Method and system for monitoring brain entropy in patient with alzheimer's disease
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-02-07
- Publication Date
- 2026-08-13
Smart Images
Figure CN2025076140_13082026_PF_FP_ABST
Abstract
Description
A method and system for monitoring brain entropy in Alzheimer's patients Technical Field
[0001] This invention belongs to the field of brain entropy monitoring technology, and in particular relates to a method and system for monitoring brain entropy in Alzheimer's patients. Background Technology
[0002] In monitoring brain entropy in Alzheimer's patients, it is necessary to analyze and process the patient's brain imaging data to determine various indicators reflecting brain structure and function. However, since the patient's brain imaging data usually contains multiple modalities, such as T1-weighted images and FLAIR images, the differences between different modalities in imaging principles, contrast, and resolution make image registration and fusion difficult. Furthermore, the varying degrees of degenerative changes in the brain structure and function of Alzheimer's patients, such as brain atrophy and white matter lesions, make accurate segmentation and quantification of brain structure challenging. At the same time, individual differences such as patient age and gender can also affect the calculation of brain entropy; how to consider and eliminate the influence of these confounding factors during the calculation process is an urgent problem to be solved. Therefore, how to achieve effective fusion of multimodal brain images, accurately segment and quantify brain structure, and eliminate the influence of individual differences when calculating brain entropy indicators are key technical problems facing brain entropy monitoring methods and systems for Alzheimer's patients. Summary of the Invention
[0003] This invention proposes a method and system for monitoring brain entropy in Alzheimer's patients to address the problems existing in the prior art.
[0004] To achieve the above objectives, the present invention provides a method for monitoring brain entropy in Alzheimer's patients, comprising the following steps:
[0005] Multimodal brain imaging data of patients are acquired, and the multimodal brain imaging data are spatially aligned using a registration algorithm to obtain a registered multimodal image dataset.
[0006] A deep learning-based feature extraction network is used to extract discriminative feature maps from the registered multimodal image dataset, resulting in a multimodal feature map set.
[0007] The multimodal feature map set is input into a preset fusion network, and the feature maps of different modalities are weighted and fused through an adaptive weight allocation mechanism to generate a fused brain feature map.
[0008] Based on the fused brain feature map, the U-Net network architecture is used to segment the brain, and the segmentation results of gray matter, white matter and cerebrospinal fluid are obtained.
[0009] Based on the segmentation results, morphological features of brain structures are extracted, including sulcus depth, cortical thickness, and hippocampal volume. Based on these morphological features, a brain structure feature vector is constructed.
[0010] The brain structural feature vector is input into a preset brain entropy calculation model to obtain the brain entropy value.
[0011] Preferably, the registered multimodal image dataset includes:
[0012] Preprocessing techniques are employed to enhance the images of multimodal brain imaging data. Based on the preprocessed multimodal image data, mutual information features are extracted to construct a mutual information matrix. A registration algorithm based on the mutual information matrix is used to transform the spatial coordinates of the multimodal image data to achieve spatial alignment. The transformation matrix obtained through the registration algorithm is then applied to the original multimodal image data to complete the registration. Finally, the registered multimodal image data is fused to generate a fused multimodal image dataset.
[0013] Preferably, the multimodal feature map set includes:
[0014] Based on the registered multimodal image dataset, a deep learning-based convolutional neural network is used to extract features from the image data of each modality to obtain discriminative feature maps. For the extracted multimodal feature maps, the feature maps of different modalities are fused through a feature fusion module to generate a fused multimodal feature map set.
[0015] Preferably, generating the fused brain feature map includes:
[0016] The feature map sets of each modality are input into a pre-defined fusion network. The fusion network adopts an attention mechanism and performs weighted fusion of feature maps of different modalities through adaptive weight allocation. The attention mechanism in the fusion network determines the weight coefficient of each modality by calculating the similarity between feature maps of different modalities. The larger the weight coefficient, the greater the contribution of that modality to the fusion result. Based on the determined weight coefficient, the feature maps of different modalities are weighted and summed to obtain the fused multimodal feature map.
[0017] Preferably, the segmentation results of gray matter, white matter, and cerebrospinal fluid include:
[0018] Based on the fused brain feature map, a pre-trained U-Net network model was used to segment brain tissue, obtaining segmentation results for gray matter, white matter, and cerebrospinal fluid. Morphological operations and connected component analysis were performed on the segmentation results, and the volume indicators of gray matter, white matter, and cerebrospinal fluid were calculated based on the segmentation results to obtain quantitative brain tissue volume information.
[0019] Preferably, constructing the brain structure feature vector includes:
[0020] Morphological analysis methods were used to extract sulcus depth, cortical thickness, and hippocampal volume from the segmentation results to obtain detailed information on brain structure and morphology. Based on the extracted morphological feature indicators, a feature fusion algorithm was used to combine multiple indicators to construct a high-dimensional feature vector that comprehensively reflects the brain structure and morphology.
[0021] This invention also discloses a brain entropy monitoring system for Alzheimer's patients, comprising:
[0022] The data acquisition and registration module is used to acquire the patient's multimodal brain image data and perform spatial alignment processing on the multimodal brain image data using a registration algorithm to obtain the registered multimodal image dataset.
[0023] The feature extraction module is used to extract discriminative feature maps from the registered multimodal image dataset using a deep learning-based feature extraction network, thereby obtaining a multimodal feature map set.
[0024] The feature fusion module is used to input the multimodal feature map set into a preset fusion network, and perform weighted fusion processing on the feature maps of different modalities through an adaptive weight allocation mechanism to generate a fused brain feature map.
[0025] The brain segmentation module is used to segment the brain based on the fused brain feature map using a U-Net network architecture to obtain segmentation results for gray matter, white matter, and cerebrospinal fluid.
[0026] The feature vector construction module is used to extract morphological features of brain structures based on the segmentation results. The morphological features include sulcus depth, cortical thickness, and hippocampal volume. Based on the morphological features, a brain structure feature vector is constructed.
[0027] The brain entropy calculation module is used to input brain structural feature vectors into a preset brain entropy calculation model to obtain brain entropy values.
[0028] Preferably, the data acquisition and registration module includes:
[0029] The acquisition unit is used to acquire multimodal brain imaging data from patients.
[0030] The preprocessing unit is used to perform image enhancement on multimodal brain imaging data using preprocessing techniques;
[0031] The mutual information matrix construction unit is used to extract the mutual information features of the images based on the preprocessed multimodal image data and construct the mutual information matrix.
[0032] Alignment units are used to perform spatial coordinate transformation on multimodal image data using a registration algorithm based on mutual information matrices to achieve spatial alignment.
[0033] The registration unit is used to obtain the transformation matrix through the registration algorithm, and then apply the transformation matrix to the original multimodal image data to complete the registration.
[0034] The fusion unit is used to fuse the registered multimodal image data to generate a fused multimodal image dataset.
[0035] Compared with the prior art, the present invention has the following advantages and technical effects:
[0036] This invention acquires multimodal brain imaging data, including T1-weighted and FLAIR images, and uses a mutual information registration algorithm for spatial alignment. Then, a deep learning network is used to extract and fuse multimodal features. Subsequently, a U-Net network is used to segment the brain using the fused features, obtaining segmentation results for gray matter, white matter, and cerebrospinal fluid. Based on the segmentation results, morphological features such as sulcus depth, cortical thickness, and hippocampal volume are extracted to construct brain structural feature vectors. Finally, the standardized feature vectors are input into a pre-defined brain entropy calculation model, and the brain structural complexity index is calculated using information entropy theory to obtain the final brain entropy value. This invention, through multimodal image fusion and deep learning technology, achieves accurate analysis and quantitative assessment of brain structure, providing new tools and indicators for the diagnosis and research of brain diseases. Attached Figure Description
[0037] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0038] Figure 1 is a flowchart of the method according to an embodiment of the present invention. Detailed Implementation
[0039] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0040] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0041] Example 1
[0042] As shown in Figure 1, this embodiment provides a method for monitoring brain entropy in Alzheimer's patients, including the following steps:
[0043] Multimodal brain imaging data of patients are acquired, and the multimodal brain imaging data are spatially aligned using a registration algorithm to obtain a registered multimodal image dataset.
[0044] A deep learning-based feature extraction network is used to extract discriminative feature maps from the registered multimodal image dataset, resulting in a multimodal feature map set.
[0045] The multimodal feature map set is input into a preset fusion network, and the feature maps of different modalities are weighted and fused through an adaptive weight allocation mechanism to generate a fused brain feature map.
[0046] Based on the fused brain feature map, the U-Net network architecture is used to segment the brain, and the segmentation results of gray matter, white matter and cerebrospinal fluid are obtained.
[0047] Based on the segmentation results, morphological features of brain structures are extracted, including sulcus depth, cortical thickness, and hippocampal volume. Based on these morphological features, a brain structure feature vector is constructed.
[0048] The brain structural feature vector is input into a preset brain entropy calculation model to obtain the brain entropy value.
[0049] Specifically, the following steps are included:
[0050] Step S101: Acquire the patient's multimodal brain imaging data, including T1-weighted images and FLAIR images; perform image enhancement using preprocessing techniques to improve image quality; extract mutual information features from the preprocessed multimodal image data and construct a mutual information matrix; use a registration algorithm based on the mutual information matrix to transform the spatial coordinates of the multimodal image data to achieve spatial alignment; obtain the transformation matrix through the registration algorithm and apply the transformation matrix to the original multimodal image data to complete the registration; fuse the registered multimodal image data to generate a fused multimodal image dataset; use the fused multimodal image dataset for subsequent medical image analysis and disease diagnosis.
[0051] In this embodiment, the acquisition of multimodal brain imaging data is fundamental to neuroimaging research. T1-weighted images provide high-resolution anatomical information about brain structures, while FLAIR images better reveal white matter lesions. For example, in Alzheimer's disease research, T1 images can be used to assess the degree of hippocampal atrophy, while FLAIR images help to observe high-signal lesions in the white matter. Preprocessing techniques are crucial for improving image quality. Commonly used methods include denoising, bias field correction, and intensity normalization. Taking denoising as an example, a nonlocal mean filtering algorithm can be used, which reduces noise by searching for similar blocks throughout the image while preserving edge details. This is significant for subsequent feature extraction and registration steps. Mutual information feature extraction is key to achieving multimodal image registration. The mutual information matrix reflects the statistical dependencies between different modalities. For example, for T1 and FLAIR images, their joint histogram can be calculated, and then the mutual information can be calculated based on this. High mutual information values indicate a strong correlation between the two modalities, which helps determine the optimal registration parameters. Mutual information-based registration algorithms find the optimal spatial transformation by maximizing mutual information. Commonly used transformations include rigid body transformations, affine transformations, and nonlinear transformations. Taking affine transformations as an example, image alignment can be achieved by optimizing rotation, translation, and scaling parameters. This is crucial for correcting spatial inconsistencies caused by changes in patient position across different scans. The application of the transformation matrix is the final step in the registration process. For example, for 3D images, a 4x4 transformation matrix can be used to describe the spatial transformation. Applying this matrix to each voxel coordinate of the original image yields the registered image. This step ensures accurate spatial correspondence between images of different modalities.
[0052] Step S102: Based on the registered multimodal image dataset, a deep learning-based convolutional neural network is used to extract features from the image data of each modality, resulting in discriminative feature maps. Then, the feature fusion module fuses the feature maps from different modalities to generate a fused multimodal feature map set.
[0053] In this embodiment, feature extraction and fusion of multimodal brain imaging data is a complex process involving several key steps. First, convolutional neural networks from deep learning are used to extract features from the image data of each modality. Taking T1-weighted images as an example, a VGG16 network can be used as the basic architecture, extracting feature maps containing information such as texture and shape through multi-layer convolution and pooling operations. For FLAIR images, a ResNet50 network can be used, leveraging its residual structure to better capture the features of lesion areas. The feature fusion module is crucial for integrating feature maps from different modalities. An effective method is to employ attention mechanisms, such as a dual attention mechanism combining channel attention and spatial attention. This mechanism can adaptively assign weights to features from different modalities, highlighting important information. For example, in brain tumor detection tasks, FLAIR images are more sensitive to displaying edema areas and may therefore be assigned higher weights. The fused feature maps undergo nonlinear transformation through fully connected layers, mapping the features to a high-dimensional space. This step can use the ReLU activation function to increase the model's nonlinear expressive power. In the high-dimensional feature space, the Support Vector Machine (SVM) algorithm is used for sample classification. SVM distinguishes different categories by finding the optimal hyperplane, making it particularly suitable for handling high-dimensional data. For example, in the task of classifying benign and malignant brain tumors, SVM can effectively differentiate based on features such as the shape, size, and location of the tumor. The model training process uses the backpropagation algorithm, continuously adjusting network parameters by calculating the gradient of the loss function. In practice, cross-entropy can be used as the loss function, and the Adam optimizer can be used for parameter updates. During training, a validation set can be set to monitor model performance and avoid overfitting. For example, the dataset can be divided into training, validation, and test sets in a 7:2:1 ratio. Iterative training is key to improving model performance. An early stopping strategy can be implemented, stopping training when the performance on the validation set has not improved for several consecutive epochs. Simultaneously, a learning rate decay strategy can be used, gradually reducing the learning rate as training progresses to help the model converge to the optimal solution more effectively. The advantage of this multimodal feature extraction and fusion method lies in its ability to fully utilize the complementary information of different modalities of imagery.
[0054] Step S103: Input the feature map sets of each modality into the preset fusion network. The fusion network adopts an attention mechanism and performs weighted fusion of feature maps of different modalities through adaptive weight allocation. The attention mechanism in the fusion network determines the weight coefficient of each modality by calculating the similarity between feature maps of different modalities. The larger the weight coefficient, the greater the contribution of the modality to the fusion result. According to the determined weight coefficient, the feature maps of different modalities are weighted and summed to obtain the fused multimodal feature map.
[0055] In this embodiment, the multimodal brain medical image data includes structural magnetic resonance imaging (MRI), functional magnetic resonance imaging (fMRI), and positron emission tomography (PET). Different feature extraction algorithms can be used for each modality. For example, wavelet transform can be used to extract multi-scale texture features for MRI images; independent component analysis (ICA) can be used to extract time-series features for fMRI data; and gray-level co-occurrence matrix (GLCM) can be used to extract statistical features for PET images. These features constitute the feature map set for each modality. The attention mechanism in the fusion network is crucial. Assuming there are three modalities—MRI, fMRI, and PET—the attention mechanism first calculates the similarity between them. Cosine similarity can be used, such as 0.7 for MRI and fMRI, 0.5 for MRI and PET, and 0.6 for fMRI and PET. Based on the similarity, weight coefficients are determined, for example, 0.4 for MRI, 0.35 for fMRI, and 0.25 for PET. The weights reflect the degree of contribution of each modality to the final fusion result. In the weighted fusion process, the feature map of each modality is multiplied by its corresponding weight coefficient and then summed at the pixel level. This results in a fused feature map that retains the unique information of each modality while highlighting the contributions of important modalities. For example, in brain tumor diagnosis, MRI provides a clearer depiction of tumor morphology and may therefore receive a higher weight. The fused feature map is then input into a pre-defined convolutional neural network, such as U-Net or DeepLab. The network extracts higher-level semantic features progressively through multiple layers of convolution and pooling operations. Shallow features may include low-level information such as edges and textures, while deep features may represent specific brain region structures or functional patterns. The final brain feature map integrates multimodal information and multi-scale features. Segmentation algorithms can utilize deep learning-based methods, such as fully convolutional networks (FCNs) or maskR-CNNs. These algorithms can effectively segment brain feature maps into different anatomical structures or functional regions. For example, they can segment key brain regions such as the cerebral cortex, white matter, gray matter, and hippocampus. Based on the segmentation results, parameters such as the volume and shape of each brain region can be calculated. Volume can be obtained by calculating the number of voxels in the segmented region; shape can be described by indices such as ellipticity or complexity. These parameters are of great significance for the diagnosis of brain diseases. For example, in Alzheimer's disease research, a reduction in hippocampal volume is an important biomarker. By comparing the hippocampal volume of a patient with the normal range, doctors can make more accurate diagnoses. The advantage of this multimodal fusion method lies in its full utilization of complementary information from different imaging techniques. MRI provides high-resolution anatomical structures, fMRI reflects brain region activity, and PET shows metabolic activity. Through intelligent fusion, more comprehensive and accurate brain information can be obtained, improving the sensitivity and specificity of disease diagnosis. This has significant clinical value for the early detection and accurate diagnosis of brain diseases, such as tumors and neurodegenerative diseases.
[0056] Step S104: Based on the fused brain feature map, a pre-trained U-Net network model is used to segment the brain tissue, obtaining segmentation results for gray matter, white matter, and cerebrospinal fluid. Post-processing of the segmentation results is performed, including morphological operations and connected component analysis, to improve segmentation accuracy and robustness. Based on the segmentation results, volume indicators of gray matter, white matter, and cerebrospinal fluid are calculated to obtain quantitative brain tissue volume information.
[0057] In this embodiment, the U-Net network, due to its encoder-decoder structure and skip connections, is particularly suitable for medical image segmentation tasks. By pre-training on a large-scale labeled dataset, U-Net can learn general features of brain tissue, improving segmentation accuracy. During segmentation, multi-scale feature fusion can be introduced to simultaneously capture local details and global structural information. Post-segmentation processing further enhances the quality of the results. Morphological operations such as opening can remove small mis-segmented regions, while closing can fill small holes. Connected component analysis helps remove isolated erroneous labels, ensuring the continuity of the segmentation results. These operations can significantly improve the robustness of segmentation, especially for images with high noise or low contrast. Volume calculation provides quantitative evidence for clinical diagnosis. By statistically analyzing the number of voxels in different tissue types and combining this with image resolution information, accurate volume data can be obtained. For example, for a 65-year-old male patient, if the calculated hippocampal volume is 20% lower than the normal range, this may be an indication of early Alzheimer's disease. Comparing the calculated results with the normal range is a crucial diagnostic step. The normal range is typically based on large-scale population statistics and considers factors such as age and gender. For example, for the patient mentioned above, a possible diagnostic recommendation would be: Since the hippocampal volume is significantly lower than the average for its age group, further neuropsychological evaluation and cerebrospinal fluid biomarker testing are recommended to confirm the presence of neurodegenerative changes. Finally, the generated analysis report should include rich visualizations. A three-dimensional reconstructed brain tissue model can visually demonstrate the spatial relationships of various structures. Comparison images of segmentation results from different sections help doctors quickly locate abnormal areas. Bar charts or radar charts of volume indicators help patients understand the differences between their own condition and normal values. These visualization techniques greatly improve the readability and clinical applicability of the report.
[0058] Step S105: Based on the segmentation results, multiple morphological feature indicators such as sulcus depth, cortical thickness, and hippocampal volume are extracted using morphological analysis methods to obtain detailed brain structural morphology information. Based on the extracted morphological feature indicators, a feature fusion algorithm is used to combine multiple indicators to construct a high-dimensional feature vector that comprehensively reflects the brain structural morphology. By performing dimensionality reduction processing on the feature vector and removing redundant information, a more refined and effective brain structural feature vector is obtained.
[0059] In this embodiment, morphological feature extraction is an important means of quantifying brain structure. Sulcus depth can be measured by calculating the distance from the cortical surface to the bottom of the sulcus, while cortical thickness can be obtained by calculating the distance between the inner and outer surfaces of gray matter. Hippocampal volume can be calculated by voxel counting of the segmented hippocampal region. These indicators can reflect subtle changes in brain structure and are of great significance for early disease diagnosis. Feature fusion algorithms can integrate multiple morphological indicators into a comprehensive feature vector. For example, principal component analysis (PCA) can be used to combine indicators such as sulcus depth, cortical thickness, and hippocampal volume into a high-dimensional feature vector. This method can retain the important information of each indicator while reducing data redundancy. Dimensionality reduction helps extract the most representative features. The t-SNE (t-distributed random neighborhood embedding) algorithm can map high-dimensional feature vectors to a low-dimensional space, preserving the local structure of the data.
[0060] Step S106: The standardized brain structure feature vector is input into a pre-defined brain entropy calculation model. The model uses an algorithm based on information entropy theory to calculate the complexity index of the brain structure. Based on the brain structure complexity index, the brain entropy value corresponding to the complexity index is obtained by querying a pre-established brain entropy value mapping table. It is determined whether the brain entropy value exceeds a normal threshold. If it exceeds the threshold, it indicates a brain structure abnormality, triggering an early warning mechanism; if it does not exceed the threshold, the brain structure is considered normal, and no early warning is triggered. For abnormal brain structures that trigger early warnings, key features are extracted from their feature vectors, and historical case data similar to the key features are retrieved from the feature vector database. A clustering algorithm is used to group the retrieved similar cases, and the disease type of the current abnormal brain structure is determined based on the clustering results, obtaining a preliminary diagnostic conclusion. The preliminary diagnostic conclusion is correlated with other medical data of the patient. By comprehensively assessing the degree of brain structure abnormality and overall physical condition, personalized treatment plan suggestions are provided to assist doctors in clinical decision-making.
[0061] This embodiment also discloses a brain entropy monitoring system for Alzheimer's patients, including:
[0062] The data acquisition and registration module is used to acquire patients' multimodal brain image data and to perform spatial alignment processing on the multimodal brain image data using a registration algorithm to obtain a registered multimodal image dataset.
[0063] The feature extraction module is used to extract discriminative feature maps from the registered multimodal image dataset using a deep learning-based feature extraction network, thereby obtaining a multimodal feature map set.
[0064] The feature fusion module is used to input a set of multimodal feature maps into a preset fusion network, and to perform weighted fusion processing on the feature maps of different modalities through an adaptive weight allocation mechanism to generate a fused brain feature map.
[0065] The brain segmentation module is used to segment the brain based on the fused brain feature map using the U-Net network architecture, and obtain segmentation results of gray matter, white matter and cerebrospinal fluid.
[0066] The feature vector construction module is used to extract morphological features of brain structures based on the segmentation results. The morphological features include sulcus depth, cortical thickness, and hippocampal volume. Based on the morphological features, a brain structure feature vector is constructed.
[0067] The brain entropy calculation module is used to input brain structural feature vectors into a preset brain entropy calculation model to obtain brain entropy values.
[0068] Preferably, the data acquisition and registration module includes:
[0069] The acquisition unit is used to acquire multimodal brain imaging data from patients.
[0070] The preprocessing unit is used to perform image enhancement on multimodal brain imaging data using preprocessing techniques;
[0071] The mutual information matrix construction unit is used to extract the mutual information features of the images based on the preprocessed multimodal image data and construct the mutual information matrix.
[0072] Alignment units are used to perform spatial coordinate transformation on multimodal image data using a registration algorithm based on mutual information matrices to achieve spatial alignment.
[0073] The registration unit is used to obtain the transformation matrix through the registration algorithm, and then apply the transformation matrix to the original multimodal image data to complete the registration.
[0074] The fusion unit is used to fuse the registered multimodal image data to generate a fused multimodal image dataset.
[0075] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for monitoring brain entropy in Alzheimer's patients, characterized in that, Includes the following steps: Multimodal brain imaging data of patients are acquired, and the multimodal brain imaging data are spatially aligned using a registration algorithm to obtain a registered multimodal image dataset. A deep learning-based feature extraction network is used to extract discriminative feature maps from the registered multimodal image dataset, resulting in a multimodal feature map set. The multimodal feature map set is input into a preset fusion network, and the feature maps of different modalities are weighted and fused through an adaptive weight allocation mechanism to generate a fused brain feature map. Based on the fused brain feature map, the U-Net network architecture is used to segment the brain, and the segmentation results of gray matter, white matter and cerebrospinal fluid are obtained. Based on the segmentation results, morphological features of brain structures are extracted, including sulcus depth, cortical thickness, and hippocampal volume. Based on these morphological features, a brain structure feature vector is constructed. The brain structural feature vector is input into a preset brain entropy calculation model to obtain the brain entropy value.
2. The method according to claim 1, characterized in that, The obtained registered multimodal image dataset includes: Preprocessing techniques are employed to enhance the images of multimodal brain imaging data. Based on the preprocessed multimodal image data, mutual information features are extracted to construct a mutual information matrix. A registration algorithm based on the mutual information matrix is used to transform the spatial coordinates of the multimodal image data to achieve spatial alignment. The transformation matrix obtained through the registration algorithm is then applied to the original multimodal image data to complete the registration. Finally, the registered multimodal image data is fused to generate a fused multimodal image dataset.
3. The method according to claim 1, characterized in that, The resulting set of multimodal feature maps includes: Based on the registered multimodal image dataset, a deep learning-based convolutional neural network is used to extract features from the image data of each modality to obtain discriminative feature maps. For the extracted multimodal feature maps, the feature maps of different modalities are fused through a feature fusion module to generate a fused multimodal feature map set.
4. The method according to claim 1, characterized in that, The generated fused brain feature map includes: The feature map sets of each modality are input into a pre-defined fusion network. The fusion network adopts an attention mechanism and performs weighted fusion of feature maps of different modalities through adaptive weight allocation. The attention mechanism in the fusion network determines the weight coefficient of each modality by calculating the similarity between feature maps of different modalities. The larger the weight coefficient, the greater the contribution of that modality to the fusion result. Based on the determined weight coefficient, the feature maps of different modalities are weighted and summed to obtain the fused multimodal feature map.
5. The method according to claim 1, characterized in that, The segmentation results for gray matter, white matter, and cerebrospinal fluid include: Based on the fused brain feature map, a pre-trained U-Net network model was used to segment brain tissue, obtaining segmentation results for gray matter, white matter, and cerebrospinal fluid. Morphological operations and connected component analysis were performed on the segmentation results, and the volume indicators of gray matter, white matter, and cerebrospinal fluid were calculated based on the segmentation results to obtain quantitative brain tissue volume information.
6. The method according to claim 1, characterized in that, Constructing brain structural feature vectors includes: Morphological analysis methods were used to extract sulcus depth, cortical thickness, and hippocampal volume from the segmentation results to obtain detailed information on brain structure and morphology. Based on the extracted morphological feature indicators, a feature fusion algorithm was used to combine multiple indicators to construct a high-dimensional feature vector that comprehensively reflects the brain structure and morphology.
7. A brain entropy monitoring system for Alzheimer's patients, characterized in that, include: The data acquisition and registration module is used to acquire the patient's multimodal brain image data and perform spatial alignment processing on the multimodal brain image data using a registration algorithm to obtain the registered multimodal image dataset. The feature extraction module is used to extract discriminative feature maps from the registered multimodal image dataset using a deep learning-based feature extraction network, thereby obtaining a multimodal feature map set. The feature fusion module is used to input the multimodal feature map set into a preset fusion network, and perform weighted fusion processing on the feature maps of different modalities through an adaptive weight allocation mechanism to generate a fused brain feature map. The brain segmentation module is used to segment the brain based on the fused brain feature map using a U-Net network architecture to obtain segmentation results for gray matter, white matter, and cerebrospinal fluid. The feature vector construction module is used to extract morphological features of brain structures based on the segmentation results. The morphological features include sulcus depth, cortical thickness, and hippocampal volume. Based on the morphological features, a brain structure feature vector is constructed. The brain entropy calculation module is used to input brain structural feature vectors into a preset brain entropy calculation model to obtain brain entropy values.
8. The system according to claim 7, characterized in that, The data acquisition and registration module includes: The acquisition unit is used to acquire multimodal brain imaging data from patients. The preprocessing unit is used to perform image enhancement on multimodal brain imaging data using preprocessing techniques; The mutual information matrix construction unit is used to extract the mutual information features of the images based on the preprocessed multimodal image data and construct the mutual information matrix. Alignment units are used to perform spatial coordinate transformation on multimodal image data using a registration algorithm based on mutual information matrices to achieve spatial alignment. The registration unit is used to obtain the transformation matrix through the registration algorithm, and then apply the transformation matrix to the original multimodal image data to complete the registration. The fusion unit is used to fuse the registered multimodal image data to generate a fused multimodal image dataset.