Alzheimer's disease course prediction method and system based on multi-scale features

By using improved ResNet block and multi-scale feature fusion methods in Alzheimer's disease course prediction, the problems of insufficient accuracy and high computational cost are solved, and higher diagnostic accuracy and lower computational complexity are achieved. Suitable for clinical auxiliary diagnosis.

CN119028585APending Publication Date: 2024-11-26HUAQIAO UNIVERSITY

Patent Information

Application Number
CN202411155731.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The identification accuracy of Alzheimer's disease course in the prior art is insufficient and the calculation cost is high, which limits its widespread use in clinical practice.

Method used

Using the Alzheimer's disease course prediction method based on multi-scale features, the improved ResNet block uses 3D hollow convolution and attention mechanism to extract features of different scales and perform multi-scale feature fusion to reduce the amount of parameters and reduce the calculation cost.

Benefits of technology

It improves the diagnostic accuracy of the Alzheimer's disease course and the accuracy of localizing pathological areas, while reducing the complexity of the model and computational cost, and is suitable for environments with insufficient computing resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028585A_ABST
    Figure CN119028585A_ABST
Patent Text Reader

Abstract

The invention discloses an Alzheimer's disease course prediction method and system based on multi-scale features, and the method comprises the following steps: obtaining and preprocessing an original brain image, and obtaining a three-dimensional whole-brain image as a to-be-predicted image; inputting a to-be-predicted image into the Alzheimer's disease 3D feature extraction model to realize disease course prediction; the Alzheimer's disease 3D feature extraction model comprises an input layer, a plurality of improved ResNet blocks arranged in sequence, an average pooling module and a full connection module; according to the improved ResNet block, feature extraction of different scales is carried out by adopting 3D cavity convolution, and multi-scale feature fusion is carried out by adopting an attention mechanism. Brain image features are extracted through a multi-scale method, important pathological features of a disease can be captured more accurately, and the accuracy of locating a pathological area is improved; brain image features of different scales are fused through multi-scale features, expression of important features can be better enhanced, irrelevant information can be inhibited, and the disease course state of the Alzheimer's disease can be effectively diagnosed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence model technology, and in particular to a method and system for predicting the course of Alzheimer's disease based on multi-scale features. Background Art

[0002] Alzheimer's disease is a progressive neurodegenerative disease with an insidious onset. Artificial intelligence technology can help doctors diagnose Alzheimer's disease more accurately by analyzing and learning from a large amount of neuroimaging data to identify subtle changes and patterns. Compared with traditional diagnostic methods, artificial intelligence can detect potential signs of lesions more quickly and accurately. Since the pathology of Alzheimer's disease is complex, different patients may have different pathological areas, and the degree of lesions in different pathological areas is also different. Therefore, considering features of different scales when designing auxiliary diagnosis algorithms may help the model capture pathological areas of different scales. At the same time, the attention mechanism is similar to the human visual mechanism. In the field of computer vision, it is usually manifested as the ability to actively focus on important areas in the image and ignore those unimportant areas, which improves the interpretability of the model. In the medical field, the interpretability and accuracy of the model are crucial for doctors and patients. The Chinese patent application with publication number CN114820524A discloses a 3D feature recognition method for Alzheimer's disease in MRI images. The method adds a 3D asymmetric convolution block to each ResNet block, and the 3D asymmetric convolution block extracts discriminative features in the 3D MRI image; an attention feature fusion addition module is set after each ResNet block to fuse local and global feature contexts and perform multi-scale channel attention feature fusion; after the feature extraction of multiple ResNet blocks, the adaptive average pooling module and the fully connected module are used to output the final recognition result. This technical solution not only prevents the loss of feature information, but also improves the extraction of more discriminative features. It also combines the characteristics of Alzheimer's disease image MRI to better integrate features with various semantics and scales in the attention mechanism, and improves the fusion of 3D features, which further improves the accuracy of feature recognition of Alzheimer's disease. However, the asymmetric convolution used in the model of this method is slightly insufficient in the ability to extract large-scale features, resulting in limited improvement in accuracy, and the computational cost and number of parameters of the model are high, which may limit its widespread use in clinical practice. Summary of the invention

[0003] The purpose of the present invention is to solve the problem of insufficient accuracy and high computational cost in identifying the course of Alzheimer's disease in the prior art.

[0004] The technical solution adopted by the present invention to solve the technical problem is: to provide a method for predicting the course of Alzheimer's disease based on multi-scale features, comprising the following steps:

[0005] Acquire the original brain image and perform preprocessing to obtain a three-dimensional whole-brain image as the image to be predicted;

[0006] The image to be predicted is input into the Alzheimer's disease 3D feature extraction model to achieve disease course prediction;

[0007] The Alzheimer's disease 3D feature extraction model includes an input layer, multiple improved ResNet blocks arranged in sequence, an average pooling module and a fully connected module; the input layer receives the image to be predicted and extracts preliminary features; the preliminary features are sequentially input into multiple improved ResNet blocks, multiple feature extractions and fusions are performed, and fused features are output; the fused features are input into the average pooling module and the fully connected module to obtain the disease course prediction results; the improved ResNet block uses 3D hole convolution to extract features of different scales, and uses the attention mechanism to perform multi-scale feature fusion.

[0008] Preferably, the improved ResNet block includes a 3×3×3 convolutional layer and a multi-scale feature module, and the process of extracting and fusing multi-scale features by the multi-scale feature module includes the following steps:

[0009] 3D dilated convolution with different dilation rates is used to extract features of different scales in the input features;

[0010] The attention mechanism is used to fuse multi-scale channel attention features of features of different scales to obtain fused features.

[0011] Preferably, the extracting features of different scales from input features using 3D dilated convolutions with different dilation rates comprises the following steps:

[0012] For input features Use a 1×1×1 convolutional layer to reduce the dimension, and get C, H, W, D represent the number of channels, height, width, and depth of X;

[0013] F is input into a 1×1×1 convolution and several dilated convolutions with different dilation rates to obtain feature maps of different scales.

[0014] Preferably, F is input into a 1×1×1 convolution and several dilated convolutions with different dilated rates, respectively, as follows:

[0015] F is input into a 1×1×1 convolution and three dilated convolutions with dilation rates of 1, 2, and 3, respectively, to obtain feature maps of four different scales.

[0016] Preferably, the attention mechanism is used to perform multi-scale channel attention feature fusion on features of different scales, features of the same scale are sorted by importance to obtain initial weights, and all features are sorted by importance to obtain corresponding weights of features of different scales.

[0017] Preferably, the use of the attention mechanism to perform multi-scale channel attention feature fusion on features of different scales comprises the following steps:

[0018] Use pooling operations to calculate the initial weights of features of different scales;

[0019] After concatenating the initial weights, they are fed into a 1×1×1 convolution, and the Sigmoid function is used to obtain the importance of each channel, expressed as:

[0020]

[0021] Among them, W n Feature M n The corresponding initial weights are represents a 1×1×1 convolution operation, and N represents the number of features;

[0022] The importance W′ is divided to obtain the corresponding weights W of features of different scales n ′;

[0023] The features of different scales are weighted using the corresponding weights, and then all weighted results are added element by element to obtain the fused feature Y, which is expressed as:

[0024]

[0025] Among them, W n ′ and M n represents a set of corresponding weights and features, and ⊙ represents a weighted operation.

[0026] Preferably, the pooling operation is used to calculate the initial weights of features of different scales, and two processing channels are used to receive the features M n The output of the two processing channels is then used with the Sigmoid function to obtain the corresponding feature M n The initial weight W n ; Two processing channels include:

[0027] The first processing channel includes a global average pooling layer, a 1×1×1 convolutional layer, a ReLU activation layer, and a 1×1×1 convolutional layer connected in sequence;

[0028] The second processing channel includes a global maximum pooling layer, a 1×1×1 convolutional layer, a ReLU activation layer, and a 1×1×1 convolutional layer connected in sequence.

[0029] Preferably, the fused features are subjected to a pooling operation before entering the next multi-scale feature module or the average pooling module to refine the discriminative features.

[0030] Preferably, the fused features pass through two residual layers before entering the average pooling module, and each residual layer contains two 3×3×3 convolutions.

[0031] The present invention also provides an Alzheimer's disease course prediction system based on multi-scale features, comprising:

[0032] An image acquisition module acquires the original brain image and performs preprocessing to obtain a three-dimensional whole-brain image as the image to be predicted;

[0033] The image prediction module inputs the image to be predicted into the Alzheimer's disease 3D feature extraction model to achieve disease course prediction;

[0034] The Alzheimer's disease 3D feature extraction model includes an input layer, multiple improved ResNet blocks arranged in sequence, an average pooling module and a fully connected module; the input layer receives the image to be predicted and extracts preliminary features; the preliminary features are sequentially input into multiple improved ResNet blocks, multiple feature extractions and fusions are performed, and fused features are output; the fused features are input into the average pooling module and the fully connected module to obtain the disease course prediction results; the improved ResNet block uses 3D hole convolution to extract features of different scales, and uses the attention mechanism to perform multi-scale feature fusion.

[0035] The present invention has the following beneficial effects:

[0036] (1) The present invention performs data preprocessing on three-dimensional brain image data, which can effectively standardize brain image data; the three-dimensional whole brain image is input into the model and dilated convolution with different dilation rates is used to extract discriminative features, which not only extracts more discriminative features, but also takes into account the spatial relationship of Alzheimer's brain images, and can more accurately capture the important pathological features of the disease, thereby improving the accuracy of locating the pathological area;

[0037] (2) When the present invention uses the attention mechanism to fuse different features, it not only obtains the attention weight on a single scale feature map, but also uses the attention mechanism to fuse features of different scales, refines the discriminative features, and improves the diagnostic accuracy of the model;

[0038] (3) The present invention uses a multi-scale feature fusion method to fuse brain image features at different scales, which can better enhance the expression of important features while suppressing irrelevant information, thereby effectively diagnosing the course of Alzheimer's disease;

[0039] (4) The present invention uses a dilated convolution, which can reduce the number of parameters without reducing the receptive field compared to other convolutions; the present invention uses a multi-scale feature module to replace the third module of the original resnet, and then halves the number of channels of the network to achieve the purpose of lightweight; the use of a lightweight network can reduce model complexity and computational cost, and can be used in an environment with scarce computing resources as an important indicator to assist in the prediction of Alzheimer's disease, improve doctors' control over the disease and thus formulate a diagnostic plan early; at the same time, the present invention can ensure stable performance.

[0040] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A method step diagram of an embodiment of the present invention;

[0042] Figure 2 This is a structural diagram of a 3D feature extraction model for Alzheimer's disease according to an embodiment of the present invention;

[0043] Figure 3 This is a structural diagram of a multi-scale feature module according to an embodiment of the present invention;

[0044] Figure 4 4 is a system structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0045] See also Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0046] S101, obtaining an original brain image and performing preprocessing to obtain an image to be predicted;

[0047] S102, the image to be predicted is input into the Alzheimer's disease 3D feature extraction model to achieve disease course prediction;

[0048] Specifically, the preprocessing described in S101 is to extract brain region morphological features from brain image data, including skull, registration, segmentation, standardization, smoothing and brain region feature extraction; specifically including: stripping all skulls from the image, then registering the images with each other, segmenting the main structures of the brain; normalizing and smoothing all images to a standard size; then extracting the volume, cortical thickness and surface area features of the region of interest from the Alzheimer's disease image, including the standard deviation of cortical thickness, the mean value of cortical thickness, white matter volume, cortical surface area and cortical volume. The method of this embodiment is based on the above medical morphological features, and analyzes the ability of each medical morphological feature to identify the development process of Alzheimer's disease, thereby improving the balance of data distribution and the performance of the algorithm.

[0049] For details, see Figure 2 As shown, the Alzheimer's disease 3D feature extraction model includes an input layer, multiple improved ResNet blocks arranged in sequence, and an average pooling module; the 3D MRI image is received by the input layer, and a 7×7×7 convolution is used to initially extract features; each improved ResNet block includes a 3×3×3 convolution layer and a multi-scale feature module, and the 3*3*3 convolution is the convolution provided by the original 3Dresnet block, which is used to capture the spatial features in the volume data; the multi-scale feature module includes two parts: multi-scale feature extraction and multi-scale feature fusion. The multi-scale feature extraction part uses 3D hole convolutions with different hole rates to extract spatial information features and discriminative features in the 3D MRI image. After each multi-scale feature extraction part, a multi-scale feature attention fusion part is set to fuse the extracted different scale features and global feature contexts, and perform multi-scale channel attention feature fusion; after feature extraction and fusion of multiple improved ResNet blocks, the global average pooling module and the fully connected module are used to output the final diagnosis result.

[0050] For details, see Figure 3 The figure shows the structure of the multi-scale feature module, which is divided into a multi-scale feature extraction part and a multi-scale feature fusion part. The multi-scale feature extraction part selects three different scales of hole convolution and a pixel-by-pixel convolution to extract features from the processed image according to the deep neural network method. The multi-scale feature fusion part inputs the acquired features into a weight adaptive module based on the attention mechanism to obtain weights, giving greater weights to important features while suppressing irrelevant information. The weights are combined together and processed by the activation function before being segmented again and fused with the extracted features.

[0051] Specifically, the multi-scale feature extraction part is an input (C, H, W, D represent the number of channels, height, width and length of X) using a 1×1×1 convolutional layer to reduce the dimension. Then F is input into a 1×1×1 and three dilated convolutions with dilation rates of 1, 2, and 3 respectively to obtain feature maps of four different scales.

[0052] Specifically, in order to fuse features of different scales, the present invention ranks the features of the same scale by importance and ranks all features by importance. Taking feature M1 as an example, the calculation process of the initial weight is explained. In the multi-scale feature fusion part, the feature map is input into the weight adaptation module, and then input into a global average pooling layer and a global maximum pooling layer for pooling operation to obtain the global information description G. avg and G max ; Then G avg and G maxThe input is sent to two stacked 1×1×1 convolutional layers (with a ReLU activation layer in the middle) to model the relationship between different channels of the feature M1 of this scale. Finally, the Sigmoid function is used to obtain the importance of each feature map, that is, the initial weight W1.

[0053] Specifically, four sets of initial weights are concatenated and then fed into an additional 1×1×1 convolution and all values ​​are normalized to between 0 and 1 using the Sigmoid function to obtain the importance of each channel, expressed as:

[0054]

[0055] Among them, W1, W2, W3 and W4 represent the four groups of initial weights corresponding to the four groups of features. Represents a 1×1×1 convolution operation.

[0056] All the obtained feature importances W′ are divided to obtain 4 groups of weights corresponding to the features of 4 groups of different scales. The weights and corresponding features are weighted respectively, and then all the results are added element by element to obtain the output Y as the fusion feature, which is expressed as:

[0057]

[0058] The fused feature Y is input into two residual layers, which contain two 3×3×3 convolutions to refine the discriminative features. Finally, the corresponding Alzheimer's disease diagnosis result is obtained after the fully connected layer algorithm and activation function.

[0059] See also Figure 4 FIG. 1 is a structural diagram of a system for predicting the course of Alzheimer's disease based on multi-scale features according to an embodiment of the present invention, comprising:

[0060] An image acquisition module 401 acquires and preprocesses the original brain image to obtain a three-dimensional whole-brain image as an image to be predicted;

[0061] Image prediction module 401, inputs the image to be predicted into the Alzheimer's disease 3D feature extraction model to achieve disease course prediction;

[0062] The Alzheimer's disease 3D feature extraction model includes an input layer, multiple improved ResNet blocks arranged in sequence, an average pooling module and a fully connected module; the input layer receives the image to be predicted and extracts preliminary features; the preliminary features are sequentially input into multiple improved ResNet blocks, multiple feature extractions and fusions are performed, and fused features are output; the fused features are input into the average pooling module and the fully connected module to obtain the disease course prediction results; the improved ResNet block uses 3D hole convolution to extract features of different scales, and uses the attention mechanism to perform multi-scale feature fusion.

[0063] The present invention extracts brain image features through a multi-scale method, which can more accurately capture the important pathological features of the disease and improve the accuracy of locating the pathological area; by fusing brain image features of different scales through a multi-scale feature fusion method, the expression of important features can be better enhanced while irrelevant information is suppressed, thereby effectively diagnosing the course of Alzheimer's disease; it can serve as an important indicator to assist in the prediction of Alzheimer's disease, improve doctors' control over the disease and thus formulate diagnostic plans early, while ensuring stable performance.

[0064] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the course of Alzheimer's disease based on multi-scale features. It is characterized in that The following steps are involved: Acquire the original brain image and perform preprocessing to obtain a three-dimensional whole-brain image as the image to be predicted; The image to be predicted is input into the Alzheimer's disease 3D feature extraction model to achieve disease course prediction; The Alzheimer's disease 3D feature extraction model includes an input layer, a plurality of sequentially arranged improved ResNet blocks, an average pooling module, and a fully connected module; the input layer receives the image to be predicted and extracts preliminary features; the preliminary features are sequentially input into a plurality of improved ResNet blocks, multiple feature extractions and fusions are performed, and fusion features are output; The fused features are input into the average pooling module and the fully connected module to obtain the disease course prediction result; the improved ResNet block uses 3D hole convolution to extract features of different scales, and uses the attention mechanism to perform multi-scale feature fusion.

2. The method for predicting the course of Alzheimer's disease based on multi-scale features according to claim 1, It is characterized in that The improved ResNet block includes a 3×3×3 convolutional layer and a multi-scale feature module. The process of extracting and fusing multi-scale features by the multi-scale feature module includes the following steps: 3D dilated convolution with different dilation rates is used to extract features of different scales in the input features; The attention mechanism is used to fuse multi-scale channel attention features of features of different scales to obtain fused features.

3. The method for predicting the course of Alzheimer's disease based on multi-scale features according to claim 2, It is characterized in that The method of extracting features of different scales from input features by using 3D dilated convolution with different dilation rates includes the following steps: For input features Use a 1×1×1 convolutional layer to reduce the dimension, and get C, H, W, D represent the number of channels, height, width, and depth of X; F is input into a 1×1×1 convolution and several dilated convolutions with different dilation rates to obtain feature maps of different scales.

4. The method for predicting the course of Alzheimer's disease based on multi-scale features according to claim 3, It is characterized in that The method of inputting F into a 1×1×1 convolution and several dilated convolutions with different dilated rates is as follows: F is input into a 1×1×1 convolution and three dilated convolutions with dilation rates of 1, 2, and 3, respectively, to obtain feature maps of four different scales.

5. The method for predicting the course of Alzheimer's disease based on multi-scale features according to claim 2, It is characterized in that The attention mechanism is used to perform multi-scale channel attention feature fusion on features of different scales, features of the same scale are ranked by importance to obtain initial weights, and all features are ranked by importance to obtain corresponding weights of features of different scales.

6. The method for predicting the course of Alzheimer's disease based on multi-scale features according to claim 5, It is characterized in that The method of using the attention mechanism to perform multi-scale channel attention feature fusion on features of different scales includes the following steps: Use pooling operations to calculate the initial weights of features of different scales; After concatenating the initial weights, they are fed into a 1×1×1 convolution, and the Sigmoid function is used to obtain the importance of each channel, expressed as: Among them, W n Feature M n The corresponding initial weights are represents a 1×1×1 convolution operation, and N represents the number of features; The importance W′ is divided to obtain the corresponding weights W of features of different scales n ′; The features of different scales are weighted using the corresponding weights, and then all weighted results are added element by element to obtain the fused feature Y, which is expressed as: Among them, W n ′ and M n represents a set of corresponding weights and features, Represents a weighted operation.

7. The method for predicting the course of Alzheimer's disease based on multi-scale features according to claim 6, It is characterized in that The pooling operation is used to calculate the initial weights of features of different scales, and two processing channels are used to receive the features M n The output of the two processing channels is then used with the Sigmoid function to obtain the corresponding feature M n The initial weight W n ; Two processing channels include: The first processing channel includes a global average pooling layer, a 1×1×1 convolutional layer, a ReLU activation layer, and a 1×1×1 convolutional layer connected in sequence; The second processing channel includes a global maximum pooling layer, a 1×1×1 convolutional layer, a ReLU activation layer, and a 1×1×1 convolutional layer connected in sequence.

8. The method for predicting the course of Alzheimer's disease based on multi-scale features according to claim 2, It is characterized in that Before the fused features enter the next multi-scale feature module or the average pooling module, a pooling operation is also performed to refine the discriminative features.

9. The method for predicting the course of Alzheimer's disease based on multi-scale features according to claim 2, It is characterized in that The fused features pass through two residual layers before entering the average pooling module, and each residual layer contains two 3×3×3 convolutions.

10. A system for predicting the course of Alzheimer's disease based on multi-scale features. It is characterized in that include: An image acquisition module acquires the original brain image and performs preprocessing to obtain a three-dimensional whole-brain image as the image to be predicted; The image prediction module inputs the image to be predicted into the Alzheimer's disease 3D feature extraction model to predict the course of the disease; The Alzheimer's disease 3D feature extraction model includes an input layer, a plurality of sequentially arranged improved ResNet blocks, an average pooling module, and a fully connected module; the input layer receives the image to be predicted and extracts preliminary features; the preliminary features are sequentially input into a plurality of improved ResNet blocks, multiple feature extractions and fusions are performed, and fusion features are output; The fused features are input into the average pooling module and the fully connected module to obtain the disease course prediction result; the improved ResNet block uses 3D hole convolution to extract features of different scales, and uses the attention mechanism to perform multi-scale feature fusion.

Citation Information

Patent Citations

  • 3D feature recognition method for Alzheimer's disease in MRI image

    CN114820524A

Cited By

  • Alzheimer's disease prediction method and system based on biomics feature topological adjustment

    CN120108656A