An alzheimer's disease classification method based on attention combined with specific structure

Abstract

For the classification of medical images, the application designs an Alzheimer's disease classification method based on attention combined with specific structure. The method involves extracting specific substructure from morphology, extracting effective information of imaging and fusing information in two dimensions. The method includes four parts of data set preprocessing, network model design, network optimal model training and network model performance evaluation. The application fully extracts morphological specific features and effective information of imaging, and designs a self-attention encoding-decoding network, which has broad application prospect in medical image feature extraction and classification.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning, specifically to the field of deep learning-based medical image classification, including medical image substructure extraction and removal of redundant information from the brain, and especially combines information from both image and morphological dimensions to classify images based on the fused information. Background Technology

[0002] In recent years, the classification of Alzheimer's disease (AD) based on magnetic resonance imaging (MRI) has attracted much attention. Studies have shown that Alzheimer's disease is closely related to the morphology of brain substructures. Existing research typically focuses on specific substructures in medical analysis, such as the hippocampus and amygdala, without comprehensively screening for specific substructures. AD is a major public health problem worldwide and one of the most prevalent neurodegenerative diseases. AD is incurable because there are no diagnostic or treatment methods to slow its progression or stop its onset. AD is caused by damage and destruction of nerve cells in brain regions related to memory, with the most common symptoms being memory loss and cognitive decline. A common practice in AD imaging-assisted diagnosis is the use of MRI images, which capture brain structures through magnetic fields and radio waves. Machine learning and deep learning methods have been proposed for the automated identification of dementia, with deep learning showing particular strength in computer-aided diagnosis of AD and MCI. In recent research, convolutional neural networks (CNNs) have demonstrated excellent performance in the automated diagnosis of cognitive disorders from brain MRI images. Compared to 2D convolution on slices, 3D convolution across the entire MRI scan can capture latent 3D structural information, which may be key to recognition. Based on existing work, AD classification methods can be broadly categorized into three types: voxel-based, region of interest (ROI)-based, and region-block-based. In voxel-based methods, features are simply extracted through statistical analysis or voxel selection. A proposal has been made to utilize deep learning combined with brain networks and clinically relevant textual information for early diagnosis of Alzheimer's disease. However, voxel-based features typically have higher dimensionality and noise, which may be irrelevant to the disease. Therefore, techniques such as smoothing, downsampling, and feature selection are needed to reduce the dimensionality of voxel-based features to improve the efficiency of the classifier. In ROI-based methods, brain MRI images are segmented into different tissue ROIs, and then ROI-based feature vectors or relationship vectors between these ROIs are used to describe the MRI images for classifying AD patients.

[0003] Most current research uses 3D convolutions to classify MRI images. Previous papers have proposed classification based on brain MRI substructure images, but none have explored the specificity of the entire brain substructure. This paper uses machine learning methods to extract volumetric feature substructures, then incorporates this information into a deep learning network with an attention module to extract effective information from MRI images. This invention proposes a self-attention classification network combining brain feature substructures, fusing morphological and image information to achieve optimal classification accuracy. Summary of the Invention

[0004] This invention specifically selects morphological information from substructures in medical images and combines a self-attention network to encode and decode the original medical images to extract effective information. It also provides an Alzheimer's disease classification method based on attention combined with specific substructures. The main structure of this method is a Self-Attention based Specific Substructures of Brain (CASSB) model, which innovatively incorporates specific substructure information and introduces a self-attention module. The specific substructures are classified using a machine learning decision tree method, and the top six specific substructures of the decision tree are output for feature selection. The attention module primarily processes the original image through autoencoding, extracting effective information from the original image through an attention mechanism. The principle and block diagram structure of the proposed brain-specific substructure self-attention network are shown below. Figure 1 As shown.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] 1. A classification method for Alzheimer's disease based on attention-based specific structures, characterized by the following steps:

[0007] Step 1: Dataset preprocessing. The brain is divided into 132 substructures according to a specific segmentation algorithm. The number of pixels in each substructure is counted as the volume parameter of the substructure. The volume parameters of the 132 substructures are used as feature vectors and fed into a decision tree for classification and output of brain-specific substructures.

[0008] Step 2: Network model design. A self-attention network model based on specific substructures of brain (CASSB) is proposed. In this model, two new modules, attention submodule and specific substructure volume, are designed to improve the accuracy of the classification network.

[0009] Step 3: Model training. The specific substructures extracted in Step 1 and the original brain images are input into the CASSB network in Step 2 for training. During the training process, parameters are adjusted and optimized based on the performance.

[0010] Step 4: Model performance testing. The test set is fed into the model trained in Step 3 for classification, and the classification metrics are quantitatively evaluated. The main quantitative metrics are the following two: accuracy and recall.

[0011] The following explains the dataset preprocessing steps in step one:

[0012] This paper employs a 3D whole-brain segmentation network based on spatial block maps to extract brain substructures. This method divides the brain MRI into 27 independent blocks, then uses a 3D fully convolutional network to perform whole-brain segmentation on the high-resolution MRI blocks. Each network learns segmentation information in a fixed space. Finally, the segmented sub-blocks are merged to reconstruct the entire brain MRI, completing the whole-brain structural segmentation. This method can segment the brain into 132 substructures, exhibiting better segmentation performance compared to multi-map segmentation methods, while reducing computation time from 30 hours to 15 minutes.

[0013] As the Self Attention based Specific Substructures of Brain (CASSB) model in step two, the attention submodule is described as follows:

[0014] The attention residual module is the main module we improved in our network. We introduced branch attention modules on top of the residual network. The attention residual module is as follows: Figure 2 As shown, the generated attention map is used to mask redundant regions by weighting the feature map extracted through convolution. The attention residual module consists of three parts: 1) attention module, 2) convolution module, and 3) original module. First, the 3D feature input is represented by M. M is first fed into the attention residual network, and the output is A(Ma). Simultaneously, M is fed into the convolutional network module to extract and output features R(Mr). Finally, the results of the two branches are aggregated with the original feature vector using the principle of residual networks. After passing through the attention residual network, 3D image features are extracted, suppressing information irrelevant to Alzheimer's disease.

[0015] (1)

[0016] The attention sub-network is integrated into the attention residual module to generate attention feature maps of 3D features, removing redundant information from MRI images and improving the accuracy of classification tasks. The network structure diagram of the attention sub-network is shown below. Figure 3 As shown, the original MRI image is passed through the convolutional layers of a 3D convolutional neural network. The feature maps extracted by the convolutional layers are downsampled using max pooling to preserve the high-dimensional information of the feature maps. Then, the reverse downsampling process is performed to restore the image to the same size as the original. Further feature extraction is then performed, and the resulting features are input into a softmax function to scale each element in the feature map to the range (0, 1), obtaining an attention-weighted mask. Finally, the original feature map is weighted using the mask to obtain the weighted attention feature map.

[0017] In the next stage, the feature map is added to the output of the remaining block to refine redundant features in its output. Notably, the attention feature map can be used not only as a feature selector in the pre-inference process but also as a gradient update filter in the backpropagation process. Therefore, the attention subnetwork can prevent incorrect gradients from updating parameters. The predicted attention map can be fused with the feature map output through the remaining block, effectively focusing network activation on salient regions.

[0018] The parameters for training the network model in step three are explained below:

[0019] The CASSB network requires adjustment of hyperparameters based on the changing trend and characteristics of the loss function during multiple training sessions. It also requires multiple verification and adjustments through experiments to finally obtain the optimal trained model.

[0020] The following is an explanation of the performance testing of the network model in step four:

[0021] Model performance testing involves feeding the test set into the model trained in step three for classification and quantitatively evaluating the classification metrics, primarily the following two: accuracy and recall. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the principle framework of the brain-specific substructure self-attention network proposed in this invention.

[0023] Figure 2 This is the attention residual submodule of the present invention.

[0024] Figure 3 This is a diagram of the attention sub-network structure in the attention residual sub-module proposed in this invention. Detailed Implementation

[0025] The Alzheimer's disease classification method based on attention combined with specific structures mainly consists of four steps: dataset preprocessing, generative network model construction, training of the optimal network model, and network model performance verification and evaluation.

[0026] In data preprocessing, the data used in this invention came from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The ADNI-SEG dataset used in this study was based on subjects, with 765 subjects and 2294 MRI brain images evenly selected from the ADNI dataset; among them, 407 subjects had AD and 358 had CN; for brain MRI, 1223 cases had AD and 1071 cases had CN. Based on the subjects' age, the mean and variance of the Mini-Mental State Examination (MMSE), Clinical Dementia Rating Scale (CDR), and Clinical Dementia Sum Boxes (CDR-SB), the subjects were evenly divided into training and test sets in a 7:3 ratio.

[0027] The open-source algorithms and mainstream methods were reproduced and their accuracy was evaluated on ADNI-SEG. SVM, ResNet18-3D, and ResNet34-3D were used as the base comparison networks. The output layer of the classification network adopted a two-layer fully connected layer to obtain the final prediction results.

[0028] Table 1. Algorithm accuracy (%) based on the ADNI-SEG dataset

[0029]

Claims

1. A classification method for Alzheimer's disease based on attention-based specific structures, characterized in that... Includes the following steps: Step 1: Dataset preprocessing. A 3D whole-brain segmentation network based on spatial block maps is used to divide the brain MRI into 27 independent spatial blocks. Each spatial block is then segmented into substructures using a 3D fully convolutional network. The segmented sub-blocks are then fused to restore the whole-brain segmentation result, resulting in 132 brain substructures. The number of pixels in each substructure is counted as the volume parameter of that substructure, resulting in a 132-dimensional volume attribute vector. The volume attribute vector is then input into a decision tree for classification, and the top 6 brain-specific substructures are output based on feature importance. Step 2: Network Model Design. A Self Attention based Specific Substructures of Brain (CASSB) network is constructed. This network includes a residual network backbone for extracting 3D features from brain MRI, as well as an attention submodule and a specific substructure volume feature fusion module. The attention submodule is an attention residual module, which is divided into three parts based on three branches: the attention subnetwork branch (attention module), the convolution branch (convolution module), and the original feature branch (original module). The 3D feature input is represented by M. First, M is fed into the attention subnetwork, and the output is A(Ma). Simultaneously, M is fed into the convolution network module to extract and output features R(Mr). Finally, the results of the two branches are aggregated with the original feature vector using the principle of the residual network. The attention sub-network extracts feature maps from the original MRI image through the convolutional layers of a 3D convolutional neural network. Then, it performs max pooling downsampling to preserve the high-dimensional information of the feature maps, followed by a reverse downsampling process to restore them to the same size as the original image. Further feature extraction is then performed, and the resulting features are input into a softmax function to scale each element in the feature map to the range (0, 1), obtaining an attention-weighted mask. Finally, the original feature map is weighted using the mask to obtain a weighted attention feature map. The specific substructure volume feature fusion module is used to fuse the volume parameters of the six brain-specific substructures with the image features extracted by the residual network backbone for classification network model design. Step 3: Model training. The six brain-specific substructure volume parameters obtained in Step 1 and the original brain MRI images are input into the CASSB network in Step 2 for training. During the training process, the parameters are adjusted and optimized based on the performance. Step 4: Model performance testing. The test set is fed into the model trained in Step 3 to output classification results, and classification evaluation metrics are calculated. The evaluation metrics include at least accuracy and recall.

Images