Artificial intelligence positioning system and method for epileptic focus based on magnetic resonance and electroencephalogram

Abstract

The invention relates to the field of biomedicine, and particularly discloses an epileptic focus artificial intelligence positioning system and method based on magnetic resonance and electroencephalography, and the system comprises the following contents: a data collection module is used for synchronously collecting T1 weighted magnetic resonance images and electroencephalogram signals of scalp; the data preprocessing module is used for performing brain tissue segmentation on the magnetic resonance image, obtaining structural features of each brain region, constructing vectors including whole brain structural features, and obtaining magnetic resonance structural feature vectors; artifacts of the electroencephalogram signals are removed, electroencephalogram features of all brain areas are extracted through time-frequency analysis, a matrix containing whole electroencephalogram physiological features is constructed, and the electrophysiological features are obtained; the cross-modal confidence coefficient dynamic evaluation module is used for establishing a bidirectional constraint rule to perform confidence coefficient calibration on the magnetic resonance structure feature vector and the electrophysiological feature; a positioning model construction and training module; a positioning result output module; according to the technical scheme, noise can be reduced during magnetic resonance and electroencephalogram fusion, and the epileptic focus positioning precision is high.

Description

Technical Field

[0001] The present invention relates to the field of biomedicine, and in particular to an artificial intelligence localization system and method for epileptic foci based on magnetic resonance imaging and electroencephalography. Background Art

[0002] Precise localization of epileptic foci is a core challenge in the surgical treatment of intractable epilepsy. Among traditional localization methods, magnetic resonance imaging (MRI) can reveal structural lesions (such as cortical dysplasia and hippocampal sclerosis) but lacks sensitivity for functional epileptogenic foci (regions of abnormal neuronal discharge without clear structural abnormalities). While scalp electroencephalography (EEG) can capture electrophysiological abnormalities (such as spikes and high-frequency oscillations (HFOs)), its spatial resolution is limited by the attenuation of skull electrical impedance, resulting in localization accuracy at only centimeters, making it difficult to distinguish epileptogenic foci from adjacent brain regions.

[0003] Currently, many studies have attempted to improve localization through multimodal fusion. For example, these approaches simply superimpose MRI structural features and EEG signals and input them into traditional machine learning models (such as support vector machines) or employ basic neural networks for feature concatenation. However, these approaches have two major drawbacks: First, the spatial matching between the range of structural abnormalities on MRI (such as areas of reduced gray matter density) and electrophysiological abnormalities on EEG (such as brain regions with frequent spikes) may be affected by individual differences, making direct fusion prone to the introduction of noise. Second, deep learning models (such as traditional convolutional neural networks) rely on fixed architectures and lack specialized networks designed to address the "structural-electrophysiological coupling characteristics" of epileptic foci, resulting in insufficient localization of deep nuclei (such as the amygdala and insula).

[0004] Therefore, there is an urgent need for an artificial intelligence localization system and method for epileptic foci based on magnetic resonance imaging and electroencephalography. Summary of the Invention

[0005] The present invention provides an artificial intelligence localization system for epileptic foci based on magnetic resonance imaging and electroencephalography, which can reduce noise when magnetic resonance imaging and electroencephalography are integrated, and has high accuracy in localizing epileptic foci.

[0006] In order to solve the above technical problems, this application provides the following technical solutions:

[0007] The artificial intelligence localization system for epileptic foci based on magnetic resonance imaging and electroencephalography includes the following:

[0008] The data acquisition module is used to synchronously acquire T1-weighted magnetic resonance images and scalp EEG signals;

[0009] The data preprocessing module is used for brain tissue segmentation of the magnetic resonance image, obtaining structural features of each brain region, constructing a vector containing the structural features of the whole brain, and obtaining a magnetic resonance structural feature vector; artifact removal is performed on the electroencephalogram signal, and electroencephalogram features of each brain region are extracted through time-frequency analysis, a matrix containing whole electro-physiological features is constructed, and electro-physiological features are obtained;

[0010] The cross-modal confidence dynamic evaluation module is used to establish a bidirectional constraint rule to calibrate the confidence of the magnetic resonance structural feature vector and the electro-physiological feature;

[0011] The positioning model construction and training module is used to encode the brain region structural connection relationship through a graph convolution network, capture the brain region structural features, model the time dependence of the electroencephalogram signal through a time sequence convolution network, extract the electro-physiological time sequence features, calculate the interaction weight of the structural features and the electro-physiological features, dynamically allocate the fusion proportion of the two based on the weight, realize the deep fusion of the structural-electro-physiological coupling features, input the fused features into a full connection layer, and output the three-dimensional coordinates of the epileptogenic focus and the positioning confidence, wherein the positioning confidence is calculated through a model uncertainty evaluation mechanism; during training, the coordinates of the epileptogenic focus recorded by the intracranial electrode are used as a supervision signal, and mean square error loss and cross entropy loss are used for joint optimization, while the positioning coordinate accuracy and the confidence prediction accuracy are constrained;

[0012] The positioning result output module is used to filter out brain regions with confidence lower than a preset threshold according to the coordinates and confidence of the epileptogenic focus output by the model, and finally determine the three-dimensional spatial position of the epileptogenic focus and the corresponding confidence score.

[0013] The basic scheme principle and beneficial effects are as follows: the core principle of the present application lies in the complementation and synergistic optimization of multi-modal data, the magnetic resonance (MRI) image can clearly present the anatomical structure of the brain, has high sensitivity to structural lesions (such as hippocampal sclerosis and cortical dysplasia), but it is difficult to capture real-time electro-physiological information of abnormal neuron discharge; while the electroencephalogram (EEG) signal can directly record the electro-activity related to epilepsy (such as spiky wave and high frequency oscillation), but it has the defects of low spatial resolution, being easily affected by skull resistance impedance attenuation, and single electroencephalogram signal being easily interfered by environmental noise and physiological artifacts, resulting in fuzzy positioning.

[0014] The present invention forms a complementarity between structural information and electrophysiological information by synchronously acquiring MRI and EEG signals. In the data preprocessing stage, brain tissue segmentation and structural feature extraction are performed on the MRI image, and artifact removal and time-frequency analysis are performed on the EEG signal to convert the original signal into a structured feature that can be used for analysis. Subsequently, a two-way constraint rule is established in the cross-modal confidence dynamic evaluation link, and the confidence of the EEG electrophysiological features is calibrated using the MRI structural features, and vice versa. For example, when MRI shows that there is a structural abnormality in a certain brain area, the confidence of the "false positive" electrophysiological features caused by non-epileptic factors (such as artifacts) in the EEG signal of the area is reduced; when EEG detects strong epilepsy-related electrical activities such as high-frequency oscillations, the credibility of the MRI structural features of the corresponding brain area is enhanced, thereby eliminating interference between the two modal signals.

[0015] Finally, during the localization model construction and training phase, a graph convolutional network encodes the structural connectivity of brain regions, while a temporal convolutional network models the temporal dependencies of EEG signals. By integrating these two features, deep learning is used to explore potential correlations between MRI and EEG data, resulting in highly accurate localization of epileptic foci. During this process, the model learns the synergistic effects of the two modalities based on extensive clinical data, further optimizing localization accuracy.

[0016] MRI alone can easily miss functional epileptogenic foci without structural abnormalities, while EEG alone is susceptible to artifacts and misinterpretations. This method utilizes bidirectionally constrained cross-modality confidence calibration, leveraging MRI structural information to verify the authenticity of EEG electrophysiological signals. It also uses abnormal EEG electrical activity to assist in localizing areas of functional abnormalities not captured by MRI, significantly reducing the risk of misinterpretation of both modalities. For example, when EEG detects widespread spikes, combining MRI structural features can eliminate false positives caused by poor scalp electrode contact and accurately locate the true epileptogenic foci, improving localization accuracy compared to single-modality methods.

[0017] Integrate the anatomical structural information of MRI with the electrophysiological dynamic information of EEG to avoid positioning deviations caused by the limitations of a single modality. In complex cases (such as MRI-negative epilepsy and multifocal epilepsy), the two modal data verify and complement each other, providing clinicians with a more comprehensive diagnostic basis. For example, for patients with no obvious structural abnormalities on MRI, the high-frequency oscillation characteristics of EEG can be combined with the MRI brain area connection information to locate potential epileptogenic networks, making the diagnostic results more reliable and clinically valuable. The full-process automation design from data acquisition, preprocessing to positioning output reduces the subjectivity and time-consuming nature of manual analysis. The deep learning model automatically extracts key features of MRI and EEG. Compared with traditional manual reading and signal analysis, the time for localizing epileptic foci can be shortened from hours to minutes, significantly improving clinical work efficiency, especially for emergency or large-scale screening scenarios.

[0018] The output epileptogenic focus positioning result is accompanied by a confidence score, which helps doctors intuitively judge the reliability of the positioning result. When formulating a surgical plan, high-confidence positioning results can be directly used as the basis for resection, while low-confidence results suggest that further examination or adjustment of the diagnostic strategy is needed, reducing the risk of surgery and improving the precision and safety of the treatment plan.

[0019] In summary, the application can reduce noise when fusing magnetic resonance and electroencephalogram, and has high accuracy in epileptic focus positioning.

[0020] Further, in the data acquisition module:

[0021] The magnetic resonance image acquisition uses a 3D-MPRAGE sequence, with a layer thickness of ≤0.8mm and a matrix size of ≥256×256, to obtain a full-brain gray matter density matrix Where N is the number of brain region divisions, divided into 90 brain regions based on the AAL template, and a T2-FLAIR sequence is simultaneously acquired to exclude cerebrospinal fluid artifacts, constructing a multi-contrast structural feature vector X mri = [[ρ, t, FA], where t is the cortical thickness vector and FA is the anisotropy fraction vector calculated based on the DTI sequence, with the formula:

[0022]

[0023] Where λ1≥λ2≥λ3 are the eigenvalues of the diffusion tensor;

[0024] The electroencephalogram signal acquisition uses 128 head electrodes, i.e., an extension of the international 10-20 electrode system, with a sampling frequency f s = 1024Hz, a 50Hz notch filter and a 0.1-500Hz bandpass filter are configured, and a reference electrode signal V ref (t) is recorded synchronously, environmental noise is eliminated through a common-mode rejection algorithm, and the corrected electroencephalogram signal V cal (t) = V(t) - V ref (t) is obtained, where V(t) is the original lead signal;

[0025] The space-time synchronization calibration aligns the magnetic resonance scan trigger signal and the electroencephalogram acquisition clock in time through a GPS clock module, with a timestamp error Δt≤1ms, constructing a synchronous time mapping function τ: t mri → t eeg The spatial coordinates of the magnetic resonance structural image and the head electrode positions of the electroencephalogram signal are mapped to the same standard brain space through a boundary element model, with a resolution of 1mm 3 ;

[0026] High-frequency oscillation pre-screening: Short-time Fourier transform (STFT) was performed on the EEG signals with parameters of STFT, 50ms window length, 75% overlap rate, and Kaiser window with β = 6. The frequency domain energy distribution of the equivalent dipole source in each brain region was calculated, and the definition was:

[0027] Fast ripple, 200-500Hz, power

[0028] Ripple, 80-200Hz, power

[0029] And extract the oscillation duration D = ∑(P HFO >0.5×P max )Δt, where P max is the maximum oscillation power of a single lead, Δt is the time resolution, and the high-frequency oscillation feature vector H = [P fast ,P ripple ,D].

[0030] Furthermore, in the data preprocessing module:

[0031] For magnetic resonance images, the U-Net++ network based on deep learning is used to jointly segment the collected T1-weighted magnetic resonance images and T2-FLAIR images; let the input T1-weighted image be I T1 , T2-FLAIR image is I T2-FLAIR , concatenate the two in the channel dimension to get I input =[I T1 ;I T2-FLAIR ];

[0032] The loss function of the U-Net++ network adopts the weighted sum of Dice loss and cross entropy loss, that is, L total =αL Dice +(1-α)L CE , where α∈[0,1] is the weight coefficient;

[0033] The Dice loss is defined as where p i is the value of the i-th pixel in the brain tissue segmentation probability map predicted by the network, g i is the value of the i-th pixel in the true brain tissue segmentation label;

[0034] The cross entropy loss is defined as

[0035] By minimizing L total , obtain accurate brain tissue segmentation results, dividing it into gray matter, white matter and cerebrospinal fluid areas;

[0036] For the segmented gray matter area, calculate the gray matter density ρ of each brain region i , where i = 1, 2, ..., N, N is the number of brain regions, and the calculation formula is where R i represents the pixel set of the i-th brain region, |R i | is the number of elements in the set, I T1 (j) is the value of the jth pixel in the T1-weighted image;

[0037] At the same time, the cortical thickness t of each brain region was calculated using a surface deformation-based method. i ; Combined with the fractional anisotropy FA calculated based on DTI sequence i , construct the magnetic resonance structure feature vector X mri =[ρ1,…,ρ N ,t1,…,t N ,FA1,…,FA N ] T ;

[0038] For EEG signals, independent component analysis (ICA) is used, combined with the adaptive threshold method to remove artifacts in EEG signals. First, the corrected EEG signal V cal (t) Perform ICA decomposition to obtain the independent component matrix S and the mixing matrix A, namely V cal (t) = AS;

[0039] For each independent component S k (t), where k = 1, 2, ..., M, M is the number of independent components, and its kurtosis K is calculated. k and correlation coefficient r k The kurtosis calculation formula is: where μ k It's S k The mean of (t), σ k is its standard deviation; the correlation coefficient r k It's S k (t) Correlation coefficient with typical electrooculogram and electromyography artifact template signals;

[0040] Set the adaptive threshold θ K and θ r , when K k >θ K and|r k |>θ r When , the independent component is determined to be an artifact component, it is set to zero, and then V clean (t) = AS clean Reconstruct the EEG signal V after removing artifacts clean (t), where S clean is the independent component matrix after setting the artifact components to zero;

[0041] After removing artifacts, the EEG signal V clean (t) Perform wavelet packet transform (WPT); select Daubechies wavelet as the wavelet basis, and the decomposition layer is L; for each lead of the EEG signal, decompose it into different frequency bands f l , where l=1,2,…,2 L ; Calculate the energy E of each frequency band l , the formula is where c l (n) is the wavelet packet coefficient of the lth frequency band, N s is the number of sampling points of the signal; combined with the extracted high-frequency oscillation feature vector H = [P fast ,P ripple ,D], construct the electrophysiological feature matrix X eeg , where each row represents the electrophysiological characteristics of a brain region.

[0042] Furthermore, in the cross-modal confidence dynamic evaluation module, the bidirectional constraint rules include:

[0043] Construct the Structure Abnormality Index, or SAI:

[0044]

[0045] in:

[0046] ρ i , t i , FA i is the extracted gray matter density, cortical thickness and anisotropy fraction of the i-th brain region;

[0047] is the average structural parameter of the whole brain, σ ρ , σ t , σ FA is the corresponding standard deviation;

[0048] α1, α2, and α3 are weight coefficients, satisfying α1+α2+α3=1, which are optimized through clinical data to α1=0.5, α2=0.3,

[0049] α3=0.2;

[0050] Dynamically adjust EEG feature confidence based on SAI:

[0051]

[0052] in:

[0053] is the confidence of the electrophysiological characteristics of the i-th brain region;

[0054] β is the adjustment parameter that controls the degree of attenuation of electrophysiological confidence by structural abnormalities;

[0055] Norm(E i ) is the constructed electrophysiological feature matrix X eeg The normalized result of the i-th row.

[0056] Furthermore, the two-way constraint rule also includes:

[0057] Constructing the Electrophysiological Seizure Index (ESI):

[0058]

[0059] in:

[0060] P fast,i 、P ripple,i 、D i is the extracted fast ripple power, ripple power and oscillation duration of the i-th brain region;

[0061] is the average oscillation duration of the whole brain, BurstRate i is the spike firing rate (detected by wavelet transform);

[0062] γ1=0.6, γ2=0.3, and γ3=0.1 are clinical optimization weights;

[0063] Dynamically adjust the confidence level of magnetic resonance features based on ESI:

[0064]

[0065] in:

[0066] is the confidence of the magnetic resonance structural feature of the i-th brain region;

[0067] δ = 5 is the steepness parameter of the Sigmoid function, and θ = 0.5 is the threshold;

[0068] Norm(M i ) is the constructed magnetic resonance structure feature vector X mri The normalized result of the i-th element.

[0069] Furthermore, the cross-modal confidence dynamic evaluation module is further configured to:

[0070] Constructing a spatiotemporal correlation matrix The element C ijRepresents the cross-modal spatiotemporal correlation between brain regions i and j:

[0071]

[0072] in:

[0073] ρ spatial (i, j) is the inverse of the spatial distance between brain regions i and j (calculated based on the AAL template);

[0074] τ i , τ j is the timestamp after time-space synchronization;

[0075] λ = 50ms is the time window parameter;

[0076] Propagate confidence through graph convolutional networks, or GCNs:

[0077]

[0078] in:

[0079] A=C+I is the adjacency matrix (I is the identity matrix), and D is the degree matrix;

[0080] is the initial confidence vector;

[0081] W (l) is the learnable weight matrix, σ is the ReLU activation function;

[0082] The final cross-modal confidence matrix is:

[0083] Conf final =Softmax(H (L) )

[0084] Where L is the number of GCN layers.

[0085] Furthermore, in the positioning model construction and training module:

[0086] Construct a hierarchical structural encoding module to perform multi-scale characterization of brain region structural connectivity:

[0087]

[0088] in:

[0089] is the l-th layer structural feature representation, and the initial input is the constructed X mri ;

[0090] A k is the k-th scale adjacency matrix, which is generated by thresholding the structural covariance matrix, with thresholds of 0.2, 0.5, and 0.8 respectively; αk is a learnable attention weight that satisfies

[0091] is a scale-specific weight matrix, K = 3 represents three scales;

[0092] Design a causal temporal module with dilated convolution to capture the long-range temporal dependencies of EEG signals:

[0093]

[0094] in:

[0095] is the electrophysiological feature representation at time t, and the initial input is the constructed X eeg ;

[0096] * indicates dilated convolution operation, dilation factor d = 2 l (l is the number of layers);

[0097] W f 、W g are the weights of feature extraction and gating units respectively, and ⊙ is the element-wise product;

[0098] b is the generated cross-modal confidence matrix Conf final The corresponding row vector of is used to dynamically adjust the weight of time series feature extraction.

[0099] Furthermore, in the positioning model construction and training module:

[0100] Construct a cross-modal interactive attention module to calculate the dynamic fusion weights of structural features and electrophysiological features:

[0101]

[0102] The similarity function is defined as:

[0103]

[0104] are the structural features of the i-th brain region and the electrophysiological features of the j-th time window output by MS-GCN and Causal-TCN respectively;

[0105] W a 、W s 、W e 、v a is a learnable parameter;

[0106] The coupled features after fusion are expressed as:

[0107]

[0108] wherein denotes the feature concatenation operation, combined with the spatio-temporal correlation matrix C, the final fused feature is:

[0109]

[0110] wherein, is the feature vector of the ith brain region after fusion operation; C i,: is the row vector composed of all elements of the ith row in the spatio-temporal correlation matrix C, which is used to assign a weight to each element in .

[0111] Further, in the positioning model construction and training module:

[0112] Monte Carlo Dropout is introduced to realize the model uncertainty quantification, and the localization confidence is calculated as:

[0113] Conf loc = exp(-Var MC (p pred ))

[0114] wherein:

[0115] p pred is the probability distribution of the three-dimensional coordinates of the epileptogenic focus predicted by the model;

[0116] Var MC (·) is the variance calculated by M forward propagation (M = 50):

[0117]

[0118] A loss function containing localization accuracy and uncertainty consistency is designed:

[0119]

[0120] wherein:

[0121] is the mean square error loss;

[0122] is the cross-entropy loss;

[0123] is the uncertainty calibration loss, wherein is the true confidence verified based on the intracranial electrode, calculated by the distance between the electrode and the predicted point;

[0124] λ1 = 0.6, λ2 = 0.3, λ3 = 0.1 are balance coefficients, optimized by the validation set. BRIEF DESCRIPTION OF DRAWINGS

[0125] Figure 1 This is a logic block diagram of an embodiment of an artificial intelligence localization system for epileptic foci based on magnetic resonance imaging and electroencephalography;

[0126] Figure 2 This is the workflow diagram of the data acquisition module;

[0127] Figure 3 This is the workflow diagram of the data preprocessing module;

[0128] Figure 4 This is the workflow diagram of the cross-modal confidence dynamic evaluation module;

[0129] Figure 5 Workflow diagram for building and training modules for localization models. DETAILED DESCRIPTION

[0130] The following is further described in detail through specific implementation methods:

[0131] Artificial intelligence localization system for epileptic focus based on magnetic resonance imaging and EEG (such as Figure 1 ), including the following:

[0132] The data acquisition module is used to synchronously acquire T1-weighted magnetic resonance images and scalp EEG signals;

[0133] The data preprocessing module is used to segment brain tissue on magnetic resonance images, obtain the structural features of each brain region, construct a vector containing the structural features of the whole brain, and obtain the magnetic resonance structural feature vector; remove artifacts from the EEG signal, extract the EEG features of each brain region through time-frequency analysis, construct a matrix containing the electrophysiological features of the whole brain, and obtain the electrophysiological features;

[0134] The cross-modal confidence dynamic assessment module is used to establish bidirectional constraint rules to calibrate the confidence of magnetic resonance structural feature vectors and electrophysiological features;

[0135] The localization model construction and training module is used to encode the structural connectivity of brain regions through a graph convolutional network to capture the structural characteristics of brain regions; the temporal dependence of EEG signals is modeled through a temporal convolutional network to extract electrophysiological temporal features; the interaction weights between structural and electrophysiological features are calculated, and the fusion ratio of the two is dynamically allocated based on this weight to achieve deep fusion of structural-electrophysiological coupling features; the fused features are input into a fully connected layer, and the three-dimensional coordinates of the epileptic focus and the localization confidence are output, where the localization confidence is calculated using a model uncertainty assessment mechanism; during training, the coordinates of the epileptic focus recorded by intracranial electrodes are used as the supervision signal, and a joint optimization is adopted using the mean square error loss and the cross entropy loss, while constraining the localization coordinate accuracy and the confidence prediction accuracy.

[0136] The localization result output module is used to filter out brain regions with confidence scores below a preset threshold based on the coordinates and confidence scores of the epileptogenic focus output by the model, combined with the cross-modal confidence matrix, to ultimately determine the three-dimensional spatial position of the epileptogenic focus and the corresponding confidence score.

[0137] This example uses a 28-year-old patient with medically refractory epilepsy as an example to illustrate the complete workflow of an AI-based epilepsy focus localization system based on MRI and EEG. The patient had been receiving medication for three years with limited success, and the clinician decided to use this method for preoperative epilepsy focus localization to assess surgical feasibility.

[0138] In the data acquisition module (such as Figure 2 shown):

[0139] Magnetic resonance imaging was acquired using a 3D-MPRAGE sequence with a slice thickness of ≤0.8 mm and a matrix size of ≥256 × 256 to obtain a whole-brain gray matter density matrix. Where N is the number of brain regions divided, which is divided into 90 brain regions based on the AAL template, and T2-FLAIR sequences are collected simultaneously to exclude cerebrospinal fluid artifacts and construct a multi-contrast structural feature vector X mri =[ρ,t,FA], where t is the cortical thickness vector and FA is the fractional anisotropy vector calculated based on the DTI sequence. The calculation formula is:

[0140]

[0141] Where λ1≥λ2≥λ3 is the eigenvalue of the diffusion tensor.

[0142] EEG signals were collected using 128-lead scalp electrodes, an extension of the international 10-20 lead system, with a sampling frequency of f s =1024Hz, equipped with a 50Hz notch filter and a 0.1-500Hz bandpass filter, and synchronously recording the reference electrode signal V ref (t), the environmental noise is eliminated by the common mode rejection algorithm to obtain the corrected EEG signal V cal (t)=V(t)-V ref (t), where V(t) is the original lead signal;

[0143] The time-space synchronization calibration uses the GPS clock module to align the magnetic resonance scan trigger signal with the EEG acquisition clock, with a timestamp error of Δt≤1ms, and construct a synchronous time mapping function τ:t mri →t eeg The spatial coordinates of the magnetic resonance image and the scalp lead position of the EEG signal are mapped to the same standard brain space through the boundary element model with a resolution of 1mm. 3 ;

[0144] High-frequency oscillation pre-screening: Short-time Fourier transform (STFT) was performed on the EEG signals with parameters of STFT, 50ms window length, 75% overlap rate, and Kaiser window with β = 6. The frequency domain energy distribution of the equivalent dipole source in each brain region was calculated, and the definition was:

[0145] Fast ripple, 200-500Hz, power

[0146] Ripple, 80-200Hz, power

[0147] And extract the oscillation duration D = ∑(P HFO >0.5×P max )Δt, where P max is the maximum oscillation power of a single lead, Δt is the time resolution, and the high-frequency oscillation feature vector H = [P fast ,P ripple ,D].

[0148] Specific use: Use a 3.0T magnetic resonance imaging machine, use 3D-MPRAGE sequence for T1 weighted imaging, the parameters are set as follows: layer thickness 0.8mm, matrix size 256×256, and scanning time is about 5 minutes. At the same time, use T2-FLAIR sequence to exclude cerebrospinal fluid artifacts, and the scanning time is about 3 minutes. Based on the AAL (Automated Anatomical Labeling) template, the whole brain is divided into 90 brain regions to obtain the gray matter density matrix

[0149] In addition, the diffusion data were collected using the DTI (Diffusion Tensor Imaging) sequence to calculate the fractional anisotropy (FA) of each brain region. For example, the FA value of a brain region in the left temporal lobe of a patient was calculated as follows: the eigenvalue of the diffusion tensor of this brain region is known to be λ1 = 1.5 × 10 -3 mm 2 / s,λ2=0.8×10 -3 mm 2 / s,λ3=0.5×10 -3 mm 2 / s, according to the formula

[0150]

[0151] Substituting into the calculation, we can get:

[0152]

[0153] Combine the cortical thickness vector t and the FA vector to construct the multi-contrast structural feature vector X mri =[ρ,t,FA].

[0154] The EEG signals were collected using 128-lead scalp electrodes (expanded from the international 10-20 lead system) with a sampling frequency of f s =1024Hz, equipped with 50Hz notch filter and 0.1-500Hz bandpass filter. Synchronously record the reference electrode signal V ref (t), the corrected EEG signal V is obtained through the common mode rejection algorithm cal (t)=V(t)-V ref (t).

[0155] The GPS clock module is used to align the MRI scan trigger signal with the EEG acquisition clock to ensure that the timestamp error Δt is less than 1ms. The boundary element model (BEM) is used to map the spatial coordinates of the MRI structure image and the scalp lead position of the EEG signal to the MNI standard brain space (resolution 1mm). 3 The EEG signal was subjected to short-time Fourier transform (STFT) with the following parameters: window length 50ms, overlap rate 75%, and Kaiser window (β=6). Taking a certain brain area in the right frontal lobe as an example, the fast ripple (200-500Hz) power P was calculated. fast :Assuming that the STFT result of the brain area in a certain period of time is |STFT(f)|, then

[0156]

[0157] Similarly, calculate the ripple (80-200Hz) power P ripple , and extract the oscillation duration D, construct the high-frequency oscillation feature vector H = [P fast ,P ripple ,D].

[0158] In the data preprocessing module (such as Figure 3 shown):

[0159] For magnetic resonance images, the U-Net++ network based on deep learning is used to jointly segment the collected T1-weighted magnetic resonance images and T2-FLAIR images; let the input T1-weighted image be I T1 , T2-FLAIR image is I T2-FLAIR , concatenate the two in the channel dimension to get I input =[I T1 ;I T2-FLAIR ];

[0160] The loss function of the U-Net++ network adopts the weighted sum of Dice loss and cross entropy loss, that is, L total =αL Dice +(1-α)L CE , where α∈[0,1] is the weight coefficient;

[0161] The Dice loss is defined as where p i is the value of the i-th pixel in the brain tissue segmentation probability map predicted by the network, g i is the value of the i-th pixel in the true brain tissue segmentation label;

[0162] The cross entropy loss is defined as

[0163] By minimizing L total , obtain accurate brain tissue segmentation results, dividing it into gray matter, white matter and cerebrospinal fluid areas;

[0164] For the segmented gray matter area, calculate the gray matter density ρ of each brain region i , where i = 1, 2, ..., N, N is the number of brain regions, and the calculation formula is where R i represents the pixel set of the i-th brain region, |R i | is the number of elements in the set, I T1 (j) is the value of the jth pixel in the T1-weighted image;

[0165] At the same time, the cortical thickness t of each brain region was calculated using a surface deformation-based method. i ; Combined with the fractional anisotropy FA calculated based on DTI sequence i , construct the magnetic resonance structure feature vector X mri =[ρ1,…,ρ N ,t1,…,t N ,FA1,…,FA N ] T ;

[0166] For EEG signals, independent component analysis (ICA) is used, combined with the adaptive threshold method to remove artifacts in EEG signals. First, the corrected EEG signal V cal (t) Perform ICA decomposition to obtain the independent component matrix S and the mixing matrix A, namely V cal (t) = AS;

[0167] For each independent component S k (t), where k = 1, 2, ..., M, M is the number of independent components, and its kurtosis K is calculated. k and correlation coefficient r k The kurtosis calculation formula is: where μ k It's S k The mean of (t), σ k is its standard deviation; the correlation coefficient r k It's S k(t) Correlation coefficient with typical electrooculogram and electromyography artifact template signals;

[0168] Set the adaptive threshold θ K and θ r , when K k >θ K and|r k |>θ r When , the independent component is determined to be an artifact component, it is set to zero, and then V clean (t) = AS clean Reconstruct the EEG signal V after removing artifacts clean (t), where S clean is the independent component matrix after setting the artifact components to zero;

[0169] After removing artifacts, the EEG signal V clean (t) Perform wavelet packet transform (WPT); select Daubechies wavelet as the wavelet basis, and the decomposition layer is L; for each lead of the EEG signal, decompose it into different frequency bands f l , where l=1,2,…,2 L ; Calculate the energy E of each frequency band l , the formula is where c l (n) is the wavelet packet coefficient of the lth frequency band, N s is the number of sampling points of the signal; combined with the extracted high-frequency oscillation feature vector H = [P fast ,P ripple ,D], construct the electrophysiological feature matrix X eeB , where each row represents the electrophysiological characteristics of a brain region.

[0170] Specific use: Use U-Net++ network to T1 weighted image I T1 and T2-FLAIR images I T2-FLAIr Perform joint segmentation. The two are spliced ​​in the channel dimension to obtain I input =[I T1 ;I T2-FLAIR ], as the network input.

[0171] The loss function of the U-Net++ network adopts the weighted sum of Dice loss and cross entropy loss, with a weight coefficient α = 0.7. Assume that at a certain pixel point, the brain tissue segmentation probability map value p predicted by the network is i =0.8, the true label value g i =1, then: Dice loss:

[0172]

[0173] Cross Entropy Loss:

[0174] L CE =-[1×log(0.8)+(1-1)×log(1-0.8)]≈0.22

[0175] Total loss L total =0.7×0.09+(1-0.7)×0.22=0.123.

[0176] By minimizing L total The brain tissue segmentation results are divided into gray matter, white matter and cerebrospinal fluid areas. The gray matter density ρ of each brain area is calculated. i For example, the gray matter density of the left hippocampus: the pixel set R of this brain region is known to be i Contains 100 pixels, I T1 The sum of the pixel values ​​of (j) is 500, then

[0177]

[0178] Combined cortical thickness t i and FA value, construct the magnetic resonance structure feature vector X mri .

[0179] The corrected EEG signal V cal (t) Perform independent component analysis (ICA). Assume that 50 independent components are obtained by decomposition, and the kth independent component S k (t), calculate its kurtosis K k and correlation coefficient r k .

[0180] Known S k The mean μ of (t) k =0.2, standard deviation σ k =0.1, its fourth-order central moment E[(S k (t)-μ k ) 4 ]=0.005, then the kurtosis

[0181]

[0182] If the correlation coefficient r between this component and the electrooculogram artifact template signal k =0.8, set the adaptive threshold θ K =3,θ r =0.5, due to K k >θ K and|r k |>θ r , determine that the component is an artifact component and set it to zero.

[0183] By V clean (t) = AS cleanReconstruct the EEG signal V after removing artifacts clean (t).

[0184] V clean (t) Perform wavelet packet transform (WPT), select Daubechies wavelet, and decompose the layer number L = 5. Decompose each lead signal into 2 5 = 32 frequency bands, calculate the energy E of each frequency band l For example, the wavelet packet coefficient c of the lth frequency band of a certain lead l (n) The sum of squares is 20, then E l = 20. Combined with the high-frequency oscillation eigenvector H, the electrophysiological feature matrix X is constructed. eeg .

[0185] Cross-modal confidence dynamic evaluation module (such as Figure 4 As shown), the two-way constraint rules include:

[0186] Construct the Structure Abnormality Index, or SAI:

[0187]

[0188] in:

[0189] ρ i , t i , FA i is the extracted gray matter density, cortical thickness and anisotropy fraction of the i-th brain region;

[0190] is the average structural parameter of the whole brain, σ ρ , σ t , σ FA is the corresponding standard deviation;

[0191] α1, α2, and α3 are weight coefficients, satisfying α1+α2+α3=1, which are optimized through clinical data to α1=0.5, α2=0.3,

[0192] α3=0.2;

[0193] Dynamically adjust EEG feature confidence based on SAI:

[0194]

[0195] in:

[0196] is the confidence of the electrophysiological characteristics of the i-th brain region;

[0197] β is the adjustment parameter that controls the degree of attenuation of electrophysiological confidence by structural abnormalities;

[0198] Norm(E i ) is the constructed electrophysiological feature matrix X eeg The normalized result of the i-th row.

[0199] Bidirectional constraint rules also include:

[0200] Construct the Electrophysiological Seizure Index, or ESI:

[0201]

[0202] in:

[0203] P fast,i 、P ripple,i 、D i is the extracted fast ripple power, ripple power and oscillation duration of the i-th brain region;

[0204] is the average oscillation duration of the whole brain, BurstRate i is the spike firing rate (detected by wavelet transform);

[0205] γ1=0.6, γ2=0.3, and γ3=0.1 are clinical optimization weights;

[0206] Dynamically adjust the confidence level of magnetic resonance features based on ESI:

[0207]

[0208] in:

[0209] is the confidence of the magnetic resonance structural feature of the i-th brain region;

[0210] δ = 5 is the steepness parameter of the Sigmoid function, and θ = 0.5 is the threshold;

[0211] Norm(M i ) is the constructed magnetic resonance structure feature vector X mri The normalized result of the i-th element.

[0212] The cross-modal confidence dynamic evaluation step also includes:

[0213] Constructing a spatiotemporal correlation matrix The element C ij Represents the cross-modal spatiotemporal correlation between brain regions i and j:

[0214]

[0215] in:

[0216] ρ spatial (i, j) is the inverse of the spatial distance between brain regions i and j (calculated based on the AAL template);

[0217] τ i , τ j is the timestamp after time-space synchronization;

[0218] λ = 50ms is the time window parameter;

[0219] Propagate confidence through graph convolutional networks, or GCNs:

[0220]

[0221] in:

[0222] A=C+I is the adjacency matrix (I is the identity matrix), and D is the degree matrix;

[0223] is the initial confidence vector;

[0224] W (l) is the learnable weight matrix, σ is the ReLU activation function;

[0225] The final cross-modal confidence matrix is:

[0226] Conf final =Softmax(H (L) )

[0227] Where L is the number of GCN layers (take L = 2), which is used for threshold filtering in the subsequent S5 step.

[0228] Specific use: Calculate the structural abnormality index (SAI) of a brain region in the left temporal lobe: the gray matter density ρ of the brain region is known i =4, average gray matter density of the whole brain Standard deviation σ ρ =0.5; cortical thickness t i =2, average cortical thickness of the whole brain Standard deviation σ t =0.3; FA value FA i =0.6, average FA value of the whole brain Standard deviation σ FA =0.1.

[0229]

[0230] Confidence of electrophysiological characteristics of this brain region:

[0231]

[0232] Calculate the electrophysiological epilepsy index (ESI) of a brain region in the right frontal lobe: the fast ripple power P in the brain region is known fast,i =15, ripple power P ripple,i =5, oscillation duration D i =100, average oscillation duration of the whole brain BurstRate i =2.

[0233]

[0234] Confidence of the MRI structural features of this brain region:

[0235]

[0236] Construct the spatiotemporal correlation matrix C, taking the left temporal lobe brain area i and the right frontal lobe brain area j as an example: the inverse of the spatial distance between the two is known spatial (i, j) = 0.2, timestamp difference |τ i -τ j |=20ms,λ=50ms,

[0237]

[0238] The confidence is propagated through the graph convolutional network (GCN), and the final cross-modal confidence matrix Conf is obtained after 2 layers of GCN final .

[0239] Positioning model construction and training module (such as Figure 5 (shown):

[0240] Construct a hierarchical structural encoding module to perform multi-scale characterization of brain region structural connectivity:

[0241]

[0242] in:

[0243] is the l-th layer structural feature representation, and the initial input is the constructed X mri ;

[0244] A k is the k-th scale adjacency matrix, which is generated by thresholding the structural covariance matrix, with thresholds of 0.2, 0.5, and 0.8 respectively;

[0245] α k is a learnable attention weight that satisfies

[0246] is a scale-specific weight matrix, K = 3 represents three scales;

[0247] Design a causal temporal module with dilated convolution to capture the long-range temporal dependencies of EEG signals:

[0248]

[0249] in:

[0250] is the electrophysiological feature representation at time t, and the initial input is the constructed X eeg ;

[0251] * indicates dilated convolution operation, dilation factor d = 2 l (l is the number of layers);

[0252] W f 、W g are the weights of feature extraction and gating units respectively, and ⊙ is the element-wise product;

[0253] b is the generated cross-modal confidence matrix Conf final The corresponding row vector of is used to dynamically adjust the weight of time series feature extraction.

[0254] In the positioning model construction and training module:

[0255] Construct a cross-modal interactive attention module to calculate the dynamic fusion weights of structural features and electrophysiological features:

[0256]

[0257] The similarity function is defined as:

[0258]

[0259] are the structural features of the i-th brain region and the electrophysiological features of the j-th time window output by MS-GCN and Causal-TCN respectively;

[0260] W a 、W s 、W e 、v a is a learnable parameter;

[0261] The coupled features after fusion are expressed as:

[0262]

[0263] in Represents the feature splicing operation, combined with the spatiotemporal correlation matrix C, the final fusion feature is:

[0264]

[0265] where, is the feature vector of the ith brain region after fusion operation; C i,: is the row vector composed of all elements of the ith row in the spatiotemporal correlation matrix C, which is used to assign a weight to each element in .

[0266] In the positioning model construction and training module:

[0267] Monte Carlo Dropout is introduced to quantify the uncertainty of the model, and the localization confidence is calculated as:

[0268] Conf loc = exp(-Var MC (p pred ))

[0269] where,

[0270] p pred is the probability distribution of the 3D coordinates of the epileptogenic zone predicted by the model;

[0271] Var MC (·) is the variance calculated by M forward propagation (M = 50):

[0272]

[0273] A loss function is designed to include localization accuracy and uncertainty consistency:

[0274]

[0275] where,

[0276] is the mean square error loss;

[0277] is the cross-entropy loss;

[0278] is the uncertainty calibration loss, where is the true confidence verified by intracranial electrodes, calculated by the distance between the electrode and the predicted point;

[0279] λ1 = 0.6, λ2 = 0.3, λ3 = 0.1 are balance coefficients, optimized by the validation set.

[0280] Specific use: Multi-scale representation of brain region structural connectivity, taking the left hippocampus as an example, in the first layer of MS-GCN:

[0281]

[0282] in is the initial structural feature vector of the brain region, A k is the adjacency matrix under different thresholds (0.2, 0.5, 0.8), α k are learnable weights.

[0283] By capturing the time dependency of EEG signals through Causal-TCN, at a certain time t, the electrophysiological characteristics of the right frontal lobe are updated as follows:

[0284]

[0285] The expansion factor d = 2 l , b is Conf final corresponds to a row vector.

[0286] Calculate the fusion weight of the structural features of the left hippocampus and the electrophysiological features of a certain time window in the right frontal lobe:

[0287]

[0288] Assumptions but

[0289]

[0290] The coupled features after fusion:

[0291] Final fusion features:

[0292]

[0293] Monte Carlo Dropout is introduced to quantify uncertainty, and the variance is calculated through 50 forward propagations:

[0294]

[0295] Positioning reliability:

[0296] Conf loc =exp(-Var MC (p pred ))

[0297] The model is trained using a joint loss function:

[0298]

[0299] in They are mean square error loss, cross entropy loss and uncertainty calibration loss respectively.

[0300] According to the epileptogenic focus coordinates and confidence output by the model, combined with the cross-modal confidence matrix Conffinal , set the threshold value as 0.7. If the confidence of a brain area is lower than the threshold value, the brain area is filtered. Finally, it is determined that the patient's epileptogenic focus is located in the left temporal lobe, and the confidence score is 0.85, which provides accurate positioning basis for subsequent surgical treatment.

[0301] Through the test of the clinical data of 100 epilepsy patients, the positioning accuracy of the method reaches 89%, which is improved by 25% compared with the traditional single modality method; the average positioning error is 4.2mm, which meets the clinical operation demand. At the same time, the positioning confidence output by the model has a high correlation with the actual operation verification result (correlation coefficient 0.82), which effectively assists the doctor to evaluate the reliability of the positioning result.

[0302] The above is only an embodiment of the application, and the application is not limited to this embodiment. The application is not limited to the field involved in the embodiment, and the common knowledge of specific structures and characteristics in the scheme is not described too much. The ordinary skilled person in the art knows all the ordinary technical knowledge in the technical field of the application before the application date or the priority date, can know all the prior art in the field, and has the ability to apply conventional experimental means before that date. The ordinary skilled person in the art can improve and implement the scheme under the guidance of the disclosure, and some typical known structures or known methods should not be an obstacle for the ordinary skilled person in the art to implement the application. It should be noted that for those skilled in the art, without departing from the structure of the application, a number of modifications and improvements can be made, which should also be considered as the protection scope of the application, and these will not affect the effect and practicality of the application. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. An artificial intelligence localization system for epileptic foci based on magnetic resonance imaging and electroencephalography, characterized by: Includes the following: The data acquisition module is used to synchronously acquire T1-weighted magnetic resonance images and scalp EEG signals; The data preprocessing module is used to segment brain tissue on magnetic resonance images, obtain the structural features of each brain region, construct a vector containing the structural features of the whole brain, and obtain the magnetic resonance structural feature vector; remove artifacts from the EEG signal, extract the EEG features of each brain region through time-frequency analysis, construct a matrix containing the electrophysiological features of the whole brain, and obtain the electrophysiological features; The cross-modal confidence dynamic assessment module is used to establish bidirectional constraint rules to calibrate the confidence of magnetic resonance structural feature vectors and electrophysiological features; The localization model construction and training module is used to encode the structural connectivity of brain regions through graph convolutional networks and capture the structural characteristics of brain regions; the temporal dependence of EEG signals is modeled through temporal convolutional networks and electrophysiological temporal characteristics are extracted; The interaction weights between structural and electrophysiological features are calculated, and the fusion ratio between the two is dynamically allocated based on these weights, achieving deep fusion of structural-electrophysiological coupling features. The fused features are input into a fully connected layer, which outputs the three-dimensional coordinates of the epileptogenic focus and the localization confidence, where the localization confidence is calculated using a model uncertainty assessment mechanism. During training, the coordinates of the epileptogenic focus recorded by intracranial electrodes are used as the supervision signal, and a joint optimization method using mean square error loss and cross entropy loss is used to simultaneously constrain the localization coordinate accuracy and confidence prediction accuracy. The localization result output module is used to filter out brain regions with confidence scores below a preset threshold based on the coordinates and confidence scores of the epileptogenic focus output by the model, combined with the cross-modal confidence matrix, to ultimately determine the three-dimensional spatial position of the epileptogenic focus and the corresponding confidence score.

2. The artificial intelligence localization system for epileptic focus based on magnetic resonance and EEG according to claim 1, characterized in that: In the data acquisition module: Magnetic resonance imaging was acquired using a 3D-MPRAGE sequence with a slice thickness of ≤0.8 mm and a matrix size of ≥256 × 256 to obtain a whole-brain gray matter density matrix. Where N is the number of brain regions divided, which is divided into 90 brain regions based on the AAL template, and T2-FLAIR sequences are collected simultaneously to exclude cerebrospinal fluid artifacts and construct a multi-contrast structural feature vector X mri =[ρ,t,FA], where t is the cortical thickness vector and FA is the fractional anisotropy vector calculated based on the DTI sequence. The calculation formula is: Where λ1≥λ2≥λ3 is the eigenvalue of the diffusion tensor; EEG signals were collected using 128-lead scalp electrodes, an extension of the international 10-20 lead system, with a sampling frequency of f s =1024Hz, equipped with a 50Hz notch filter and a 0.1-500Hz bandpass filter, and synchronously recording the reference electrode signal V ref (t), the environmental noise is eliminated by the common mode rejection algorithm to obtain the corrected EEG signal V cal (t)=V(t)-V ref (t), where V(t) is the original lead signal; The time-space synchronization calibration uses the GPS clock module to align the magnetic resonance scan trigger signal with the EEG acquisition clock, with a timestamp error of Δt≤1ms, and construct a synchronous time mapping function τ:t mri →t eeg The spatial coordinates of the magnetic resonance image and the scalp lead position of the EEG signal are mapped to the same standard brain space through the boundary element model with a resolution of 1mm. 3 ; High-frequency oscillation pre-screening: Short-time Fourier transform (STFT) was performed on the EEG signals with parameters of STFT, 50ms window length, 75% overlap rate, and Kaiser window with β = 6. The frequency domain energy distribution of the equivalent dipole source in each brain region was calculated, and the definition was: Fast ripple, 200-500Hz, power Ripple, 80-200Hz, power And extract the oscillation duration D = ∑(P HFO >0.5×P max )Δt, where P max is the maximum oscillation power of a single lead, Δt is the time resolution, and the high-frequency oscillation feature vector H = [P fast ,P ripple ,D].

3. The artificial intelligence localization system for epileptic focus based on magnetic resonance and EEG according to claim 2, characterized in that: In the data preprocessing module: For magnetic resonance images, the U-Net++ network based on deep learning is used to jointly segment the collected T1-weighted magnetic resonance images and T2-FLAIR images; let the input T1-weighted image be I T1 , T2-FLAIR image is I T2-FLAIR , concatenate the two in the channel dimension to get I input =[I T1 ;I T2-FLAIR ]; The loss function of the U-Net++ network adopts the weighted sum of Dice loss and cross entropy loss, that is, L total =αL Dice +(1-α)L CE , where α∈[0,1] is the weight coefficient; The Dice loss is defined as where p i is the value of the i-th pixel in the brain tissue segmentation probability map predicted by the network, g i is the value of the i-th pixel in the true brain tissue segmentation label; The cross entropy loss is defined as By minimizing L total , obtain accurate brain tissue segmentation results, dividing it into gray matter, white matter and cerebrospinal fluid areas; For the segmented gray matter area, calculate the gray matter density ρ of each brain region i , where i = 1, 2, ..., N, N is the number of brain regions, and the calculation formula is where R i represents the pixel set of the i-th brain region, |R i | is the number of elements in the set, I T1 (j) is the value of the jth pixel in the T1-weighted image; At the same time, the cortical thickness t of each brain region was calculated using a surface deformation-based method. i ; Combined with the fractional anisotropy FA calculated based on DTI sequence i , construct the magnetic resonance structure feature vector X mri =[ρ1,…,ρ N ,t1,…,t N ,FA1,…,FA N ] T ; For EEG signals, independent component analysis (ICA) is used, combined with the adaptive threshold method to remove artifacts in EEG signals. First, the corrected EEG signal V cal (t) Perform ICA decomposition to obtain the independent component matrix S and the mixing matrix A, namely V cal (t) = AS; For each independent component S k (t), where k = 1, 2, ..., M, M is the number of independent components, and its kurtosis K is calculated. k and correlation coefficient r k The kurtosis calculation formula is: where μ k It's S k The mean of (t), σ k is its standard deviation; the correlation coefficient r k It's S k (t) Correlation coefficient with typical electrooculogram and electromyography artifact template signals; Set the adaptive threshold θ K and θ r , when K k >θ K and|r k |>θ r When , the independent component is determined to be an artifact component, it is set to zero, and then V clean (t) = AS clean Reconstruct the EEG signal V after removing artifacts clean (t), where S clean is the independent component matrix after setting the artifact components to zero; After removing artifacts, the EEG signal V clean (t) Perform wavelet packet transform (WPT); select Daubechies wavelet as the wavelet basis, and the decomposition layer is L; for each lead of the EEG signal, decompose it into different frequency bands f l , where l=1,2,…,2 L ; Calculate the energy E of each frequency band l , the formula is where c l (n) is the wavelet packet coefficient of the lth frequency band, N s is the number of sampling points of the signal; combined with the extracted high-frequency oscillation feature vector H = [P fast ,P ripple ,D], construct the electrophysiological feature matrix X eeg , where each row represents the electrophysiological characteristics of a brain region.

4. The artificial intelligence localization system for epileptic focus based on magnetic resonance and EEG according to claim 3, characterized in that: In the cross-modal confidence dynamic evaluation module, the bidirectional constraint rules include: Construct the Structure Abnormality Index, or SAI: in: ρ i , t i , FA i is the extracted gray matter density, cortical thickness and anisotropy fraction of the i-th brain region; is the average structural parameter of the whole brain, σ ρ , σ t , σ FA is the corresponding standard deviation; α1, α2, and α3 are weight coefficients, satisfying α1+α2+α3=1, which are optimized to α1=0.5, α2=0.3, and α3=0.2 through clinical data; Dynamically adjust EEG feature confidence based on SAI: in: is the confidence of the electrophysiological characteristics of the i-th brain region; β is the adjustment parameter that controls the degree of attenuation of electrophysiological confidence by structural abnormalities; Norm(E i ) is the constructed electrophysiological feature matrix X eeg The normalized result of the i-th row.

5. The artificial intelligence localization system for epileptic focus based on magnetic resonance and EEG according to claim 4, characterized in that: The two-way constraint rules also include: Construct the Electrophysiological Seizure Index, or ESI: in: P fast,i 、P ripple,i 、D i is the extracted fast ripple power, ripple power and oscillation duration of the i-th brain region; is the average oscillation duration of the whole brain, BurstRate i is the spike firing rate (detected by wavelet transform); γ1=0.6, γ2=0.3, and γ3=0.1 are clinical optimization weights; Dynamically adjust the confidence level of magnetic resonance features based on ESI: in: is the confidence of the magnetic resonance structural feature of the i-th brain region; δ = 5 is the steepness parameter of the Sigmoid function, and θ = 0.5 is the threshold; Norm(M i ) is the constructed magnetic resonance structure feature vector X mri The normalized result of the i-th element.

6. The artificial intelligence localization system for epileptic focus based on magnetic resonance and EEG according to claim 5, characterized in that: The cross-modal confidence dynamic evaluation module is further used to: Constructing a spatiotemporal correlation matrix The element C ij Represents the cross-modal spatiotemporal correlation between brain regions i and j: in: ρ spatial (i, j) is the inverse of the spatial distance between brain regions i and j (calculated based on the AAL template); τ i , τ j is the timestamp after time-space synchronization; λ = 50ms is the time window parameter; Propagate confidence through graph convolutional networks, or GCNs: in: A=C+I is the adjacency matrix (I is the identity matrix), and D is the degree matrix; is the initial confidence vector; W (l) is the learnable weight matrix, σ is the ReLU activation function; The final cross-modal confidence matrix is: Conf final =Softmax(H (L) ) Where L is the number of GCN layers.

7. The artificial intelligence localization system for epileptic focus based on magnetic resonance and EEG according to claim 6, characterized in that: In the positioning model construction and training module: Construct a hierarchical structural encoding module to perform multi-scale characterization of brain region structural connectivity: in: is the l-th layer structural feature representation, and the initial input is the constructed X mri ; A k is the k-th scale adjacency matrix, which is generated by thresholding the structural covariance matrix, with thresholds of 0.2, 0.5, and 0.8 respectively; α k is a learnable attention weight that satisfies is a scale-specific weight matrix, K = 3 represents three scales; Design a causal temporal module with dilated convolution to capture the long-range temporal dependencies of EEG signals: in: is the electrophysiological feature representation at time t, and the initial input is the constructed X eeg ; * indicates dilated convolution operation, dilation factor d = 2 l (l is the number of layers); W f 、W g are the weights of feature extraction and gating units respectively, and ⊙ is the element-wise product; b is the generated cross-modal confidence matrix Conf final The corresponding row vector of is used to dynamically adjust the weight of time series feature extraction.

8. The artificial intelligence localization system for epileptic focus based on magnetic resonance and electroencephalogram according to claim 7, characterized in that: In the positioning model construction and training module: Construct a cross-modal interactive attention module to calculate the dynamic fusion weights of structural features and electrophysiological features: The similarity function is defined as: are the structural features of the i-th brain region and the electrophysiological features of the j-th time window output by MS-GCN and Causal-TCN respectively; W a 、W s 、W e 、v a is a learnable parameter; The coupled features after fusion are expressed as: Where ⊕ represents the feature concatenation operation. Combined with the spatiotemporal correlation matrix C, the final fusion feature is: in, is the feature vector of the i-th brain region obtained after the fusion operation; C i,: It is a row vector composed of all elements in the i-th row of the spatiotemporal correlation matrix C, and its function is to Each element in is assigned a weight.

9. The artificial intelligence localization system for epileptic focus based on magnetic resonance and electroencephalogram according to claim 8, characterized in that: In the positioning model construction and training module: Monte Carlo Dropout is introduced to quantify model uncertainty, and the position confidence is calculated as: cf. loc =exp(-Var MC (p pred )) in: p pred is the probability distribution of the three-dimensional coordinates of the epileptogenic focus predicted by the model; Var MC (·) is the variance calculated by M forward propagations (M=50): Design a loss function that includes positioning accuracy and uncertainty consistency: in: is the mean square error loss; is the cross entropy loss; is the uncertainty calibration loss, where To provide the true confidence based on intracranial electrode verification, the distance between the electrode and the predicted point is calculated; λ1=0.6, λ2=0.3, and λ3=0.1 are balance coefficients, which are tuned through the validation set.

10. An artificial intelligence method for localizing epileptic foci based on magnetic resonance imaging and electroencephalography, characterized in that: A system as claimed in any one of claims 1 to 9 is employed.

Images