Brain source imaging method and system based on learnable regularization
By designing a learnable regularized brain-source imaging method, and utilizing a network framework for residual connections and texture reconstruction, the threshold and regularization intensity are adaptively adjusted. This solves the problem of low reconstruction quality in existing brain-source imaging technologies, achieving higher reconstruction accuracy and stability, and is applicable to brain source localization and temporal reconstruction of EEG signals.
Patent Information
- Application Number
- CN202511888919.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-01-23
AI Technical Summary
Existing deep unfolding networks based on iterative shrinkage thresholding algorithms have low reconstruction quality in brain-source imaging, and their reliance on fixed thresholding strategies leads to poor regularization intensity adjustment.
A learnable regularization-based brain-source imaging method is designed. By acquiring training EEG observation signals and real brain-source signals, the model is trained using a learnable regularized unfolded network with residual connections, texture reconstruction, and structural reconstruction. The threshold and regularization intensity are adaptively adjusted, including multiple cascaded deep unfolded network layers. Feature processing is performed using convolutional blocks, convolutional residual blocks, texture and structural regularization blocks, and linear transformation layers.
It achieves higher reconstruction accuracy and stability, has adaptive regularization adjustment capability, can effectively suppress redundancy under low signal-to-noise ratio and few electrode conditions, retain key source activities, improves the localization performance and interpretability of brain source imaging, and reduces the number of model parameters.
Smart Images
Figure CN121370191A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of brain imaging technology, and more particularly to a brain imaging method and system based on learnable regularization. Background Technology
[0002] Brain source imaging (BSI) aims to reconstruct neural activity in the source space based on sensor-observed signals and lead matrices. Since the number of electrodes / sensors is usually much smaller than the scale of cortical source points, the inverse problem is highly underdetermined and sensitive to noise. Therefore, reasonable regularization / priors need to be introduced under the constraint of physical consistency to achieve stable, accurate and interpretable source activity reconstruction and localization.
[0003] With the development of deep learning, iterative solution methods such as the Iterative Shrinkage-Thresholding Algorithm (ISTA) have been deeply unfolded into neural network structures (DUNs). By mapping each iteration to a layer (or a stage) of the network, end-to-end training is used to adaptively learn the step size, threshold, and proximal operators in order to achieve stronger fitting ability and better generalization. However, existing DUNs still rely on fixed threshold strategies to adjust the regularization strength, resulting in low quality of brain-derived imaging reconstruction. Summary of the Invention
[0004] This invention provides a brain-source imaging method and system based on learnable regularization, which aims to improve the technical problem of low reconstruction quality in brain-source imaging when depth unfolding based on iterative shrinkage threshold algorithm is low.
[0005] The first aspect of this invention provides a brain-source imaging method based on learnable regularization, comprising:
[0006] Acquire training EEG observation signals and real brain-source signals;
[0007] After preprocessing the training EEG observation signals and obtaining the initial brain source signals through the inversion matrix, the signals are input into the trainable learnable regularized unfolded network based on residual connections, texture reconstruction, structure reconstruction, and so on for model training. The target learnable regularized unfolded network is determined by minimizing the loss function based on the real brain source signals.
[0008] After preprocessing and inversion of the collected measured EEG signals, the target brain source signal is output through the target learnable regularized expansion network.
[0009] Further, the learnable regularization unfolding network comprises a plurality of cascaded deep unfolding network layers, the deep unfolding network layers comprising convolution blocks, convolution residual blocks, texture and structure regularization blocks, and linear transformation layers, the texture and structure regularization blocks comprising cascaded structure reconstruction modules and texture reconstruction modules;
[0010] The processing procedure of the deep unfolding network layers comprises:
[0011] The input brain source signal input into the deep unfolding module is sequentially processed by the convolution blocks, the convolution residual blocks, the texture and structure regularization blocks, and the linear transformation layers, and then is added to the input brain source signal element by element to determine an output brain source signal.
[0012] Further, after the training electroencephalogram observation signal is preprocessed and an initial training brain source signal is obtained through an inversion matrix, the initial training brain source signal is input into a to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction, and structure reconstruction to perform model training, and a target learnable regularization unfolding network is determined by minimizing a loss function based on the real brain source signal, comprising:
[0013] The training electroencephalogram observation signal is band-pass filtered, power interference suppressed, baseline corrected, amplitude normalized, and time-sliced to determine a training preprocessed electroencephalogram observation signal;
[0014] The preprocessed training electroencephalogram observation signal is mapped to a cortical source space through an inversion matrix to determine an initial training brain source signal;
[0015] The initial training brain source signal is solved by the to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction, and structure reconstruction to obtain an output reconstructed brain source signal;
[0016] The loss function is calculated using the reconstructed brain source signal and the real brain source signal;
[0017] When the loss function does not satisfy an iteration stop condition, the model parameters are updated to minimize the loss function until the loss function satisfies the iteration stop condition, and then a target learnable regularization unfolding network is determined.
[0018] Further, the calculation procedure of the loss function comprises:
[0019]
[0020] In the formula, L represents the loss function, Y represents the reconstructed brain source signal, X represents the real brain source signal, ||Y||2 represents the square of the L2 norm, ||Y||1 represents the L1 norm, and denotes a sparse regularization term coefficient, denotes a spatial gradient of the reconstructed brain source signal, denotes a spatial smoothing term coefficient.
[0021] The second aspect of the present application provides a brain source imaging system based on learnable regularization, comprising:
[0022] A data acquisition module is configured to acquire training electroencephalogram observation signals and real brain source signals.
[0023] A model training module is configured to input the training electroencephalogram observation signals after preprocessing, and obtain training initial brain source signals through an inversion matrix, into a to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction structure reconstruction, and model training, and determine a target learnable regularization unfolding network by minimizing a loss function based on the real brain source signals.
[0024] A brain source imaging module is configured to output a target brain source signal through the target learnable regularization unfolding network after preprocessing and inversion of the collected measured electroencephalogram observation signals.
[0025] Further, the learnable regularization unfolding network comprises a plurality of cascaded deep unfolding network layers, the deep unfolding network layers comprise convolution blocks, convolution residual blocks, texture and structure regularization blocks, and linear transformation layers, and the texture and structure regularization blocks comprise cascaded structure reconstruction modules and texture reconstruction modules.
[0026] The processing procedure of the deep unfolding network layer comprises:
[0027] The input brain source signal input into the deep unfolding module is sequentially subjected to feature processing through the convolution blocks, the convolution residual blocks, the texture and structure regularization blocks, and the linear transformation layers, and then is added element by element with the input brain source signal to determine an output brain source signal.
[0028] Further, the model training module is specifically configured to:
[0029] The training electroencephalogram observation signals are subjected to band-pass filtering, power interference suppression, baseline correction, amplitude normalization, and time slicing to determine training preprocessed electroencephalogram observation signals.
[0030] The preprocessed training electroencephalogram observation signals are mapped to a cortical source space through an inversion matrix to determine training initial brain source signals.
[0031] The training initial brain source signals are solved by the to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction, and structure reconstruction to output reconstructed brain source signals.
[0032] The loss function is calculated using the reconstructed brain source signals and the real brain source signals.
[0033] When the loss function does not satisfy the iteration stop condition, the model parameters are updated to minimize the loss function until the loss function satisfies the iteration stop condition, and then the target learnable regularization unfolding network is determined.
[0034] The third aspect of the present application provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the brain source imaging method based on the learnable regularization.
[0035] The fourth aspect of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed to implement the brain source imaging method based on the learnable regularization.
[0036] The fifth aspect of the present application provides a computer program product, comprising computer programs / instructions, which are executed by a processor to implement the brain source imaging method based on the learnable regularization.
[0037] From the above technical solutions, the present application has the following advantages:
[0038] The above scheme of the present application provides a brain source imaging method based on learnable regularization, comprising: obtaining training electroencephalogram observation signals and real brain source signals; after preprocessing the training electroencephalogram observation signals and obtaining the training initial brain source signals through the inversion matrix, inputting the to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction structure reconstruction into the model training, determining the target learnable regularization unfolding network by minimizing the loss function based on the real brain source signals; after preprocessing and inversion of the collected real electroencephalogram observation signals, outputting the target brain source signal through the target learnable regularization unfolding network. Based on the above scheme, through the designed brain source imaging unfolding network framework with adaptive regularization operator selection capability, through the inversion matrix, the structure-texture decomposition capability, the threshold learnable regularization adjustment mechanism and the residual steady-state optimization characteristics are possessed while the physical constraint of "data consistency" is retained, the threshold and the regularization strength are adaptively adjusted from end to end according to the data characteristics, so that higher reconstruction accuracy is realized. BRIEF DESCRIPTION OF DRAWINGS
[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.
[0040] Figure 1 A step flow chart of a brain source imaging method based on learnable regularization provided for the first embodiment of the present application is shown in the figure.
[0041] Figure 2 A structural schematic diagram of the reconstruction network of the ISTA-Net in the kth stage provided for the first embodiment of the present application is shown in the figure.
[0042] Figure 3 A structural schematic diagram of the deep unfolding network layer of the learnable regularization unfolding network in the kth stage provided for the first embodiment of the present application is shown in the figure.
[0043] Figure 4 A structural schematic diagram of the structure reconstruction module provided for the first embodiment of the present application is shown in the figure.
[0044] Figure 5 A structural schematic diagram of the texture reconstruction module provided for the first embodiment of the present application is shown in the figure.
[0045] Figure 6 A schematic diagram of the brain source imaging processing process provided for the first embodiment of the present application is shown in the figure.
[0046] Figure 7 A comparison schematic diagram of the parameter quantity under different methods provided for the first embodiment of the present application is shown in the figure.
[0047] Figure 8 A comparison schematic diagram of the AUC of different methods provided for the first embodiment of the present application is shown in the figure.
[0048] Figure 9 A structural block diagram of a brain source imaging system based on learnable regularization provided for the second embodiment of the present application is shown in the figure. DETAILED DESCRIPTION
[0049] The embodiments of the present application provide a brain source imaging method and system based on learnable regularization, which are used to solve the technical problem of low reconstruction quality of deep unfolding based on an iterative shrinkage threshold algorithm in brain source imaging.
[0050] In order to make the purpose, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the embodiments described below are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0051] Please refer to Figure 1 The brain source imaging method based on learnable regularization provided by the first embodiment of the present application comprises the following steps.
[0052] Step 101, obtaining training electroencephalogram observation signals and true brain source signals.
[0053] It should be noted that the brain source signal is used to reflect the distribution of the bioelectric signal generated by the neuron activity of the cerebral cortex, and the electroencephalogram observation signal refers to the electrical signal collected by the electrode or sensor placed on the surface of the cerebral cortex. The training set is constructed by using the training electroencephalogram observation signal and the true brain source signal, so as to facilitate subsequent network training.
[0054] In an implementation manner, one or more active regions can be set on the cortical grid based on a known lead-field, and then the corresponding brain source distribution, i.e., the true brain source signal, is generated The training electroencephalogram observation signal is obtained by the lead-field and superimposing noise conforming to the actual working condition, that is , is the electroencephalogram observation signal, is the lead-field, is the brain source signal, is the noise and modeling error, and a training sample pair of different signal-to-noise ratios can be formed .
[0055] Step 102, after the training electroencephalogram observation signal is preprocessed and the training initial brain source signal is obtained through the inversion matrix, the target learnable regularization unfolding network is determined by inputting the training initial brain source signal into the to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction structure reconstruction, and minimizing the loss function based on the true brain source signal.
[0056] In one specific implementation manner of the embodiment, step 102 includes the following sub-steps:
[0057] The training electroencephalogram observation signal is band-pass filtered, power interference suppressed, baseline corrected, amplitude normalized, and time-sliced to determine the training preprocessed electroencephalogram observation signal;
[0058] The preprocessed training electroencephalogram observation signal is mapped to the cortical source space through the inversion matrix to determine the training initial brain source signal;
[0059] The training initial brain source signal is solved and output by the to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction, and structure reconstruction to obtain a reconstructed brain source signal;
[0060] The loss function is calculated by using the reconstructed brain source signal and the true brain source signal;
[0061] When the loss function does not satisfy the iteration stop condition, the model parameters are updated to minimize the loss function until the loss function satisfies the iteration stop condition, and then the target learnable regularization unfolding network is determined.
[0062] In one specific implementation of the embodiment, the learnable regularized unfolding network comprises a plurality of cascaded deep unfolding network layers, the deep unfolding network layer comprises a convolution block, a convolution residual block, a texture and structure regularization block, and a linear transformation layer, the texture and structure regularization block comprises a cascaded structure reconstruction module and a texture reconstruction module;
[0063] The processing procedure of the deep unfolding network layer comprises:
[0064] The input brain source signal input into the input deep unfolding module is sequentially subjected to feature processing by the convolution block, the convolution residual block, the texture and structure regularization block, and the linear transformation layer, and then is subjected to element-by-element addition with the input brain source signal to determine an output brain source signal.
[0065] It should be noted that the preprocessing can comprise band-pass filtering, power interference suppression, baseline correction, amplitude normalization, and time slicing (which can comprise uniform time window slicing or single time point slicing), the training electroencephalogram observation signal is preprocessed to form a network input tensor, so as to train the preprocessed electroencephalogram observation signal, and the inversion matrix The training preprocessed electroencephalogram observation signal is mapped to a cortical source space to obtain a training initial brain source signal to provide a "data consistency" physical constraint, for example, the inversion matrix can be constructed by using an algorithm based on MNE, and specific details can be referred to in the prior art;
[0066] The embodiment designs a learnable regularized unfolding network on the basis of the ISTA-Net framework, performs steady-state optimization by introducing a residual connection to ensure continuity of features between stages and numerical stability, explicitly distinguishes "structural redundant information" and "texture redundant information" based on texture reconstruction and structure reconstruction to separate and accurately process subtle redundant information, adjusts thresholds and regularization strengths end-to-end according to data characteristics, and improves information density by reducing redundancy (rather than simply increasing capacity), and the method of improving the ISTA-Net is to "stack deeper / broader", which is easy to cause parameter explosion; It can be understood that the "structural redundant information" is related to spatial continuity and functional block boundaries on the cortical grid, for example, redundant components of repeated edges, outline blocks, and smooth areas, and the "texture redundant information" more reflects fine noise and high-frequency disturbance, for example, fine high-frequency noise and repeated textures;
[0067] It can be understood that the ISTA-Net framework can be referred to as Figure 2 as shown, and the specific steps comprise: 1, a data consistency step, which corresponds to the linear gradient descent process of ISTA: , denotes the intermediate feature vector after gradient descent of the stage, denotes the intermediate feature vector after gradient descent of the brain-derived signals output in the stage, denotes the learnable step size in the stage, denotes the lead matrix, denotes the transpose of the lead matrix, denotes the lead matrix; the physical meaning of this step is to correct the reconstruction error of the previous stage while maintaining the consistency of observation. Unlike traditional algorithms, ISTA-Net adjusts the update amplitude of each stage adaptively through learning instead of artificial setting, which improves the convergence efficiency and stability; 2, the proximal mapping step, which corresponds to the sparse regularization solution of ISTA. ISTA-Net uses the "analysis-threshold-synthesis (Analysis-Threshold-Synthesis)" framework to approximate the nonlinear sparse transformation with a convolution operator: denotes the thresholded sparse coefficient output by the analysis transformation, denotes the intermediate feature vector after gradient descent in the stage, denotes the learnable analysis transformation (analysis convolution layer), denotes the soft threshold operator, denotes the threshold (learned adaptively by the network), denotes the thresholded sparse feature vector, denotes the synthesis transformation (decoding convolution layer), denotes the brain-derived signals output in the stage; the existing scheme introduces symmetry constraint (Symmetry Constraint) in training to ensure , thereby improving the reversibility and stability; 3, ISTA-Net usually adopts the "untied parameters" strategy, so that each stage can learn the optimal step size and threshold for different reconstruction depths; the entire network minimizes the reconstruction error in an end-to-end manner, achieving interpretable depth unfolding;
[0068] The learnable regularization unfolding network designed based on the ISTA-Net framework in this embodiment includes multiple cascaded depth unfolding network layers, which map each iteration of ISTA to a stage of depth unfolding network layer, and the first The network structure of the deep unfolding network layer of each stage includes a convolution block (CB), a convolution residual block (RB), a texture and structure regularization block, and a linear transformation layer; wherein the convolution block can include 3x3 convolution+ReLU activation function, the convolution residual block can include 3x3 convolution with residual structure, ReLU activation function and 3x3 convolution, that is, the input of the convolution residual block is connected in residual with the output of the second 3x3 convolution, the texture and structure regularization block includes cascaded structure reconstruction module and texture reconstruction module, and the linear transformation layer can adopt 1x1 convolution; for the input brain source signal of the input deep unfolding module of each stage, after being sequentially processed by the convolution block, the convolution residual block, the texture and structure regularization block and the linear transformation layer, the output brain source signal is determined by element-by-element addition with the input brain source signal as the input of the next stage.
[0069] After the above-mentioned learnable regularization unfolding network is built, the initial learning rate is set to 0.001, the training batch size is 32, the training period is 200 epochs, the step, the threshold value and the linear transformation and weight in the deep unfolding network layer are set as learnable parameters, the initial brain source signal is input into the trained learnable regularization unfolding network, the multi-scale texture features are extracted and enhanced through the convolution structure of the convolution block and the convolution residual block, the texture details and the spatial structure information are simultaneously depicted on the output of the convolution residual block through the texture and structure regularization block, the brain source distribution is gradually refined and constrained, the output brain source signal of the final stage of the network is taken as the reconstructed brain source signal, the loss function is calculated based on the reconstructed brain source signal and the real brain source signal, the small batch random gradient descent and the Adam optimization algorithm are used based on the loss function for iterative training until the loss function meets the iteration stop condition to determine the target learnable regularization unfolding network; it can be understood that the brain source label is arranged in the order of the cortical grid and normalized, which corresponds to the lead matrix and the network output one by one, and the iteration stop condition can be set as that the number of training times meets the number threshold or the loss converges, which is not limited in the embodiment.
[0070] Exemplarily, in a specific implementation, the network structure of the structure reconstruction module (Structure Reconstruction block) is as shown in Figure 4 , which retains key information and compresses repeated structures through "separation-reconstruction" operation:
[0071] 1. Feature normalization and weight extraction: the structure input feature is processed by group normalization (GN) to calculate the mean and variance of each channel, so as to obtain the normalized feature:
[0072]
[0073] wherein, denotes the normalized feature, denotes the group normalization layer, denotes the learnable scaling parameter, denotes the learnable bias parameter, denotes a very small constant;
[0074] Subsequently, the weight distribution is extracted from the output of the GN layer to reflect the spatial saliency of different channels, which is achieved by channel normalization:
[0075]
[0076] wherein, denotes the weight of the th channel, denotes the channel index, denotes the number of channels;
[0077] 2. Information channel separation (Separate Operation): According to the weight distribution , the adaptive separation of information channels is achieved through the Sigmoid activation and gate function: When the gate value is higher than the threshold T, the corresponding channel is determined as the information salient channel, forming the weight ; those lower than the threshold are non-salient channels, forming the weight ;
[0078] 3. Channel weighting and feature division: the input feature is multiplied by the weights , respectively, obtaining two types of weighted features: , , contains rich structure and spatial information, represents low information or redundant features. This operation realizes the adaptive distinction between salient and non-salient information;
[0079] 4. Feature reconstruction (Reconstruction Operation): To reduce structural redundancy and enhance structural consistency, a cross-reconstruction mechanism is introduced. The information feature and the redundant feature are cross-fused according to the spatial patch, and the four groups of reconstructed features are multiplied and summed by channel to form the structure-optimized feature mapping: , , , , ;
[0080] 5. Structure optimization output: structure reconstruction feature , which can maintain global feature consistency and gradient stability and provide purer input for the subsequent texture reconstruction module.
[0081] Exemplarily, in a specific implementation, the network structure of the texture reconstruction module (Texture Reconstruction block) is as shown in Figure 5 , through channel segmentation, light convolution and attention fusion, to filter out truly useful texture details:
[0082] 1. Channel segmentation: divide the texture input feature into two groups of sub-features in the channel direction: one group is the first channel (denoted as ), and the other group is the remaining channel (denoted as ), wherein is a settable constant or a learnable parameter of the segmentation ratio. The purpose of this segmentation operation is to separate the texture channels with high information quantity and strong expression from the channels with relatively low information quantity, which may contain redundancy or noise, so as to facilitate subsequent differentiated compression. After segmentation, 1x1 convolution is used for channel compression;
[0083] 2. Local texture compression and channel transformation: 1x1 point-wise convolution (PWC) is used to compress the two sub-feature branches obtained by segmentation and compression, to reduce the channel dimension of each branch and suppress the redundant expression of repeated textures. At the same time, in order to balance the correlation modeling across channels and improve the calculation efficiency, group-wise convolution (GWC) is further introduced in one of the branches to perform convolution operation on the features by groups, thereby reducing the parameter quantity and limiting unnecessary cross-group coupling. Thus, two groups of intermediate features compressed and transformed by texture can be obtained, denoted as: , , represents the upper texture sub-feature, represents the convolution mapping based on GWC, represents the convolution mapping based on PWC, represents the lower texture sub-feature; it can be understood that the process of obtaining the upper rich feature embodies the use of group convolution (emphasizing group structure and sparsity) and point-wise convolution (emphasizing full-channel reorganization) within the same branch to construct representative texture sub-features.
[0084] 3. To achieve adaptive reservation of useful texture components, global statistical pooling (such as global average pooling or global weighted pooling) is applied on each branch to obtain global description vectors of each branch and , then these global description vectors are mapped by Softmax normalization to generate two attention weights and , representing the contribution degree of high-information texture branches, representing the supplementary value of low-information texture branches;
[0085] 4. The texture sub-features of the corresponding branches are weighted and fused element by element using the above attention weights to obtain texture reconstruction features .
[0086] In one specific embodiment of the present embodiment, the loss function includes a reconstruction loss (MSE), an L1 sparse regularization term and a spatial smoothness (gradient consistency) constraint, and the calculation process includes:
[0087]
[0088] In the formula, represents the loss function, represents the reconstructed brain source signal, represents the true brain source signal, represents the square of the L2 norm, represents the L1 norm, represents the sparse regularization term coefficient, represents the spatial gradient of the reconstructed brain source signal, represents the spatial smoothness term coefficient; wherein the MSE reconstruction loss is used to ensure that the output signal approximates the true brain source signal in amplitude, the L1 sparse regularization term is used to emphasize the inherent sparsity of the neural source signal to improve the distinguishability of the source position, and the gradient consistency constraint is used to suppress abnormal spikes and maintain the spatial continuity of the signal on the cortical grid; specifically, The optimal value in the range of 0.001-0.01 can be selected by experiment.
[0089] Step 103, after pre-processing and inversion of the collected measured electroencephalogram observation signal, output the target brain source signal through the target learnable regularization unfolding network.
[0090] It should be noted that, as shown in Figure 6 , after pre-processing and inversion of the collected measured electroencephalogram observation signal (such as an EEG signal), the target brain source signal is input into the target learnable regularization unfolding network and output through forward propagation.
[0091] To illustrate the effect of the present embodiment, experimental verification is carried out:
[0092] Reference is made to Figure 7 The present embodiment is compared with the existing ISTA-Net (Interpretable Optimization-Inspired Deep Network for Image Compressive Sensing), AMP-Net (Denoising-Based Deep Unfolding for Compressive Image Sensing), DPA-Net (Dual-Path Attention Network for Compressed Sensing Image Reconstruction), and TRanCS (A Transformer-Based Hybrid Architecture for Image Compressed Sensing), and the peak signal-to-noise ratio (PSNR) is used as an objective indicator to evaluate the quality of signal reconstruction. The indicator reflects the ability of the reconstruction model to restore signal details by measuring the error between the reconstruction result and the reference true value. From Figure 7 It can be seen that when the number of parameters of the method of the present embodiment is about 1M (about 0.8M-1.2M), the PSNR has reached or exceeded the level that other networks can reach when a larger scale (2M, 3M, 4M or even more) is required. This shows that under almost the same computing / storage overhead, the present embodiment can produce a higher PSNR, i.e. a higher quality reconstruction result, and can more accurately restore the neural source signal and more effectively suppress noise interference.
[0093] Reference is made to Figure 8As shown, AUC (Area Under the Receiver Operating Characteristic Curve) represents the overall discrimination ability of the model for true source locations and non-source locations at each decision threshold, and the closer the value is to 1, the better the recognition performance. The AUC can comprehensively evaluate the source point detection ability and false alarm suppression effect of the method under different noise intensities, thereby verifying the reliability and robustness of the method. In comparison with VSSI-Net (a variant network of VSSI-Lp), LORETA (Low resolution electromagnetic tomography: a new method for localizing electrical activity in the brain), SBL (A unified Bayesian framework for MEG / EEG source imaging) and VSSI-Lp (EEG Extended Source Imaging with Variation Sparsity and Lp-Norm Constraint), the AUC of the method of the embodiment remains optimal throughout the training phase and steadily rises with the epoch: starting from about 0.75 at the early epoch = 1, which is significantly higher than VSSI-Net (≈0.60), VSSI-Lp (≈0.70), LORETA (≈0.50) and SBL (≈0.35); at Epoch = 20, the AUC of the method is about 0.90, which is about +0.05 higher than VSSI-Net and about +0.02-0.03 higher than VSSI-Lp, and is about +0.35 and +0.55 higher than LORETA and SBL, respectively; at the convergence stage of Epoch = 100, the AUC of the method is about 0.95, which is still ahead of VSSI-Net (≈0.90) and VSSI-Lp (≈0.92), and is significantly better than LORETA (≈0.60) and SBL (≈0.40) based on traditional priors.
[0094] As can be seen from the above, the method of the embodiment can effectively suppress redundancy and retain key source activity under low signal-to-noise ratio and few electrode conditions, thereby achieving higher source detection accuracy (AUC) and faster early convergence, even when the number of training rounds is small, and maintaining a stable advantage in the later stage.
[0095] In the embodiment of the present application, based on the designed brain source imaging unfolding network framework with adaptive regularization operator selection capability, while retaining the "data consistency" physical constraint, the framework has structure-texture decomposition capability, threshold learnable regularization adjustment mechanism and residual steady-state optimization characteristics, so that the threshold and regularization strength are adaptively adjusted end-to-end according to the data characteristics, thereby realizing higher reconstruction accuracy, more robust positioning performance and better interpretability, while significantly reducing the number of model parameters without sacrificing performance, suitable for brain source localization and timing reconstruction of EEG and other signals, and can be extended to different head models and task paradigms.
[0096] Please refer to Figure 9 The second embodiment of the present application provides a brain source imaging system based on learnable regularization, which comprises:
[0097] The data acquisition module 901 is configured to acquire training electroencephalogram observation signals and real brain source signals.
[0098] The model training module 902 is configured to preprocess the training electroencephalogram observation signals, obtain training initial brain source signals through an inversion matrix, and input the training initial brain source signals into a to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction and structure reconstruction for model training, and determine a target learnable regularization unfolding network by minimizing a loss function based on the real brain source signals.
[0099] The brain source imaging module 903 is configured to preprocess and invert the collected measured electroencephalogram observation signals, and output target brain source signals through the target learnable regularization unfolding network.
[0100] In one specific embodiment of the present application, the model training module 902 is specifically configured to:
[0101] The training electroencephalogram observation signals are bandpass filtered, power interference suppressed, baseline corrected, amplitude normalized and time sliced to determine training preprocessed electroencephalogram observation signals.
[0102] The preprocessed training electroencephalogram observation signals are mapped to the cortical source space through an inversion matrix to determine training initial brain source signals.
[0103] The training initial brain source signals are solved and output to obtain reconstructed brain source signals through the to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction and structure reconstruction.
[0104] The loss function is calculated using the reconstructed brain source signals and the real brain source signals.
[0105] When the loss function does not satisfy the iteration stopping condition, the model parameters are updated to minimize the loss function until the loss function satisfies the iteration stopping condition, and then the target learnable regularization unfolding network is determined.
[0106] In one specific embodiment of the present embodiment, the calculation process of the loss function comprises:
[0107]
[0108] In the formula, denotes the loss function, denotes the reconstructed brain source signal, denotes the true brain source signal, denotes the square of the L2 norm, denotes the L1 norm, denotes the sparse regularization term coefficient, denotes the spatial gradient of the reconstructed brain source signal, denotes the spatial smoothing term coefficient.
[0109] In one specific embodiment of the present embodiment, the learnable regularization unfolding network comprises a plurality of cascaded deep unfolding network layers, the deep unfolding network layer comprises a convolution block, a convolution residual block, a texture and structure regularization block and a linear transformation layer, the texture and structure regularization block comprises a cascaded structure reconstruction module and a texture reconstruction module;
[0110] The processing process of the deep unfolding network layer comprises:
[0111] The input brain source signal input into the input deep unfolding module is sequentially subjected to feature processing by the convolution block, the convolution residual block, the texture and structure regularization block and the linear transformation layer, and then is added to the input brain source signal element by element to determine the output brain source signal.
[0112] Embodiment three of the present application further provides a computer device comprising a memory and a processor, the memory storing a computer program; the computer program is executed by the processor to make the processor execute the steps of the brain source imaging method based on learnable regularization as described in the above embodiment one of the present application.
[0113] Embodiment four of the present application further provides a computer readable storage medium, which stores a computer program / instruction, the computer program / instruction is executed by the processor to realize the steps of the brain source imaging method based on learnable regularization as described in the above embodiment one of the present application.
[0114] Embodiment five of the present application further provides a computer program product comprising a computer program / instruction, the computer program / instruction is executed by the processor to realize the steps of the brain source imaging method based on learnable regularization as described in the above embodiment one of the present application.
[0115] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system and module can refer to the corresponding process in the foregoing method embodiments, which will not be described here.
[0116] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other manners. For example, the system embodiments described above are merely illustrative. For example, the division of the modules is merely logical function division. In actual implementation, another division manner can be used. For example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between modules can be indirect coupling or communication connection through some interface, device or module, and can be electrical, mechanical or in other forms.
[0117] The modules illustrated as separated components can or can not be physically separated, and the components illustrated as modules can or can not be physical modules, i.e., can be located in one place or distributed to a plurality of network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.
[0118] In addition, each functional module in each embodiment of the present application can be integrated into a processing module, or each module can exist physically independently, or two or more modules can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of a software functional module.
[0119] The integrated module, if realized in the form of a software functional module and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application, essentially or the part that contributes to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.
[0120] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalent replacements; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A method of learning-based regularization-based source imaging of the brain, characterized in that, The method comprises the following steps: Obtaining training electroencephalogram observation signals and real brain source signals; After preprocessing the training electroencephalogram observation signals and obtaining training initial brain source signals through an inversion matrix, inputting the training electroencephalogram observation signals into a to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction and structure reconstruction for model training, and determining a target learnable regularization unfolding network by minimizing a loss function based on the real brain source signals. After preprocessing and inversion of the collected real electroencephalogram observation signals, outputting a target brain source signal through the target learnable regularization unfolding network.
2. The learning-based regularization method for source imaging of brain according to claim 1, wherein, The learnable regularization unfolding network comprises a plurality of cascaded deep unfolding network layers, the deep unfolding network layers comprise a convolution block, a convolution residual block, a texture and structure regularization block and a linear transformation layer, and the texture and structure regularization block comprises a cascaded structure reconstruction module and a texture reconstruction module. The processing process of the deep unfolding network layer comprises: After sequentially processing the input brain source signal input into the deep unfolding module through the convolution block, the convolution residual block, the texture and structure regularization block and the linear transformation layer, performing element-by-element addition on the input brain source signal and the output brain source signal, and determining the output brain source signal.
3. The learnable regularization based source imaging method of claim 1, wherein, The method of inputting the training electroencephalogram observation signals into the to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction and structure reconstruction for model training, and determining the target learnable regularization unfolding network by minimizing the loss function based on the real brain source signals comprises the following steps: Band-pass filtering, power interference suppression, baseline correction, amplitude normalization and time slicing are performed on the training electroencephalogram observation signals to determine training preprocessed electroencephalogram observation signals; The preprocessed training electroencephalogram observation signals are mapped to a cortical source space through an inversion matrix to determine training initial brain source signals; The training initial brain source signals are solved through the to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction and structure reconstruction to output reconstructed brain source signals; The loss function is calculated by using the reconstructed brain source signals and the real brain source signals. When the loss function does not satisfy the iteration stopping condition, the model parameters are updated to minimize the loss function until the loss function satisfies the iteration stopping condition, and then the target learnable regularization unfolding network is determined.
4. The learning-based regularization method for source imaging according to claim 1 or 3, wherein, The calculation process of the loss function comprises the following steps: wherein denotes a loss function, denotes a reconstructed brain source signal, denotes a true brain source signal, denotes a square of an L2 norm, denotes an LI norm, denotes a sparse regularization term coefficient, denotes a spatial gradient of the reconstructed brain source signal, denotes a spatial smoothing term coefficient.
5. A learning-regularization-based source imaging system, characterized by, The method comprises the following steps: A data acquisition module is configured to obtain training electroencephalogram observation signals and real brain source signals; A model training module is configured to input the training electroencephalogram observation signals into a to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction and structure reconstruction for model training after preprocessing the training electroencephalogram observation signals and obtaining training initial brain source signals through an inversion matrix, and determine a target learnable regularization unfolding network by minimizing a loss function based on the real brain source signals. A brain source imaging module is configured to output a target brain source signal through the target learnable regularization unfolding network after preprocessing and inversion of the collected real electroencephalogram observation signals.
6. The learnable regularization based brain source imaging system of claim 5, wherein, The learnable regularization unfolding network comprises a plurality of cascaded deep unfolding network layers, the deep unfolding network layers comprising convolution blocks, convolution residual blocks, texture and structure regularization blocks, and linear transformation layers, the texture and structure regularization blocks comprising cascaded structure reconstruction modules and texture reconstruction modules; The processing procedure of the deep unfolding network layers comprises: The input brain source signal is sequentially processed by the convolution blocks, the convolution residual blocks, the texture and structure regularization blocks, and the linear transformation layers, and then is added to the input brain source signal element by element to determine an output brain source signal.
7. The learnable regularization based source imaging system of claim 5, wherein, The model training module is specifically configured to: The training electroencephalogram observation signal is band-pass filtered, power interference suppressed, baseline corrected, amplitude normalized, and time-sliced to determine a training preprocessed electroencephalogram observation signal; The preprocessed training electroencephalogram observation signal is mapped to a cortical source space by an inversion matrix to determine a training initial brain source signal; The training initial brain source signal is solved by a to-be-trained learnable regularization unfolding network based on residual connection, texture reconstruction, and structure reconstruction to determine an output reconstructed brain source signal; The reconstructed brain source signal and the real brain source signal are used to calculate a loss function; When the loss function does not satisfy an iterative stop condition, the model parameters are updated to minimize the loss function until the loss function satisfies the iterative stop condition, and then a target learnable regularization unfolding network is determined.
8. A computer device, comprising: The computer program / instructions are executed by the processor to implement the steps of the brain source imaging method based on learnable regularization according to any one of claims 1-4.
9. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the brain source imaging method based on learnable regularization according to any one of claims 1-4.
10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the brain source imaging method based on learnable regularization according to any one of claims 1-4.