A brain image processing method and system based on a deep neural network
By combining multi-scale wavelet transform and deep neural network classification prediction model, the problems of damaged brain network topology and insufficient exploration of biological pathology in existing technologies are solved, and high-accuracy Alzheimer's disease classification and neuropathological transmission pathway prediction are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2024-03-13
- Publication Date
- 2026-06-26
Smart Images
Figure CN118154546B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of machine learning and image processing, and in particular to a brain image processing method and system based on deep neural networks. Background Technology
[0002] Statistics show that approximately 50 million people worldwide are affected by Alzheimer's disease (AD), and the number is expected to steadily increase over the next decade, making the global public health burden increasingly urgent. As a neurodegenerative disease, Alzheimer's disease manifests abnormalities in certain functional brain regions during its progression. Thanks to advancements in imaging technology, techniques such as magnetic resonance imaging (MRI), structural MRI (sMRI), functional MRI (fMRI), and diffusion MRI (dMRI) can visualize the brain structure of AD patients, providing an effective method for analyzing AD-related abnormal functional brain regions. Clinically, the severity of AD is determined by observing changes in brain structure shown in MRI and other images. However, early-stage AD abnormalities are not readily apparent in various images, making them difficult for doctors to discern with the naked eye. Furthermore, related research indicates that AD is not only associated with localized disturbances in brain structure and function, but these abnormalities can also be reflected in the connectivity between brain regions in AD patients. Therefore, exploring the connectivity of brain regions that is crucial for understanding disease development is urgently needed.
[0003] With the development of deep neural network technology, researchers are considering building classification models based on deep neural networks to classify brain imaging patterns of patients, thereby assisting doctors in the clinical diagnosis of Alzheimer's disease. Current technologies mostly use Euclidean averaging to calculate the Laplacian matrix that preserves the topological relationships of all data when constructing public brain networks. However, this method is often used for two-dimensional calculations, and its application to topological structures can lead to a certain degree of structural loss. In other words, using a simple Euclidean averaging method may result in topological damage. Furthermore, existing Alzheimer's disease classification models neglect discussions and explorations at the biological pathology level, failing to discover neuropathological transmission patterns related to AD. Summary of the Invention
[0004] To address the aforementioned problems, this invention proposes a brain image processing method and system based on deep neural networks. It constructs an Alzheimer's disease classification and prediction model based on deep neural networks to classify brain imaging images of patients, improving the accuracy of the model's classification and enabling the prediction of neuropathological transmission pathways of Alzheimer's disease.
[0005] In a first aspect, embodiments of the present invention provide a brain image processing method based on a deep neural network, comprising:
[0006] Obtain the brain imaging images and brain PET images to be predicted;
[0007] Brain network data is generated by extracting and generating fiber bundles from each brain region in the brain images.
[0008] Protein deposition data for each brain region were extracted from the brain PET images.
[0009] The brain network data and protein deposition data are input into a preset classification prediction model so that the classification prediction model generates the corresponding Alzheimer's disease classification result and updates the weight matrix inside the model. The classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, and the weight matrix is the weight matrix in the deep neural network.
[0010] The neuropathological transmission pathways of Alzheimer's disease between the various brain regions are predicted based on the weight matrix.
[0011] This invention provides a brain image processing method based on deep neural networks. It acquires brain images and PET images to be predicted, extracting corresponding brain network data and protein deposition data, which are then input into a preset classification prediction model. The model generates corresponding Alzheimer's disease classification results based on the brain network data and protein deposition data, achieving classification and identification of Alzheimer's disease based on brain images and PET images. Furthermore, the model incorporates a multi-scale wavelet transform model and a deep neural network during model construction, improving classification accuracy. After each classification, the model automatically updates its internal wavelet operators and weight matrices at various scales. The wavelet operators describe the connectivity information between different brain regions, while the information in the weight matrix serves as the basis for selecting brain region pathways. Combining the weight matrix with the wavelet operators allows for further prediction of the neuropathological propagation pathways of Alzheimer's disease between these brain regions.
[0012] Furthermore, the step of extracting and generating brain network data based on the number of fiber bundles in each brain region of the brain image includes:
[0013] The number of fiber bundles in each brain region is calculated based on the brain images. Then, the correlation between the number of fiber bundles in each brain region is calculated to obtain the association information between the brain regions and construct the corresponding adjacency matrix to form the brain network data.
[0014] In one possible implementation, the step of inputting the brain network data and protein deposition data into a preset classification prediction model, so that the classification prediction model generates a corresponding Alzheimer's disease classification result, includes:
[0015] A corresponding brain network matrix is constructed based on the brain network data using a pre-defined public brain network construction model;
[0016] The brain network matrix is subjected to wavelet transform using the multi-scale wavelet transform model to obtain multiple wavelet operators of different scales.
[0017] By using wavelet operators of different scales, feature extraction is performed on the protein deposition data to obtain feature expression data of the protein deposition data at different scales.
[0018] The feature representation data is input into the deep neural network for classification to obtain the Alzheimer's disease classification result.
[0019] This invention provides a method for classifying brain network data and protein deposition data using a classification prediction model. First, a pre-defined public brain network construction model is used to construct a corresponding brain network matrix based on the brain network data. This ensures the integrity of the brain network topology information, improves the model's classification accuracy, and allows for further prediction of neuropathological transmission pathways based on the brain network topology information. Then, a multi-scale wavelet transform model is used to perform wavelet transform on the brain network matrix, extracting the brain network topology information in the form of wavelet operators at multiple different scales. These wavelet operators are then used to extract features from the protein deposition data, effectively fusing the brain network topology information and the protein deposition data to obtain feature expression data at different scales. Finally, this data is input into a deep neural network for classification. Because the input feature expression data has multiple scales and combines brain network topology information and protein deposition data, the Alzheimer's disease classification result generated by the neural network is more accurate, improving the classification accuracy of the classification prediction model.
[0020] In one possible implementation, predicting the neuropathological propagation pathways of Alzheimer's disease between the various brain regions based on the weight matrix includes:
[0021] Sort each column of the weight matrix in descending order;
[0022] Calculate the average value of the first M1 elements in each column of the sorted weight matrix to obtain the average weight vector;
[0023] Determine the M2 elements with the largest values from the average weight vector and obtain the indices corresponding to the M2 elements;
[0024] Match the corresponding wavelet basis from each wavelet operator according to the subscripts corresponding to the M2 elements;
[0025] The neuropathological propagation pathways of Alzheimer's disease between the respective brain regions are predicted based on the wavelet basis, wherein M1 and M2 are preset values.
[0026] This invention provides a method for predicting the neuropathological propagation pathway of Alzheimer's disease based on a weight matrix. By sorting and averaging the weight matrix, the top M2 elements with the largest values in the average weight vector are determined, which are the most likely propagation pathways. Then, based on the indices of the M2 elements, the corresponding wavelet bases are matched from various wavelet operators. Since the wavelet operators are generated based on the brain network matrix and contain the original brain network topology information, the neuropathological propagation pathways of Alzheimer's disease between the various brain regions can be predicted based on the wavelet bases.
[0027] In one possible implementation, the classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, including:
[0028] An initial classification prediction model is constructed based on an initial multi-scale wavelet transform model and an initial deep neural network.
[0029] Acquire several sets of training data, each set including brain network data, corresponding protein deposition data, and corresponding Alzheimer's disease diagnosis results;
[0030] The aforementioned sets of training data are input into the initial classification prediction model for training, so that the initial classification prediction model continuously updates the initial multi-scale wavelet transform model and the initial deep neural network during the training process, thereby obtaining the classification prediction model.
[0031] This invention provides a method for constructing and training the classification prediction model. An initial classification prediction model is constructed by combining a multi-scale wavelet transform model and an initial deep neural network. Multiple sets of training data are input into the initial classification prediction model. Since the training data includes Alzheimer's disease diagnosis results, based on the characteristics of deep neural networks, the deep neural network will verify the results against real diagnoses during training, thereby continuously updating its training parameters and gradually improving classification accuracy. Finally, the classification prediction model is obtained through training.
[0032] Furthermore, the step of inputting the plurality of sets of training data into the initial classification prediction model for training, so that the initial classification prediction model continuously updates the initial multi-scale wavelet transform model and the initial deep neural network during the training process, includes:
[0033] A public brain network matrix is constructed based on several brain network data in the training data using a pre-defined public brain network construction model. The public brain network matrix is used to store the brain network topology information of all training data.
[0034] The common brain network matrix is subjected to wavelet transform using the initial multi-scale wavelet transform model to obtain multiple initial wavelet operators of different scales.
[0035] Several protein deposition data in the training data are sequentially multiplied with the initial wavelet operator and then input into the initial deep neural network to obtain the corresponding Alzheimer's disease classification result. After obtaining the Alzheimer's disease classification result each time, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function.
[0036] This embodiment of the invention further illustrates the training process of the initial classification prediction model. First, a common brain network matrix is constructed based on several brain network data, and the brain network topology information of all training data is saved in the common brain network matrix. This allows the deep neural network to simultaneously receive several protein deposition data for training during subsequent training. During the training process, a multi-scale wavelet transform model is combined to automatically update and explore the propagation pathway. After each Alzheimer's disease classification result is obtained, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function to achieve automatic iterative updates and improve the training efficiency of the model.
[0037] Furthermore, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function, specifically using the following formula:
[0038]
[0039] Where L M The binary cross-entropy between the model's classification results and the actual diagnostic results is used as... express, For the classification results of the model, This is the true diagnostic result; L S The constraint term affecting the automatic update of the multi-scale wavelet transform model and the deep neural network is represented by θ1∥A∥1+θ2∥S∥2, where θ1 and θ2 are adjustable parameters, A is the current weight matrix of the neural network input layer, and S is the set of scale parameters s1, s2, ..., sk of the multi-scale wavelet transform model.
[0040] Secondly, embodiments of the present invention provide a brain image processing system based on a deep neural network, including an acquisition module, a first extraction module, a second extraction module, a classification module, and a prediction module;
[0041] The acquisition module is used to acquire the brain imaging image and brain PET image to be predicted;
[0042] The first extraction module is used to extract and generate brain network data based on the number of fiber bundles in each brain region of the brain image.
[0043] The second extraction module is used to extract protein deposition data of each brain region from the brain PET image;
[0044] The classification module is used to input the brain network data and protein deposition data into a preset classification prediction model, so that the classification prediction model generates the corresponding Alzheimer's disease classification result and updates the weight matrix inside the model. The classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, and the weight matrix is the weight matrix in the deep neural network.
[0045] The prediction module is used to predict the neuropathological transmission pathways of Alzheimer's disease between the various brain regions based on the weight matrix.
[0046] Furthermore, the first extraction module extracts brain network data from the brain image to represent the connections between various brain regions, including:
[0047] The number of fiber bundles in each brain region is calculated based on the brain images. Then, the correlation between the number of fiber bundles in each brain region is calculated to obtain the association information between the brain regions and construct the corresponding adjacency matrix to form the brain network data.
[0048] In one possible implementation, the classification module includes a brain network matrix construction unit, a wavelet transform unit, a feature extraction unit, and a classification unit;
[0049] The brain network matrix construction unit is used to construct a corresponding brain network matrix based on the brain network data using a preset public brain network construction model.
[0050] The wavelet transform unit is used to perform wavelet transform on the brain network matrix through the multi-scale wavelet transform model to obtain multiple wavelet operators of different scales.
[0051] The feature extraction unit is used to extract features from the protein deposition data using multiple wavelet operators of different scales to obtain the feature expression data of the protein deposition data at different scales.
[0052] The classification unit is used to input the feature expression data into the deep neural network for classification, and obtain the Alzheimer's disease classification result. Attached Figure Description
[0053] Figure 1 : A schematic flowchart of an embodiment of a brain image processing method based on a deep neural network provided by the present invention.
[0054] Figure 2 This is a schematic diagram illustrating the process of generating corresponding Alzheimer's disease classification results through a classification prediction model in a brain image processing method based on deep neural networks provided by the present invention.
[0055] Figure 3 This is a schematic diagram illustrating the process of predicting the neuropathological propagation pathway of Alzheimer's disease based on a weight matrix in a brain image processing method based on a deep neural network provided by the present invention.
[0056] Figure 4 This is a flowchart illustrating a practical application of the brain image processing method based on deep neural networks provided by the present invention.
[0057] Figure 5 This is a schematic diagram illustrating the process of constructing a classification prediction model in a brain image processing method based on deep neural networks provided by the present invention.
[0058] Figure 6 : A schematic diagram of an embodiment of a brain image processing system based on a deep neural network provided by the present invention.
[0059] Figure 7 : A schematic diagram of the structure of a classification module in a brain image processing system based on a deep neural network provided by the present invention.
[0060] Figure 8 : A schematic diagram of the prediction module in a brain image processing system based on a deep neural network provided by the present invention.
[0061] Figure 9 : A schematic diagram of the structure of a model building module in a brain image processing system based on a deep neural network provided by the present invention. Detailed Implementation
[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0063] It should be noted that the step numbers in this document are only for the convenience of explaining the specific embodiments and are not intended to limit the order in which the steps are performed.
[0064] Throughout this specification, the brain imaging images described include those obtained using techniques such as magnetic resonance imaging (MRI), structural magnetic resonance imaging (SMRI), functional magnetic resonance imaging (fMRI), and diffusion-weighted magnetic resonance imaging (DMRI). The PET images described are positron emission tomography (PET), a relatively advanced clinical imaging technique in nuclear medicine. It involves labeling essential substances for biological metabolism (glucose, proteins, nucleic acids, etc.) with short-lived radioactive isotopes (such as F18, carbon-11, etc.) and injecting them into the human body. The accumulation of these substances during metabolism reflects the state of metabolic activities. After labeling, positron collimation technology is used to non-invasively, quantitatively, and dynamically determine the spatial distribution, quantity, and dynamic changes of the PET imaging agent or its metabolites in vitro or in vivo. This provides imaging information at the molecular level of the biochemical, physiological, and functional metabolic changes resulting from the interaction between the PET imaging agent and the target site in vivo.
[0065] Example 1:
[0066] like Figure 1 As shown, Embodiment 1 provides a brain image processing method based on deep neural networks, including steps S1-S5:
[0067] Step S1: Obtain the brain imaging image and brain PET image to be predicted;
[0068] Step S2: Extract and generate brain network data based on the number of fiber bundles in each brain region in the brain image;
[0069] Step S3: Extract protein deposition data of each brain region from the brain PET image;
[0070] Step S4: Input the brain network data and protein deposition data into a preset classification prediction model so that the classification prediction model generates the corresponding Alzheimer's disease classification result and updates the weight matrix inside the model. The classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, and the weight matrix is the weight matrix in the deep neural network.
[0071] Step S5: Predict the neuropathological transmission pathways of Alzheimer's disease between the various brain regions based on the weight matrix.
[0072] This invention provides a brain image processing method based on deep neural networks. It acquires brain images and PET images to be predicted, extracting corresponding brain network data and protein deposition data, which are then input into a preset classification prediction model. The model generates corresponding Alzheimer's disease classification results based on the brain network data and protein deposition data, achieving classification and identification of Alzheimer's disease based on brain images and PET images. Furthermore, the model incorporates a multi-scale wavelet transform model and a deep neural network during model construction, improving classification accuracy. After each classification, the model automatically updates its internal wavelet operators and weight matrices at various scales. The wavelet operators describe the connectivity information between different brain regions, while the information in the weight matrix serves as the basis for selecting brain region pathways. Combining the weight matrix with the wavelet operators allows for further prediction of the neuropathological propagation pathways of Alzheimer's disease between these brain regions.
[0073] Furthermore, in step S2, the extraction and generation of brain network data based on the number of fiber bundles in each brain region of the brain image includes:
[0074] The number of fiber bundles in each brain region is calculated based on the brain images. Then, the correlation between the number of fiber bundles in each brain region is calculated to obtain the association information between the brain regions and construct the corresponding adjacency matrix to form the brain network data.
[0075] In one possible implementation, in step S4, the brain network data and protein deposition data are input into a preset classification prediction model, so that the classification prediction model generates a corresponding Alzheimer's disease classification result, such as... Figure 2 As shown, steps S401-S404 are included:
[0076] Step S401: Construct a corresponding brain network matrix based on the brain network data using a preset public brain network construction model;
[0077] Step S402: Perform wavelet transform on the brain network matrix using the multi-scale wavelet transform model to obtain multiple wavelet operators of different scales;
[0078] Step S403: Extract features from the protein deposition data using wavelet operators of different scales to obtain feature representation data of the protein deposition data at different scales;
[0079] Step S404: Input the feature expression data into the deep neural network for classification to obtain the Alzheimer's disease classification result.
[0080] This invention provides a method for classifying brain network data and protein deposition data using a classification prediction model. First, a pre-defined public brain network construction model is used to construct a corresponding brain network matrix based on the brain network data. This ensures the integrity of the brain network topology information, improves the model's classification accuracy, and allows for further prediction of neuropathological transmission pathways based on the brain network topology information. Then, a multi-scale wavelet transform model is used to perform wavelet transform on the brain network matrix, extracting the brain network topology information in the form of wavelet operators at multiple different scales. These wavelet operators are then used to extract features from the protein deposition data, effectively fusing the brain network topology information and the protein deposition data to obtain feature expression data at different scales. Finally, this data is input into a deep neural network for classification. Because the input feature expression data has multiple scales and combines brain network topology information and protein deposition data, the Alzheimer's disease classification result generated by the neural network is more accurate, improving the classification accuracy of the classification prediction model.
[0081] In one possible implementation, in step S5, the prediction of the neuropathological propagation pathways of Alzheimer's disease between the various brain regions based on the weight matrix, such as... Figure 3 As shown, steps S501-S505 are included:
[0082] Step S501: Sort each column of the weight matrix in descending order;
[0083] Step S502: Calculate the average value of the first M1 elements in each column of the sorted weight matrix to obtain the average weight vector;
[0084] Step S503: Determine the top M2 elements with the largest values from the average weight vector and obtain the indices corresponding to the M elements;
[0085] Step S504: Match the corresponding wavelet basis from each wavelet operator according to the subscripts corresponding to the M2 elements;
[0086] Step S505: Predict the neuropathological transmission pathway of Alzheimer's disease between the respective brain regions based on the respective wavelet bases, wherein M1 and M2 are preset values.
[0087] This invention provides a method for predicting the neuropathological propagation pathway of Alzheimer's disease based on a weight matrix. By sorting and averaging the weight matrix, the top M2 elements with the largest values in the average weight vector are determined, which are the most likely propagation pathways. Then, based on the indices of the M2 elements, the corresponding wavelet bases are matched from various wavelet operators. Since the wavelet operators are generated based on the brain network matrix and contain the original brain network topology information, the neuropathological propagation pathways of Alzheimer's disease between the various brain regions can be predicted based on the wavelet bases.
[0088] In a preferred embodiment, this method can simultaneously process brain images from multiple patients for classification and prediction, such as... Figure 4 As shown, specifically:
[0089] The number of fiber tracts in each brain region was calculated based on MRI images of M patients. Then, the correlation between the number of fiber tracts in each brain region was calculated to obtain the association information (adjacency information) between brain regions throughout the entire brain. This integrated association information between brain regions forms the structural brain network W∈R. N×N N is the number of brain regions, and W is the adjacency matrix. Protein deposition data F for each brain region of M patients is obtained from PET images of each patient. Each patient may have multiple PET images from different scan times.
[0090] Structural brain network data from M patients are input into the public brain network construction module StiefelSpace to obtain a public brain network that can store the topological structure information of the brain networks of all samples.
[0091]
[0092] Public brain network Inputting the multi-scale wavelet transform module WaveletTransform yields multiple wavelet operators T at different scales. S＝{s1,s2…sk} ,
[0093]
[0094] The protein deposition data F is input into the multi-scale wavelet transform module WaveletTransform to obtain the feature representations F of data F at different scales. T ,
[0095] F T =T S ×F
[0096] Feature representation F T The input is fed into a deep neural network for semi-supervised classification prediction, yielding the classification result y and the weight matrix A.
[0097] y,A=DNN(F T )
[0098] Take the weight matrix A after model training, sort it in descending order by column, and calculate the average of the first 10 rows of data in each column.
[0099]
[0100] Where k represents the number of scale parameters S = {s1, s2, ..., sk}.
[0101] right Sort the elements in descending order and extract the indices of the 10 elements with the largest weight values.
[0102]
[0103] Finally, the multi-scale wavelet operator T was found. S The wavelet basis T indicated by the corresponding subscript index in the middle index These small wave bases can indicate specific transmission pathways in the brain.
[0104] In one possible implementation, the classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, such as... Figure 5 As shown, steps S601-S603 are included:
[0105] Step S601: Construct an initial classification prediction model based on the initial multi-scale wavelet transform model and the initial deep neural network;
[0106] Step S602: Obtain several sets of training data, wherein each set of data includes brain network data, corresponding protein deposition data and corresponding Alzheimer's disease diagnosis results;
[0107] Step S603: Input the several sets of training data into the initial classification prediction model for training, so that the initial classification prediction model continuously updates the initial multi-scale wavelet transform model and the initial deep neural network during the training process to obtain the classification prediction model.
[0108] This invention provides a method for constructing and training the classification prediction model. An initial classification prediction model is constructed by combining a multi-scale wavelet transform model and an initial deep neural network. Multiple sets of training data are input into the initial classification prediction model. Since the training data includes Alzheimer's disease diagnosis results, based on the characteristics of deep neural networks, the deep neural network will verify the results against real diagnoses during training, thereby continuously updating its training parameters and gradually improving classification accuracy. Finally, the classification prediction model is obtained through training.
[0109] Furthermore, in step S603, the step of inputting the plurality of sets of training data into the initial classification prediction model for training, so that the initial classification prediction model continuously updates the initial multi-scale wavelet transform model and the initial deep neural network during the training process, includes:
[0110] A public brain network matrix is constructed based on several brain network data in the training data using a pre-defined public brain network construction model. The public brain network matrix is used to store the brain network topology information of all training data.
[0111] The common brain network matrix is subjected to wavelet transform using the initial multi-scale wavelet transform model to obtain multiple initial wavelet operators of different scales.
[0112] Several protein deposition data in the training data are sequentially multiplied with the initial wavelet operator and then input into the initial deep neural network to obtain the corresponding Alzheimer's disease classification result. After obtaining the Alzheimer's disease classification result each time, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function.
[0113] This embodiment of the invention further illustrates the training process of the initial classification prediction model. First, a common brain network matrix is constructed based on several brain network data, and the brain network topology information of all training data is saved in the common brain network matrix. This allows the deep neural network to simultaneously receive several protein deposition data for training during subsequent training. During the training process, a multi-scale wavelet transform model is combined to automatically update and explore the propagation pathway. After each Alzheimer's disease classification result is obtained, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function to achieve automatic iterative updates and improve the training efficiency of the model.
[0114] Furthermore, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function, specifically using the following formula:
[0115]
[0116] Where L M The binary cross-entropy between the model's classification results and the actual diagnostic results is used as... express, For the classification results of the model, This is the true diagnostic result; L S The constraint term affecting the automatic update of the multi-scale wavelet transform model and the deep neural network is represented by θ1∥A∥1+θ2∥S∥2, where θ1 and θ2 are adjustable parameters, A is the current weight matrix of the neural network input layer, and S is the set of scale parameters s1, s2, ..., sk of the multi-scale wavelet transform model.
[0117] In a preferred embodiment, the training process for the initial deep neural network specifically involves: dividing the biological protein deposition data F into a training set Fi. 1 and test set F 2 Each protein deposition data point has a corresponding actual diagnostic result. true ={CN,EMCI,LMCI}(CN: normal cognition, EMCI: early mild cognitive impairment, LMCI: late mild cognitive impairment).
[0118] 1) First, set the training set F 1 and the true labels corresponding to the training set 1) A training set is formed by randomly selecting 32 data points for each group (x=32) and input into the model for training; 2) The input data is multiplied by a multi-scale wavelet operator within the model to extract feature information at different scales; 3) This information is combined row by row and input into the neural network for training. After training, the model outputs its prediction results. Calculate the binary cross-entropy between the predicted results and the actual diagnostic results:
[0119]
[0120] To automatically explore neuropathological transmission pathways related to Alzheimer's disease (AD) during training, this invention embeds a multi-scale wavelet transform module into the deep model and designs the final loss function as follows:
[0121]
[0122] After training, a model with the ability to distinguish different developmental stages of Alzheimer's disease (AD) is obtained.
[0123] Then test set F 2 The data is fed into the model for classification prediction to obtain the classification result. Will and the real labels corresponding to the test set By comparing the models, we can evaluate their various performance indicators.
[0124] Example 2:
[0125] like Figure 6 As shown, Embodiment 2 provides a brain image processing system based on a deep neural network, including an acquisition module 10, a first extraction module 20, a second extraction module 30, a classification module 40, and a prediction module 50.
[0126] The acquisition module 10 is used to acquire the brain imaging image and brain PET image to be predicted;
[0127] The first extraction module 20 is used to extract and generate brain network data based on the number of fiber bundles in each brain region in the brain image.
[0128] The second extraction module 30 is used to extract protein deposition data of each brain region from the brain PET image;
[0129] The classification module 40 is used to input the brain network data and protein deposition data into a preset classification prediction model, so that the classification prediction model generates the corresponding Alzheimer's disease classification result and updates the weight matrix inside the model. The classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, and the weight matrix is the weight matrix in the deep neural network.
[0130] The prediction module 50 is used to predict the neuropathological transmission pathways of Alzheimer's disease between the various brain regions based on the weight matrix.
[0131] Furthermore, the first extraction module 20 extracts brain network data from the brain image to represent the connections between various brain regions, including:
[0132] The number of fiber bundles in each brain region is calculated based on the brain images. Then, the correlation between the number of fiber bundles in each brain region is calculated to obtain the association information between the brain regions and construct the corresponding adjacency matrix to form the brain network data.
[0133] In one possible implementation, such as Figure 7 As shown, the classification module 40 includes a brain network matrix construction unit 401, a wavelet transform unit 402, a feature extraction unit 403, and a classification unit 404.
[0134] The brain network matrix construction unit 401 is used to construct a corresponding brain network matrix based on the brain network data using a preset public brain network construction model.
[0135] The wavelet transform unit 402 is used to perform wavelet transform on the brain network matrix through the multi-scale wavelet transform model to obtain multiple wavelet operators of different scales.
[0136] The feature extraction unit 403 is used to extract features from the protein deposition data using multiple wavelet operators of different scales to obtain the feature expression data of the protein deposition data at different scales.
[0137] The classification unit 404 is used to input the feature expression data into the deep neural network for classification to obtain the Alzheimer's disease classification result.
[0138] In one possible implementation, such as Figure 8 As shown, the prediction module 50 includes a sorting unit 501, a calculation unit 502, an index acquisition unit 503, an index matching unit 504, and a prediction unit 505.
[0139] The sorting unit 501 is used to sort each column of the weight matrix in descending order.
[0140] The calculation unit 502 is used to calculate the average value of the first M1 elements of each column in the sorted weight matrix to obtain the average weight vector.
[0141] The subscript acquisition unit 503 is used to determine the top M2 elements with the largest values from the average weight vector and obtain the subscripts corresponding to the M2 elements;
[0142] The subscript matching unit 504 is used to match the corresponding wavelet basis from each wavelet operator according to the subscripts corresponding to the M2 elements;
[0143] The prediction unit 505 is used to predict the neuropathological propagation pathway of Alzheimer's disease between the various brain regions based on the respective wavelet bases, wherein M1 and M2 are preset values.
[0144] In one possible implementation, such as Figure 9 As shown, the brain image processing system further includes a model building module 60, which is used to build and train the classification prediction model based on a multi-scale wavelet transform model and a deep neural network, including a building unit 601, an acquisition unit 602, and a training unit 603.
[0145] The construction unit 601 is used to construct an initial classification prediction model based on the initial multi-scale wavelet transform model and the initial deep neural network.
[0146] The acquisition unit 602 is used to acquire several sets of training data, wherein each set of data includes brain network data, corresponding protein deposition data and corresponding Alzheimer's disease diagnosis results.
[0147] The training unit 603 is used to input the several sets of training data into the initial classification prediction model for training, so that the initial classification prediction model continuously updates the initial multi-scale wavelet transform model and the initial deep neural network during the training process to obtain the classification prediction model.
[0148] Furthermore, the training unit 603 inputs the plurality of sets of training data into the initial classification prediction model for training, so that the initial classification prediction model continuously updates the initial multi-scale wavelet transform model and the initial deep neural network during the training process, including:
[0149] A public brain network matrix is constructed based on several brain network data in the training data using a pre-defined public brain network construction model. The public brain network matrix is used to store the brain network topology information of all training data.
[0150] The common brain network matrix is subjected to wavelet transform using the initial multi-scale wavelet transform model to obtain multiple initial wavelet operators of different scales.
[0151] Several protein deposition data in the training data are sequentially multiplied with the initial wavelet operator and then input into the initial deep neural network to obtain the corresponding Alzheimer's disease classification result. After obtaining the Alzheimer's disease classification result each time, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function.
[0152] Furthermore, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function, specifically using the following formula:
[0153]
[0154] Where L M The binary cross-entropy between the model's classification results and the actual diagnostic results is used as... express, For the classification results of the model, This is the true diagnostic result; L S The constraint term affecting the automatic update of the multi-scale wavelet transform model and the deep neural network is represented by θ1∥A∥1+θ2∥S∥2, where θ1 and θ2 are adjustable parameters, A is the current weight matrix of the neural network input layer, and S is the set of scale parameters s1, s2, ..., sk of the multi-scale wavelet transform model.
[0155] This invention provides a brain image processing system based on a deep neural network. It acquires brain images and PET images to be predicted, extracting corresponding brain network data and protein deposition data, which are then input into a preset classification prediction model. The model generates corresponding Alzheimer's disease classification results based on the brain network data and protein deposition data, achieving classification and identification of Alzheimer's disease based on brain images and PET images. Furthermore, the system combines a multi-scale wavelet transform model and a deep neural network in the construction of the classification prediction model, improving classification accuracy. After each classification, the model automatically updates its internal wavelet operators and weight matrices at various scales. The wavelet operators describe the connectivity information between different brain regions, while the information in the weight matrix serves as the basis for selecting brain region pathways. Combining the weight matrix with the wavelet operators allows for further prediction of the neuropathological propagation pathways of Alzheimer's disease between these brain regions.
[0156] For a more detailed explanation of the working principle and procedures of this embodiment, please refer to the relevant description in Embodiment 1.
[0157] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.
Claims
1. A brain image processing method based on deep neural networks, characterized in that, include: Obtain the brain imaging images and brain PET images to be predicted; Brain network data is generated by extracting and generating fiber bundles from each brain region in the brain images. Protein deposition data for each brain region were extracted from the brain PET images. The brain network data and protein deposition data are input into a preset classification prediction model so that the classification prediction model generates the corresponding Alzheimer's disease classification result and updates the weight matrix inside the model. The classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, and the weight matrix is the weight matrix in the deep neural network. Predicting the neuropathological propagation pathway of Alzheimer's disease between various brain regions based on the weight matrix includes: sorting each column of the weight matrix in descending order; calculating the average value of the first M1 elements of each column in the sorted weight matrix to obtain an average weight vector; determining the first M2 elements with the largest values from the average weight vector and obtaining the indices corresponding to the M2 elements; matching the corresponding wavelet basis from each wavelet operator based on the indices corresponding to the M2 elements; and predicting the neuropathological propagation pathway of Alzheimer's disease between various brain regions based on each wavelet basis, wherein M1 and M2 are preset values.
2. The brain image processing method based on a deep neural network as described in claim 1, characterized in that, The step of extracting and generating brain network data based on the number of fiber bundles in each brain region of the brain image includes: The number of fiber bundles in each brain region is calculated based on the brain images. Then, the correlation between the number of fiber bundles in each brain region is calculated to obtain the association information between the brain regions and construct the corresponding adjacency matrix to form the brain network data.
3. The brain image processing method based on a deep neural network as described in claim 1, characterized in that, The step of inputting the brain network data and protein deposition data into a preset classification prediction model, so that the classification prediction model generates a corresponding Alzheimer's disease classification result, includes: A corresponding brain network matrix is constructed based on the brain network data using a pre-defined public brain network construction model; The brain network matrix is subjected to wavelet transform using the multi-scale wavelet transform model to obtain multiple wavelet operators of different scales. By using wavelet operators of different scales, feature extraction is performed on the protein deposition data to obtain feature expression data of the protein deposition data at different scales. The feature representation data is input into the deep neural network for classification to obtain the Alzheimer's disease classification result.
4. The brain image processing method based on a deep neural network as described in claim 1, characterized in that, The classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, including: An initial classification prediction model is constructed based on an initial multi-scale wavelet transform model and an initial deep neural network. Acquire several sets of training data, each set including brain network data, corresponding protein deposition data, and corresponding Alzheimer's disease diagnosis results; The aforementioned sets of training data are input into the initial classification prediction model for training, so that the initial classification prediction model continuously updates the initial multi-scale wavelet transform model and the initial deep neural network during the training process, thereby obtaining the classification prediction model.
5. The brain image processing method based on a deep neural network as described in claim 4, characterized in that, The step of inputting the plurality of sets of training data into the initial classification prediction model for training, so that the initial classification prediction model continuously updates the initial multi-scale wavelet transform model and the initial deep neural network during the training process, includes: A public brain network matrix is constructed based on several brain network data in the training data using a pre-defined public brain network construction model. The public brain network matrix is used to store the brain network topology information of all training data. The common brain network matrix is subjected to wavelet transform using the initial multi-scale wavelet transform model to obtain multiple initial wavelet operators of different scales. Several protein deposition data in the training data are sequentially multiplied with the initial wavelet operator and then input into the initial deep neural network to obtain the corresponding Alzheimer's disease classification result. After obtaining the Alzheimer's disease classification result each time, the initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function.
6. The brain image processing method based on a deep neural network as described in claim 5, characterized in that, The initial multi-scale wavelet transform model and the initial deep neural network are updated based on a preset loss function, and the specific formula is as follows: in The binary cross-entropy between the model's classification results and the actual diagnostic results is used as... express, For the classification results of the model, This is the true diagnostic result; To be used as a constraint term affecting the automatic updating of multi-scale wavelet transform models and deep neural networks, express, It is an adjustable parameter. This is the current weight matrix of the neural network input layer, and S is the scale parameter of the multi-scale wavelet transform model. A set of.
7. A brain image processing system based on a deep neural network, characterized in that, It includes an acquisition module, a first extraction module, a second extraction module, a classification module, and a prediction module; The acquisition module is used to acquire the brain imaging image and brain PET image to be predicted; The first extraction module is used to extract and generate brain network data based on the number of fiber bundles in each brain region of the brain image. The second extraction module is used to extract protein deposition data of each brain region from the brain PET image; The classification module is used to input the brain network data and protein deposition data into a preset classification prediction model, so that the classification prediction model generates the corresponding Alzheimer's disease classification result and updates the weight matrix inside the model. The classification prediction model is constructed and trained based on a combination of a multi-scale wavelet transform model and a deep neural network, and the weight matrix is the weight matrix in the deep neural network. The prediction module is used to predict the neuropathological propagation pathway of Alzheimer's disease between various brain regions based on the weight matrix, including: sorting each column of the weight matrix in descending order; calculating the average value of the first M1 elements of each column in the sorted weight matrix to obtain an average weight vector; determining the first M2 elements with the largest values from the average weight vector and obtaining the indices corresponding to the M2 elements; matching the corresponding wavelet basis from each wavelet operator based on the indices corresponding to the M2 elements; and predicting the neuropathological propagation pathway of Alzheimer's disease between various brain regions based on each wavelet basis, wherein M1 and M2 are preset values.
8. The brain image processing system based on a deep neural network as described in claim 7, characterized in that, The first extraction module extracts brain network data representing the connections between various brain regions from the brain image, including: The number of fiber bundles in each brain region is calculated based on the brain images. Then, the correlation between the number of fiber bundles in each brain region is calculated to obtain the association information between the brain regions and construct the corresponding adjacency matrix to form the brain network data.
9. A brain image processing system based on a deep neural network as described in claim 7, characterized in that, The classification module includes a brain network matrix construction unit, a wavelet transform unit, a feature extraction unit, and a classification unit; The brain network matrix construction unit is used to construct a corresponding brain network matrix based on the brain network data using a preset public brain network construction model. The wavelet transform unit is used to perform wavelet transform on the brain network matrix through the multi-scale wavelet transform model to obtain multiple wavelet operators of different scales. The feature extraction unit is used to extract features from the protein deposition data using multiple wavelet operators of different scales to obtain the feature expression data of the protein deposition data at different scales. The classification unit is used to input the feature expression data into the deep neural network for classification, and obtain the Alzheimer's disease classification result.