Brain activation map prediction method and device
Through deep neural networks and state extraction algorithms, brain states are extracted from resting state fMRI images, and combined with task coding and signal reconstruction algorithms, brain activation maps under different tasks are predicted, solving the problems of high cost and difficult to capture individual characteristics in the existing technology, and achieving efficient and individualized brain activation map prediction.
Patent Information
- Application Number
- CN202411831667.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-04-01
AI Technical Summary
When calculating task-state brain activation maps, the prior art relies on task-state fMRI data acquisition, which is costly and difficult to capture individual features.
Through the deep neural network model and state extraction algorithm, brain states are extracted from resting state fMRI images, and combined with task coding algorithms and signal reconstruction algorithms, brain activation maps under different tasks are predicted.
No task-state data collection is required, which reduces experimental costs, improves the individual analysis capabilities of brain imaging research, and can effectively predict brain activation maps under different tasks.
Smart Images
Figure CN119732671A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a brain activation map prediction method and a brain activation map prediction device. Background Art
[0002] The task-based activation map is an analysis tool based on functional magnetic resonance imaging (fMRI) that is used to display brain activity areas triggered by specific tasks or stimuli. By recording changes in blood oxygenation level dependent (BOLD) signals in different brain regions when performing specific tasks, researchers can identify which brain regions are related to the task. Brain activation maps provide important references for personalized diagnosis, treatment plan formulation, and preoperative planning, especially in complex surgeries such as brain tumor resection, helping to avoid damage to important functional areas.
[0003] At present, the calculation methods of task-state brain activation maps are mainly divided into hypothesis-driven methods and data-driven methods. The traditional generalized linear model (GLM) method relies on hypothesis-driven. GLM assumes that there is a linear relationship between BOLD signal changes and task stimulation, and regresses the task design matrix with the BOLD signal to detect brain areas that are significantly activated under task conditions. This method not only relies on a clear task model and cannot capture complex and nonlinear brain activities, but also only focuses on the group average effect and is difficult to capture individual characteristics. Independent component analysis (ICA) is a data-driven method that decomposes fMRI data into independent spatial components to identify activation patterns related to tasks or noise. ICA is often used for exploratory analysis, such as discovering unknown brain networks. This method makes it difficult to determine which components are meaningful, and the neurobiological significance of the components is not easy to clarify.
[0004] The calculation of task-based brain activation maps mainly relies on the acquisition, preprocessing, statistical modeling and multiple comparison correction of task-based fMRI (t-fMRI) data. Task-based fMRI relies on clear experimental design, different tasks may require a lot of time and resources, and individual differences and noise such as head movement may affect the results. Summary of the invention
[0005] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a brain activation map prediction method, which can use resting-state fMRI to predict brain activation maps under different tasks. It not only does not require task-state data collection, reducing experimental costs, but also improves the individualized analysis capabilities of brain imaging research.
[0006] The technical solution of the present invention is: a method for predicting brain activation maps, which comprises the following steps:
[0007] (1) Preprocessing of resting-state fMRI images;
[0008] (2) The preprocessed resting-state fMRI images are used to extract brain states through a deep neural network model and a state extraction algorithm to obtain basic state information of the brain functional network;
[0009] (3) Based on the basic state information of the brain functional network, different task paradigm stimulus information is added through the task coding algorithm to obtain the brain functional network state under specific tasks;
[0010] (4) The brain functional network state under a specific task is input into the deep neural network model and the signal reconstruction algorithm for regression reconstruction to obtain the task-state fMRI;
[0011] (5) Through the deep learning model and activation map prediction algorithm, the regressed task state
[0012] fMRI predictions were used to obtain task-state brain activation maps.
[0013] The present invention directly uses different task paradigms to guide the modification of the basic state of the brain functional network to obtain the brain functional network state under different tasks, thereby realizing the transition of the brain state from resting to a specific task, and then predicting the task-state fMRI and brain activation map through a deep learning model. The prediction of brain activation maps in different task states can be realized by modifying the task paradigm added to the basic state of the brain functional network. There is no need for the subjects to undergo multiple complex task-state fMRI scans, which reduces time and medical costs and improves the individualized analysis capabilities of brain imaging research.
[0014] A brain activation map prediction device is also provided, comprising:
[0015] a preprocessing module configured to preprocess the resting-state fMRI images;
[0016] A brain state extraction module is configured to extract brain states from preprocessed resting-state fMRI images through a deep neural network model and a state extraction algorithm to obtain basic state information of the brain functional network;
[0017] The task brain function network state acquisition module is configured based on the basic state information of the brain function network, and adds different task paradigm stimulation information through the task coding algorithm to obtain the brain function network state under a specific task;
[0018] A reconstruction module, which is configured to input the brain functional network state under a specific task into a deep neural network model and a signal reconstruction algorithm for regression reconstruction to obtain task-state fMRI;
[0019] The task-state brain activation map acquisition module is configured to obtain the task-state brain activation map from the regressed task-state fMRI prediction through a deep learning model and an activation map prediction algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 A schematic diagram showing the principle of the brain activation map prediction method according to the present invention is shown.
[0021] Figure 2 A flowchart of a brain activation map prediction method according to the present invention is shown. DETAILED DESCRIPTION
[0022] like Figure 1 , 2 As shown, a brain activation map prediction method comprises the following steps:
[0023] (1) Preprocessing of resting-state fMRI images;
[0024] (2) The preprocessed resting-state fMRI images are used to extract brain states through a deep neural network model and a state extraction algorithm to obtain basic state information of the brain functional network;
[0025] (3) Based on the basic state information of the brain functional network, different task paradigm stimulus information is added through the task coding algorithm to obtain the brain functional network state under specific tasks;
[0026] (4) The brain functional network state under a specific task is input into the deep neural network model and the signal reconstruction algorithm for regression reconstruction to obtain the task-state fMRI;
[0027] (5) Through the deep learning model and activation map prediction algorithm, the task-state brain activation map is predicted from the regressed task-state fMRI.
[0028] The present invention directly uses different task paradigms to guide the modification of the basic state of the brain functional network to obtain the brain functional network state under different tasks, thereby realizing the transition of the brain state from resting to a specific task, and then predicting the task-state fMRI and brain activation map through a deep learning model. The prediction of brain activation maps in different task states can be realized by modifying the task paradigm added to the basic state of the brain functional network. There is no need for the subjects to undergo multiple complex task-state fMRI scans, which reduces time and medical costs and improves the individualized analysis capabilities of brain imaging research.
[0029] Preferably, in step (1), the open source tool fMRIPrep is used to preprocess the resting-state fMRI images, including time correction, head motion correction, variable regression, filtering, spatial normalization, and spatial smoothing; time correction corrects the scanning time difference to ensure the synchronization of each layer of images; head motion correction is used to correct the signal error caused by the subject's head movement; variable regression is used to eliminate the influence of non-interesting signals; filtering is used to retain the frequency band signals of interest and remove non-physiological noise; spatial normalization aligns individual brain images to the standard space for group analysis; spatial smoothing uses Gaussian filtering to smooth the image to improve the signal-to-noise ratio and adapt to the differences in brain structure between individuals.
[0030] Preferably, in step (1), the preprocessed resting-state fMRI image is expanded voxel by voxel into two-dimensional data, one dimension being voxel and the other dimension being time.
[0031] Preferably, in step (4), the task-state fMRI image is preprocessed and expanded voxel by voxel into two-dimensional data, with one dimension being voxel and the other dimension being time.
[0032] Preferably, in the step (1), the task-state fMRI is used to calculate the brain activation map by the generalized linear model GLM method for supervision in the deep learning model training. The GLM method constructs a design matrix based on the experimental design, in which the task condition generates a predicted time series by convolving the hemodynamic response function (HRF) of the blood oxygen level dependent (BOLD) signal as an independent variable. By performing regression analysis on the time series signal of each voxel, the contribution (regression coefficient) of different task conditions to the BOLD signal is estimated. Subsequently, a statistical test map (such as a t-value map or a z-value map) is generated by linear comparison of the task conditions, and the false positive rate is controlled by multiple comparison correction. Finally, based on the statistical significance threshold, a brain activation map reflecting the distribution of brain area activation under specific task conditions is generated.
[0033] Preferably, in step (3), the task paradigm is based on a block design, which divides the task stimulus into a number of continuous time blocks, each block contains the same type of task, and the task blocks are presented alternately with the rest blocks.
[0034] Preferably, the state extraction algorithm and the signal reconstruction algorithm use a Transformer network, and the task encoding algorithm and the activation map prediction algorithm use a multi-layer perceptron MLP network.
[0035] Preferably, after the model training iterations until the sum of the signal reconstruction loss and the activation map prediction loss is reduced to a specified value, the trained model file is saved to obtain a model file for predicting a specific task brain activation map from resting fMRI. The signal reconstruction loss and the activation map prediction loss use the mean absolute error loss function (Mean Absolute Error). To predict a specific task brain activation map from resting fMRI, a deep learning model is constructed and trained. The model takes the resting fMRI time series as input and the task state activation map as output. During the training process, the mean absolute error (MAE) is used as the loss function to calculate the signal reconstruction loss and the activation map prediction loss respectively. The reconstructed task state signal and the actual task state signal are used to calculate the signal reconstruction loss; the predicted activation map and the actual activation map are used to calculate the activation map prediction loss. The total loss is the weighted sum of these two parts of the loss. The model is iteratively optimized and the parameters are updated to minimize the total loss. When the total loss is reduced to the set threshold, the training process stops and the model parameters are saved as a file for subsequent application.
[0036] Preferably, during the testing phase of the method, the resting-state fMRI and the task paradigm are input into the trained model to obtain the corresponding task-state brain activation map.
[0037] All methods in the present invention are applicable to voxel level and region of interest (ROI) level.
[0038] The task paradigm added to the model in the present invention includes but is not limited to task paradigm signals and text descriptions of the task paradigm.
[0039] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes the steps of the above-mentioned embodiment method, and the storage medium can be: ROM / RAM, disk, CD, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a brain activation map prediction device, which is usually represented in the form of functional modules corresponding to the steps of the method. The device includes:
[0040] a preprocessing module configured to preprocess the resting-state fMRI images;
[0041] A brain state extraction module is configured to extract brain states from preprocessed resting-state fMRI images through a deep neural network model and a state extraction algorithm to obtain basic state information of the brain functional network;
[0042] The task brain function network state acquisition module is configured based on the basic state information of the brain function network, and adds different task paradigm stimulation information through the task coding algorithm to obtain the brain function network state under a specific task;
[0043] A reconstruction module, which is configured to input the brain functional network state under a specific task into a deep neural network model and a signal reconstruction algorithm for regression reconstruction to obtain task-state fMRI;
[0044] The task-state brain activation map acquisition module is configured to obtain the task-state brain activation map from the regressed task-state fMRI prediction through a deep learning model and an activation map prediction algorithm.
[0045] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.
Claims
1. A brain activation map prediction method, characterized in that: It includes the following steps: (1) Preprocessing of resting-state fMRI images; (2) The preprocessed resting-state fMRI images are used to extract brain states through a deep neural network model and a state extraction algorithm to obtain basic state information of the brain functional network; (3) Based on the basic state information of the brain functional network, different task paradigm stimulus information is added through the task coding algorithm to obtain the brain functional network state under specific tasks; (4) The brain functional network state under a specific task is input into the deep neural network model and the signal reconstruction algorithm for regression reconstruction to obtain the task-state fMRI; (5) Through the deep learning model and activation map prediction algorithm, the task-state brain activation map is predicted from the regressed task-state fMRI.
2. The method for predicting brain activation maps according to claim 1, characterized in that: In the step (1), the open source tool fMRIPrep is used to preprocess the resting state fMRI images and the task state fMRI images, including time correction, head motion correction, variable regression, filtering, spatial normalization, and spatial smoothing; time correction corrects the scanning time difference to ensure the synchronization of each layer of images; head motion correction is used to correct the signal error caused by the subject's head movement; variable regression is used to eliminate the influence of non-interesting signals; filtering is used to retain the frequency band signals of interest and remove non-physiological noise; spatial normalization aligns individual brain images to the standard space for group analysis; spatial smoothing uses Gaussian filtering to smooth the image to improve the signal-to-noise ratio and adapt to the differences in brain structure between individuals.
3. The method for predicting brain activation maps according to claim 2, characterized in that: In the step (1), the preprocessed resting-state fMRI image is expanded voxel by voxel into two-dimensional data, one dimension is voxel and the other dimension is time.
4. The method for predicting brain activation maps according to claim 3, characterized in that: In the step (4), the task-state fMRI image is preprocessed and expanded voxel by voxel into two-dimensional data, one dimension being voxel and the other dimension being time.
5. The method for predicting brain activation maps according to claim 4, characterized in that: In the step (1), the task-state fMRI is used to calculate a brain activation map by a generalized linear model (GLM) method, which is used for supervision in deep learning model training.
6. The method for predicting brain activation maps according to claim 5, characterized in that: In step (3), the task paradigm is based on a block design, which divides the task stimulus into a number of continuous time blocks, each block contains the same type of task, and the task blocks are presented alternately with the rest blocks.
7. The method for predicting brain activation maps according to claim 6, characterized in that: The state extraction algorithm and signal reconstruction algorithm use the Transformer network, and the task encoding algorithm and activation map prediction algorithm use the multi-layer perceptron MLP network.
8. The method for predicting brain activation maps according to claim 7, characterized in that: After the model training iterations are completed until the sum of the signal reconstruction loss and the activation map prediction loss is reduced to a specified value, the trained model file is saved to obtain a model file for predicting task-specific brain activation maps from resting-state fMRI.
9. The method for predicting brain activation maps according to claim 8, characterized in that: During the testing phase, this method inputs the resting-state fMRI and task paradigm into the trained model to obtain the corresponding task-state brain activation map.
10. A brain activation map prediction device, characterized in that: It includes: a preprocessing module configured to preprocess the resting-state fMRI images; A brain state extraction module is configured to extract brain states from preprocessed resting-state fMRI images through a deep neural network model and a state extraction algorithm to obtain basic state information of the brain functional network; The task brain function network state acquisition module is configured based on the basic state information of the brain function network, and adds different task paradigm stimulation information through the task coding algorithm to obtain the brain function network state under a specific task; A reconstruction module, which is configured to input the brain functional network state under a specific task into a deep neural network model and a signal reconstruction algorithm for regression reconstruction to obtain task-state fMRI; The task-state brain activation map acquisition module is configured to obtain the task-state brain activation map from the regressed task-state fMRI prediction through a deep learning model and an activation map prediction algorithm.