Brain disease classification system based on SBM and multi-scale convolution attention

Abstract

The invention provides a brain disease classification system based on SBM and multi-scale convolution attention, and belongs to the technical field of medical image processing. The system comprises a data processing module which comprises an image data preprocessing module; the SBM module comprises SBM feature extraction; the network module comprises a multi-scale convolution attention and a 3D-CNN (Three Dimensional Convolutional Neural Network); and the classification module comprises a feature selection method and a classifier. According to the method, a channel attention mechanism and a space attention mechanism are adopted to pay attention to different space positions in a feature map; performing multi-scale feature extraction on each brain region by adopting convolution kernels of different scales; according to the method, on the basis that only grey matter data is used in a traditional method, cortex data and T1w / T2w data are integrated for classification, the available features of a classification algorithm are effectively increased, and therefore the classification accuracy is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a brain disease classification system based on SBM and multi-scale convolutional attention. Background Art

[0002] In the field of medical image processing, through in-depth analysis of structural magnetic resonance imaging (sMRI), computer algorithms can extract key brain structural features to help doctors more accurately assess the patient's cognitive status. This technology not only improves the accuracy of early diagnosis, but also provides strong support for the formulation of personalized treatment plans. With the continuous advancement of artificial intelligence and deep learning technology, the efficiency and effectiveness of computer-assisted analysis have been significantly improved. This enables higher accuracy and reliability in the classification of early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI) and Alzheimer's Disease (AD). At present, researchers generally use machine learning technology to analyze patients' sMRI images to identify the characteristics of Alzheimer's disease at different stages.

[0003] Compared with traditional methods, 3D Convolutional Neural Network (3D-CNN) has significant advantages in processing medical images. 3D-CNN shows better contextual understanding when processing volumetric data, which is crucial for identifying subtle changes in neurodegenerative diseases such as Alzheimer's disease. Through comprehensive analysis of the entire brain structure, 3D-CNN can better identify disease-related features, promote early diagnosis and personalized treatment. In contrast, traditional two-dimensional convolutional neural networks (2D-CNN) can only process information in a single plane and may ignore deep features in the image. But this also has some disadvantages, such as low feature interpretability and low data utilization. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a brain disease classification system based on SBM and multi-scale convolutional attention, which performs feature extraction, feature selection and feature classification on each brain region of the subject, thereby improving the detection and diagnosis accuracy of brain diseases.

[0005] The present invention adopts the following scheme:

[0006] A brain disease classification system based on SBM and multi-scale convolutional attention, including:

[0007] A data processing module, which includes image data preprocessing;

[0008] SBM module, which includes SBM feature extraction. ;

[0009] Network module: This module includes multi-scale convolutional attention and 3D-CNN network;

[0010] Classification module: This module includes feature selection methods and classifiers.

[0011] Furthermore, the work content of the data processing module is:

[0012] The T1-weighted and T2-weighted data were corrected for anterior commissure-posterior commissure, respectively. The preprocessing steps of T1-weighted data included skull removal and tissue segmentation to obtain cortical and gray matter images. T1w / T2w preprocessing involved radiographic registration and intensity calibration to generate T1w / T2w images.

[0013] Furthermore, the working content of the SBM module is to perform surface-based morphological measurement analysis on the sMRI data to extract cortical thickness, groove depth, cortical folds and cortical complexity.

[0014] Furthermore, the working content of the network module is: select the hippocampal ontology map template, divide the brain into 112 brain regions, and extract features through a multi-scale convolutional attention network.

[0015] Furthermore, the work content of the classification module is: to perform feature selection and feature classification on the outputs of the SBM module and the network module.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] (1) Based on the traditional method of using only gray matter data, the present invention integrates cortical data and T1w / T2w data for classification, effectively increasing the available features of the classification algorithm, thereby significantly improving the classification accuracy.

[0018] (2) The present invention adopts a channel attention mechanism to learn the importance of each channel. Different channels may contain different types of feature information. By assigning different weights to each channel, the features of important channels can be highlighted while the features of irrelevant channels can be suppressed.

[0019] (3) The present invention adopts a spatial attention mechanism to focus on different spatial locations in the feature map. In medical images, the target area may only occupy a part of the image. The spatial attention mechanism enables the model to focus on these key areas, thereby improving the ability to capture the characteristics of the target area.

[0020] (4) The present invention uses convolution kernels of different scales to perform multi-scale feature extraction on each brain region. Convolution kernels of different scales can capture features of different sizes. By fusing these multi-scale features, a more comprehensive and discriminative feature representation can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the overall framework of the system of the present invention;

[0022] Figure 2 It is a structural diagram of the data processing module in the present invention;

[0023] Figure 3 It is a schematic diagram of the structure of the SBM module in the present invention;

[0024] Figure 4 It is a structural diagram of the network module in the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0027] A brain disease classification system based on SBM and multi-scale convolutional attention, the system composition reference Figure 1 As shown, the system of the present invention includes four modules, namely a data processing module, an SBM module, a network module and a classification module.

[0028] refer to Figure 2 As shown, the structure of the data processing module includes: T1 preprocessing and T1w / T2w preprocessing.

[0029] T1w data and T2w data were preprocessed to generate cortical images, gray matter images and T1w / T2w images, respectively.

[0030] refer to Figure 3 As shown, the SBM module includes brain surface reconstruction and extraction of four indicators.

[0031] The SBM module extracts four indicators through brain surface reconstruction, namely cortical thickness, groove depth, cortical folds and cortical complexity.

[0032] The SBM module uses spherical registration of the cortical surface grid to improve the accuracy of brain registration. This method not only enhances the accuracy of positioning, but also brings many topological structures inside the brain to the surface, thereby mining more deep features and increasing the interpretability of features.

[0033] refer to Figure 4 As shown, the structure of the network module includes gray matter image, T1w / T2w image, module 1, module 2, module 3 and module 4.

[0034] Module 1 is the channel attention block (CAB), which includes pooling layer 1, pooling layer 2, convolution layer 1, activation function layer 1 and activation function layer 2.

[0035] Pooling layer 1 is Adaptive Max Pooling (AMP), which is used to adaptively select the maximum value according to the size of the input feature map, emphasize the significant features in the feature map, and enhance the model's attention to important information.

[0036] Pooling layer 2 is Adaptive Average Pooling (AAP), which reduces the dimension of feature maps by averaging the feature maps while retaining global information.

[0037] Convolutional layer 1 uses a convolution kernel of size 1×1.

[0038] The activation function layer 1 uses the ReLU (Rectified Linear Unit, ReLU) activation function. Its characteristic is that it sets the negative values ​​in the input to zero and retains the positive values. This nonlinear transformation can effectively introduce the nonlinear characteristics of the model while avoiding the problem of gradient disappearance.

[0039] The activation function layer 2 uses the Sigmoid activation function. Its output value ranges from 0 to 1. This feature makes Sigmoid very suitable for binary classification problems, especially in the output layer, where the output of the model can be interpreted as probability.

[0040] Module 2 is the Spatial Attention Block (SAB), which includes pooling layer 3, pooling layer 4, convolution layer 2, and activation function layer 2.

[0041] Pooling layer 3 is adaptive maximum pooling: an adaptive maximum pooling operation is performed along the channel dimension to obtain a feature description.

[0042] Pooling layer 4 is adaptive average pooling: adaptive average pooling operation is performed along the channel dimension to obtain another feature description.

[0043] Convolutional layer 2 uses a convolution kernel of size 7×7.

[0044] Module 3 is a multi-scale convolution block (MSCB), which includes convolution layer 1, batch normalization layer, activation function layer 3 and module 5.

[0045] The main function of the batch normalization (BN) layer is to reduce internal covariate shift and improve training speed and model stability.

[0046] The activation function layer 3 uses ReLU6, which is a variant of ReLU (Rectified Linear Unit), which limits the maximum output value to 6; this helps alleviate the gradient vanishing problem and makes the network training more stable.

[0047] Module 5 is a multi-scale dilated convolution (MSDC), which includes a convolution layer 2, a batch normalization layer, an activation function layer 3, and a channel rearrangement layer. It can simultaneously extract multi-scale features through different convolution kernels, which enables it to capture richer information, including local details and global context information.

[0048] Convolutional layer 2 is a depthwise separable convolution (Dwconv), which is computationally more efficient than standard convolution and significantly reduces the required multiplication and addition operations, thereby reducing the computational complexity of the model.

[0049] The main function of the channel rearrangement layer is to promote the interaction and fusion of information by rearranging the channels in the feature map, thereby improving the performance of the model.

[0050] Module 4 is a 3D-CNN network, including convolution layer 3, convolution layer 4, pooling layer 5 and pooling layer 6.

[0051] The convolution kernel size of convolution layer 3 is 1x1x1, which is mainly used for feature fusion between channels.

[0052] The convolution kernel size of convolution layer 4 is 3x3x3. Through the convolution operation, the spatial dimension of the input data is reduced while retaining important feature information.

[0053] Pooling layer 5 is a three-dimensional maximum pooling layer with a pooling kernel size of 2x2x2. By reducing the size of the feature map, the pooling layer helps control the complexity of the model and reduce the risk of overfitting.

[0054] Pooling layer 6 is a global average pooling layer, which extracts the global features of the channel by averaging all spatial positions of each channel. This helps to better capture the overall information of the entire input feature map.

[0055] The network module extracts the channel and spatial features of the image, which can effectively enhance the expression of key features while reducing the interference of redundant information, thus improving the performance and generalization ability of the model.

[0056] The output of the SBM module and the output of the network module are used as the input of the classification module.

[0057] The classification module includes a feature selection layer and a classifier, and finally outputs the classification probability.

[0058] The feature selection layer uses the variance method to select the most valuable features for model prediction from high-dimensional data. The basic principle is to evaluate the importance of each feature in the data by calculating its variance. After screening, the retained features will form a new feature set for subsequent model training. This process not only helps to reduce computational complexity, but also significantly improves the performance and generalization ability of the model.

[0059] The present invention selects support vector machine, random forest and naive Bayes as classifiers. In the classification tasks of AD vs NC, LMCI vs NC, EMCI vs NC, and LMCI vs EMCI, support vector machine, random forest and naive Bayes are used for classification. By evaluating the classification results of each task, the best performing model is finally selected as the final classification result of the present invention.

[0060] The present invention uses a ten-fold cross validation method, which can effectively reduce overfitting, provide more accurate performance evaluation, and make full use of limited data sets. By dividing the data into multiple subsets, the ten-fold cross validation ensures that each data point can be used for training and testing, thereby improving the generalization ability of the model. In addition, the method is also applicable to a variety of machine learning models, which helps to select the best model and hyperparameters, and effectively reduces the impact of randomness on the classification results.

[0061] The model results are shown in Table 1.

[0062] Table 1

[0063]

[0064]

[0065] Table 1 lists in detail the various indicators of the present invention, including accuracy (ACC), area under the curve (AUC), sensitivity (SEN), specificity (SPE), precision (PRE) and F1 score (F1). The results show that the present invention has achieved a significant improvement in accuracy, which indicates that its performance in classification tasks is better than that of traditional methods.

[0066] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A brain disease classification system based on SBM and multi-scale convolutional attention, characterized in that, include: A data processing module, which includes image data preprocessing; SBM module, which includes SBM feature extraction; Network module: This module includes multi-scale convolutional attention and 3D-CNN network; Classification module: This module includes feature selection methods and classifiers.

2. A brain disease classification system based on SBM and multi-scale convolutional attention according to claim 1, characterized in that: The work content of the data processing module is: The T1-weighted and T2-weighted data were corrected for anterior commissure-posterior commissure respectively; the preprocessing steps of T1-weighted data included skull removal and tissue segmentation to obtain cortical images and gray matter images; T1w / T2w preprocessing involves radioregistration and intensity calibration to generate T1w / T2w images.

3. A brain disease classification system based on SBM and multi-scale convolutional attention according to claim 1, characterized in that: The working content of the SBM module is to perform surface-based morphological measurement analysis on sMRI data to extract cortical thickness, groove depth, cortical folds and cortical complexity.

4. A brain disease classification system based on SBM and multi-scale convolutional attention according to claim 1, characterized in that: The working content of the network module is: select the hippocampal ontology map template, divide the brain into 112 brain regions, and extract features through a multi-scale convolutional attention network.

5. A brain disease classification system based on SBM and multi-scale convolutional attention according to claim 1, characterized in that: The work content of the classification module is: to perform feature selection and feature classification on the output of the SBM module and the network module.

Images