] The beneficial effects of the present invention are as follows: The present invention provides a method for predicting the onset risk of Alzheimer's disease based on the CTA framework, which can automatically identify the discriminant positions of specific participants in the whole brain image and seamlessly learn high-level feature representations to construct a regression model for clinical cognitive score estimation. The contributions of the present invention are specifically as follows:

[0122] (1) Use a 3D convolutional neural network to encode brain images;

[0123] (2) Use the Transformer attention module to allow the adaptive fusion of information between different positions or features, and dynamically adjust the weights according to importance, which means that during the encoding of brain images, the attention mechanism can adaptively fuse information from different brain regions to better represent the structural features of the whole brain.

[0124] (3) Jointly predict multiple cognitive scores through the fully connected regression layer combined with the total mean square error loss, and finally improve the prediction accuracy of clinical scores.

[0125] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of technical features. Therefore, the features defined by "first", "second", "third" may explicitly or implicitly include one or more of such features.

Claims

1. A method for predicting the onset risk of Alzheimer's disease based on the CTA framework, characterized in that, It includes the following steps: S1. Input the brain image into a 3D convolutional neural network for feature extraction to obtain a brain image feature map; S2. Input the brain image feature map into a Transformer attention module to adaptively adjust the information between different positions of the brain image feature map, and obtain a final attention feature map; S3. Convolve the final attention feature map with the brain image feature map. After global pooling of the convolution result, input it into a fully connected regression layer, and jointly predict several cognitive scores through the fully connected regression layer combined with the total mean square error loss to obtain the prediction probability of Alzheimer's disease onset; In S2, the Transformer attention module includes an encoder, a decoder, a connection layer, and a softmax activation function layer connected in sequence; Among them, the encoder includes a first multi-head attention layer and a first forward position feedback network layer. The input end of the first forward position feedback network layer is connected to the output end and the input end of the first multi-head attention layer through normalization and residual connection; The decoder includes a masked multi-head attention layer, a second multi-head attention layer, and a second forward position feedback network layer. The input end of the second multi-head attention layer is connected to the output end and the input end of the first forward position feedback network layer, as well as the output end and the input end of the masked multi-head attention layer through normalization and residual connection. The input end of the second forward position feedback network layer is connected to the output end and the input end of the second multi-head attention layer through normalization and residual connection. The input end of the connection layer is connected to the output end and the input end of the second forward position feedback network layer through normalization and residual connection; The masked multi-head attention layer, the first multi-head attention layer, and the second multi-head attention layer have the same structure, including three parallel first convolutions, several scaled dot-product attention modules, and a second convolution. The input end of each scaled dot-product attention module is connected to the three parallel first convolutions, and the output end of each scaled dot-product attention module is connected to the second convolution; S2 includes the following sub-steps: S21. Embed the brain image feature map into 3D spatial position information through position encoding to obtain an input feature map; S22. Input the input feature map into the first multi-head attention layer to obtain a first attention feature map; S23. Input the first attention feature map and the input feature map into the first forward position feedback network layer to obtain the attention feature map of the encoder; S24. Input the input feature map into the masked multi-head attention layer to obtain a second attention feature map; S25. Input the second attention feature map, the input feature map, and the attention feature map of the encoder into the second multi-head attention layer to obtain a third attention feature map; S26. Input the third attention feature map and the second attention feature map into the second forward position feedback network layer to obtain the attention feature map of the decoder; S27. Input the attention feature map of the decoder and the third attention feature map into the softmax activation function layer through the connection layer to obtain an attention weight map; S28. Fuse the attention weight map and the brain image feature map to obtain the final attention feature map.

2. The method for predicting the onset risk of Alzheimer's disease based on the CTA framework according to claim 1, wherein In S1, the 3D convolutional neural network includes first to fourth convolutional sub-networks connected in sequence; Among them, the structures of the first to third convolutional sub-networks are the same, and each includes a first convolutional layer, a second convolutional layer, and a 2×2×2 max pooling layer connected in sequence. The convolutional kernels of the first convolutional layer and the second convolutional layer are both 3×3×3; The fourth convolutional sub-network includes a third convolutional layer, a fourth convolutional layer, and a 1×1×1 global average pooling layer connected in sequence. The convolutional kernel of the third convolutional layer is 3×3×3, and the convolutional kernel of the fourth convolutional layer is 1×1×1; The outputs of the first to fourth convolutional layers are all connected to a batch normalization layer and a ReLU activation function layer in sequence, which are used to extract and represent the spatial features of brain images.

3. The method for predicting the onset risk of Alzheimer's disease based on the CTA framework according to claim 1, wherein In S21, the positional encoding is the same as the five-dimensional shape parameters of the brain image feature map. The five-dimensional shape parameters include batch size, depth, height, width, and number of channels. The expression of the positional encoding PE is specifically as follows: where sin(·) is the sine function, pos is the position index tensor, div_term is the scaling factor.

4. The method for predicting the onset risk of Alzheimer's disease based on the CTA framework according to claim 3, wherein The method by which the masked multi-head attention layer, the first multi-head attention layer, and the second multi-head attention layer calculate the output feature map based on the input feature map is the same. Specifically: SA1. Map the input feature map to queries, keys, and values through three parallel first convolutions; SA2. Input the queries, keys, and values into all scaled dot-product attention modules to obtain the calculation results of each scaled dot-product attention module; SA3. Concatenate the calculation results of all scaled dot-product attention modules through a second convolution to obtain the output feature map.

5. The method for predicting the onset risk of Alzheimer's disease based on the CTA framework according to claim 4, wherein In the SA2, obtain the calculation result of the $i$-th scaled dot-product attention module head i The specific expression is as follows: where Conv3D(·) is a 3D convolution operation, is the self-attention dot product matrix, and its expression is specifically: In the formula, softmax (·) is the softmax activation function, Q is the query, K is the key, V is the value, T is the transpose symbol, d k is the dimension of the key vector; In the SA3, the result after concatenating the calculation results of all scaled dot-product attention modules The specific expression is as follows: In the formula, is the splicing operation, head 1 is the calculation result of the first scaled dot-product attention module, head h is the h calculation result of the scaled dot-product attention module.

6. The method for predicting the onset risk of Alzheimer's disease based on the CTA framework according to claim 1, wherein In S3, the total mean squared error loss L reg has the following specific expression: In the formula, is the mean square error of the cognitive score with the scoring label ADAS11, is the mean square error of the cognitive score with the scoring label CDRSB, is the mean square error of the cognitive score with the scoring label MMSE. Among them, the expression for calculating the mean square error of any cognitive score MSE is specifically as follows: In the formula, N is the number of samples, y i is the true rating value of the n th sample, is the predicted rating value of the n th sample.

Images