Method and system for determining epileptic focus

Abstract

The present invention provides a method and system for determining an epileptic focus, comprising: identifying the type of epileptic focus by combining PET-CT with typical clinical symptoms and EEG signals for multimodal fusion; establishing a mapping relationship between the features obtained by splicing EEG neural signals and clinical typical features and the features obtained by splicing SEEG neural signals and clinical typical features through a generative adversarial network; determining the contribution weight of each SEEG electrode to epileptic seizures and the coverage range of the epileptic signal at its location by varying the magnitude of the electrical stimulation current, and obtaining a final precise range of the epileptic focus through comprehensive evaluation; and optimizing a focal epileptic focus boundary recognition model in PET-CT based on the precise range of the epileptic focus obtained from the neural signal, so that the model has the ability to recognize lesions that cannot be distinguished by the human eye and to delineate the boundaries of epileptic focuses with unclear boundaries.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical assistance technology, and in particular to a method and system for determining an epileptic focus. Background Art

[0002] Epilepsy can usually be diagnosed through consultation with medical history, clinical symptoms, cranial nerve signals, and head MRI / CT. There are many factors that cause epilepsy: focal lesions with clear boundaries, such as intracranial tumors, cerebral hemorrhage, or enlarged vascular malformations, which compress the cranial nerves and can cause epileptic seizures; epileptogenic lesions with unclear boundaries, such as those that cannot be clearly identified even with imaging methods such as PET-CT / MRI; and epileptogenic lesions that cannot be identified even with imaging methods such as PET-CT / MRI. Electroencephalograms and clinical symptoms can only roughly focus on abnormal discharges in entire functional areas or brain structures, leading to epilepsy. For example, visual stimulation can cause abnormal discharges in the entire occipital lobe neurons, auditory stimulation can cause abnormal discharges in the temporal lobe neurons, and abnormal emotions or movements can cause abnormal discharges in the frontal lobe neurons. More severe cases include diffuse epilepsy, which is also the most difficult and incurable.

[0003] Currently, the preliminary diagnosis of epilepsy can be determined based on the patient's clinical manifestations combined with electroencephalogram (EEG). However, if you want to accurately locate the epileptogenic focus, it is best to use intracranial electrodes to determine the boundaries of the epileptogenic focus through electrical stimulation and neural signal analysis. Summary of the Invention

[0004] The present invention provides an epileptic focus determination method and system, which are used to solve the defects in the prior art in accurately locating the epileptic focus of epilepsy.

[0005] In a first aspect, the present invention provides a method for determining an epileptogenic focus, comprising:

[0006] Collect EEG, body movement features, voice features, and emotional features of epilepsy patients to extract epilepsy recognition features;

[0007] The epilepsy recognition features are fused and spliced, PET-CT tomography images are extracted, an abnormal area segmentation model is trained, and the type of lesion is determined;

[0008] Obtaining a three-dimensional reconstructed image of the brain based on the PET-CT tomographic image;

[0009] Determining the implantation range of the stereoscopic electroencephalogram (SEEG) electrodes according to the three-dimensional reconstructed brain image;

[0010] Performing SEEG stimulation and signal acquisition within the implantation range of the SEEG electrode to identify typical epileptic symptoms, and performing feature fusion and splicing on the typical epileptic symptoms to obtain mapped SEEG signal features;

[0011] Determining the epileptic focus area boundary based on the mapped SEEG signal features;

[0012] Based on the abnormal region segmentation model, the boundary of the epileptic focus region is optimized.

[0013] According to a method for determining an epileptogenic focus provided by the present invention, EEG, limb movement features, voice features, and emotional features of an epileptic patient are collected to extract epilepsy identification features, including:

[0014] Collecting multiple channels of brain nerve signals from an epileptic patient through an electroencephalogram (EEG) cap, and decomposing the multiple channels of brain nerve signals into multiple frequency band signals to obtain signal features;

[0015] Decomposing the action video into consecutive frames, inputting the trained first convolutional neural network, identifying multiple consecutive images exceeding a preset confidence level, flattening the last feature layer of the first convolutional neural network, adding the multiple consecutive images, and randomly extracting features to obtain action features, where the length of the action feature is 1 greater than the length of the signal feature;

[0016] The sound signal is window-cropped and input into the trained second convolutional neural network to identify multiple consecutive window segments exceeding a preset confidence level. The last feature layer of the second convolutional neural network is flattened and then the multiple consecutive window segments are added together to randomly extract features to obtain sound features, where the length of the sound features is 1 greater than the length of the signal features.

[0017] Decomposing the emotional video into consecutive frames, inputting the frames into a trained third convolutional neural network, identifying multiple consecutive images exceeding a preset confidence level, flattening the last feature layer of the third convolutional neural network, adding the multiple consecutive images, and randomly extracting features to obtain emotional features, wherein the length of the emotional features is 1 greater than the length of the signal features;

[0018] The signal features, the motion features, the sound features, and the emotion features are averaged respectively to obtain the signal feature average, the motion feature average, the sound feature average, and the emotion feature average. The signal feature average, the motion feature average, the sound feature average, and the emotion feature average are weightedly fitted by a machine learning algorithm to obtain the epilepsy recognition feature.

[0019] According to a method for determining an epileptic focus provided by the present invention, a method for collecting brain nerve signals of an epileptic patient through multiple channels using an electroencephalogram (EEG) cap is used, and the brain nerve signals of the patient through the multiple channels are decomposed into multiple frequency band signals to obtain signal features, including:

[0020] Based on the deep learning classification model, each frequency band is trained to identify epilepsy models and obtain the frequency band with the highest accuracy;

[0021] Based on the highest accuracy frequency band, each channel is trained to identify an epilepsy model, the accuracy of each model is ranked, and several channels with the highest accuracy are selected;

[0022] Randomly extracting the plane coordinates of 4 channels from the plurality of channels each time and the model recognition accuracy corresponding to the 4 channels;

[0023] performing equivalent coordinate processing on a combination of the plane coordinates and the model recognition accuracy with the highest accuracy, and determining attribution label attributes of the equivalent coordinates, wherein the attribution label attributes include frontal lobe, parietal lobe, occipital lobe, and temporal lobe;

[0024] The last feature layer of the model corresponding to the obtained best combination is flattened and features are randomly extracted, and all features are spliced ​​together to obtain the signal features.

[0025] According to a method for determining an epileptogenic focus provided by the present invention, the epilepsy identification features are fused and spliced, a PET-CT tomographic image is extracted, an abnormal region segmentation model is trained, and the type of the causative focus is determined, including:

[0026] After concatenating the signal feature, the action feature, the sound feature, and the emotion feature, a feature square matrix is ​​formed, and each element in the feature square matrix is ​​scaled to a grayscale image within a preset grayscale range;

[0027] Obtain PET-CT tomographic images of patients confirmed to have epilepsy, train a deep learning model for normal / abnormal identification, and output abnormal PET-CT images and normal PET-CT images;

[0028] The abnormal PET-CT image is used to train an abnormal region segmentation model. If a segmentation boundary is determined to exist in the image, the abnormal region is considered to be focal with a clear boundary; otherwise, the abnormal region is considered to be focal with no clear boundary.

[0029] The features of the last convolutional layer of the normal PET-CT image are flattened and concatenated with the concatenated features, and then input into a BP neural network for binary classification recognition to obtain the type of the lesion.

[0030] According to a method for determining an epileptic focus provided by the present invention, a three-dimensional reconstructed image of the brain is obtained based on the PET-CT tomographic image, comprising:

[0031] A trained brain functional area and anatomical structure image segmentation model is obtained, and combined with the PET-CT tomographic scan image, the brain functional area and anatomical structure are segmented and the boundary lines are marked, and the three-dimensional reconstruction image of the brain is obtained using three-dimensional reconstruction software.

[0032] According to a method for determining an epileptogenic focus provided by the present invention, SEEG stimulation and signal acquisition are performed within the SEEG electrode implantation range to identify typical epileptic symptoms, and feature fusion and splicing of the typical epileptic symptoms are performed to obtain mapped SEEG signal features, including:

[0033] One SEEG electrode is provided with current stimulation at any one of several typical SEEG current values, and the remaining SEEG electrodes are responsible for receiving neural signals;

[0034] The collected electrode numbers are stored in a two-dimensional matrix to obtain several two-dimensional matrices;

[0035] Input the electrode signals in the plurality of two-dimensional matrices into the trained epilepsy recognition model respectively, remove the two-dimensional matrices that fail to recognize epilepsy signals, silence the corresponding SEEG electrodes and no longer use them, and take the remaining SEEG electrode sets;

[0036] Inputting the remaining SEEG electrode set into the trained epilepsy recognition model, flattening the last feature layer of the model and randomly extracting features, which are then concatenated with the features corresponding to the remaining SEEG electrode set to obtain new features;

[0037] The new features are processed in sequence according to the steps of signal features, action features, sound features and emotion features to obtain the mapped SEEG signal features.

[0038] According to a method for determining an epileptic focus provided by the present invention, the boundary of an epileptic focus area is determined by mapping SEEG signal features, comprising:

[0039] Determining the centroids of electrode arrangement regions corresponding to different current magnitudes based on the similarity of the same current in the mapped SEEG signal features;

[0040] Based on the preset similarity ratio and centroid, the boundaries of the epileptic lesion area with different current sizes are determined.

[0041] In a second aspect, the present invention further provides an epileptogenic focus determination system, comprising:

[0042] The acquisition module is used to collect EEG, body movement characteristics, voice characteristics and emotional characteristics of epilepsy patients and extract epilepsy recognition features;

[0043] A type determination module is used to fuse and splice the epilepsy identification features, extract positron emission tomography (PET-CT) tomography images, train an abnormal region segmentation model, and determine the type of lesion;

[0044] A reconstruction module, configured to obtain a three-dimensional reconstructed image of the brain based on the PET-CT tomographic image;

[0045] A range determination module is used to determine the implantation range of the stereoscopic brain wave signal SEEG electrode according to the three-dimensional reconstructed brain image;

[0046] A mapping module is used to perform SEEG stimulation and signal acquisition within the implantation range of the SEEG electrode, identify typical epileptic symptoms, perform feature fusion and splicing on the typical epileptic symptoms, and obtain mapped SEEG signal features;

[0047] a boundary determination module, configured to determine the boundary of the epileptic focus area based on the mapped SEEG signal features;

[0048] A boundary optimization module is used to optimize the boundary of the epileptic focus area based on the abnormal area segmentation model.

[0049] In a third aspect, the present invention further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for determining an epileptic focus as described above is implemented.

[0050] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for determining an epileptic focus.

[0051] The present invention provides an epileptic focus determination method and system that identifies the type of epileptic focus by combining PET-CT with typical clinical symptoms and EEG signals for multimodal fusion. A mapping relationship is established between the features obtained by splicing EEG neural signals and clinical typical features and the features obtained by splicing SEEG neural signals and clinical typical features through a generative adversarial network. By varying the magnitude of the electrical stimulation current, the contribution weight of each SEEG electrode to epileptic seizures and the coverage range of the epileptic signal at the location are determined, and a comprehensive evaluation is performed to determine the precise range of the final epileptic focus. Based on the precise range of the epileptic focus obtained from the neural signal, the boundary recognition model for focal epileptic focuses in PET-CT is optimized, enabling the model to recognize lesions that cannot be discerned by the human eye and to delineate the boundaries of epileptic focuses with unclear boundaries. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 This is one of the flow charts of the method for determining the epileptogenic focus provided by the present invention;

[0054] Figure 2 This is the second flow chart of the method for determining the epileptogenic focus provided by the present invention;

[0055] Figure 3 Schematic diagram of an epilepsy identification unit provided by the present invention;

[0056] Figure 4 This is a schematic diagram of the EEG electrode distribution provided by the present invention;

[0057] Figure 5 Schematic diagram of a pathogenic focus type determination unit provided by the present invention;

[0058] Figure 6 Schematic diagram of the SEEG electrode implantation position provided by the present invention;

[0059] Figure 7 Schematic diagram of the SEEG stimulation and signal acquisition unit provided by the present invention;

[0060] Figure 8 This is a schematic diagram of the 0.5mA electrode provided by the present invention for preliminarily determining the epilepsy boundary;

[0061] Figure 9 is a schematic structural diagram of the epileptogenic focus determination system provided by the present invention;

[0062] Figure 10 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0063] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0064] Figure 1 FIG. 1 is a flow chart of a method for determining an epileptic focus according to an embodiment of the present invention. Figure 1 As shown, including:

[0065] Step 100: Collect EEG, body movement features, voice features, and emotional features of epilepsy patients to extract epilepsy recognition features;

[0066] Step 200: performing feature fusion and splicing on the epilepsy recognition features, extracting PET-CT tomography images, training an abnormal region segmentation model, and determining the type of lesion focus;

[0067] Step 300: Obtaining a three-dimensional reconstructed image of the brain based on the PET-CT tomographic image;

[0068] Step 400: Determine the implantation range of the SEEG electrode based on the three-dimensional reconstructed brain image;

[0069] Step 500: Perform SEEG stimulation and signal acquisition within the SEEG electrode implantation range to identify typical epileptic symptoms, perform feature fusion and splicing on the typical epileptic symptoms, and obtain mapped SEEG signal features;

[0070] Step 600: determining the epileptic focus area boundary based on the mapped SEEG signal features;

[0071] Step 700: Optimizing the boundary of the epileptic focus region based on the abnormal region segmentation model.

[0072] The specific solution process of the embodiment of the present invention is as follows Figure 2 As shown, specifically including:

[0073] Step S1, epilepsy identification unit, functional modules such as Figure 3 As shown, it includes four functional modules: electroencephalogram (EEG) signal feature extraction module, epileptic limb movement feature extraction module, typical sound feature extraction module and typical expression feature extraction module.

[0074] Step S1.1, EEG signal feature extraction module, when the patient has an epileptic seizure, collects the patient's brain nerve signals through the EEG cap (a total of N channels, N can be 64 / 128 / 256, etc.), and decomposes the nerve signals into multiple frequency bands delta (<4Hz), theta (4-7Hz), alpha (8-13Hz), beta (14-31Hz) and gamma (>32Hz), such as Figure 4 shown.

[0075] Step S1.1.1: Build a model based on a deep learning classification model, such as 2DCNN or 2DCNN+RNN, and train the model for epilepsy recognition in each frequency band to obtain the frequency band with the highest accuracy;

[0076] Step S1.1.2: Use the optimal frequency band obtained in step S1.1.1 to perform epilepsy identification on N channels separately. The specific steps are as follows:

[0077] 1) Build a model based on deep learning classification models, such as 1DCNN or 1DCNN+RNN, and train a model for epilepsy recognition in each channel (N models in total). Rank the accuracy of each model and select the m channels with the highest accuracy, where m can be 10 / 20 or any other number.

[0078] 2) Randomly select 4 channels from the m channels each time (recording the plane coordinates of these 4 channels simultaneously), and based on the model recognition accuracy in step 1), construct models using methods such as 2DCNN and 2DCNN+RNN, and train each combination of models to recognize epilepsy, until the combination with the highest accuracy is obtained;

[0079] 3) Equivalent coordinates are calculated for the combination with the highest accuracy in step 2), and one of the label attributes of the frontal lobe, parietal lobe, occipital lobe, and temporal lobe is determined based on the coordinates;

[0080] 4) Flattening the last feature layer of the model corresponding to the best combination obtained in step 2), and then randomly extracting features to ensure that the feature length is L. In the embodiment of the present invention, L=35;

[0081] 5) Concatenate the features obtained in step 4) with those obtained in step 3) to obtain new features.

[0082] Step S1.2: Typical limb movement feature extraction module, typical movements such as limb twitching, etc.

[0083] Step S1.2.1: Obtain the time-frequency information of the patient's epileptic seizure. Video footage can be captured by mobile phone or from video surveillance. Break the video down into frames (e.g., 25 frames / s). Use a trained convolutional neural network model1 (e.g., ResNet, VGG, etc.) to identify epileptic movements. Consecutive frames (M1 consecutive images, where M1 can be 10 / 20, etc.) must be identified as epileptic movements, and the model confidence for each of these M1 consecutive images must exceed 0.95.

[0084] Step S1.2.2: The model1 obtained in S1.2.1 performs feature extraction on M images, flattens the last feature layer of the model to obtain M1 image features, then performs the add() operation on the M1 image features, and finally randomly extracts features after add() to obtain and ensure that the feature length is L+1.

[0085] Step S1.3: Typical sound feature extraction module, typical sounds such as screaming, roaring, etc.

[0086] Step S1.3.1: Obtain the patient's epileptic seizure sound signal, which can come from a mobile phone recording or a voice recorder. Crop the sound signal into windows and use a trained neural network model 2 (e.g., 1DCNN, 1DCNN+RNN, RNN, etc.) to identify epileptic movements. Continuous windows (M2 consecutive windows, where M2 can be 10 / 20, etc.) are identified as epileptic sounds, and the model confidence for each of the M2 consecutive windows exceeds 0.95.

[0087] Step S1.3.2: The model2 obtained in S1.3.1 performs feature extraction on the M2 window segments, flattens the last feature layer of the model to obtain M2 window features, then performs the add() operation on the M2 window features, and finally randomly extracts features after add() to obtain and ensure that the feature length is L+1.

[0088] Step S1.4: Typical emotion feature extraction module, typical actions such as clenched teeth and pale face.

[0089] Step S1.4.1: Obtain the time-frequency information of the patient's epileptic seizure. Video footage can be captured by mobile phone or from video surveillance. Break the video down into frames (e.g., 25 frames / s). Use a trained convolutional neural network model3 (e.g., ResNet, VGG, etc.) to identify epileptic movements. Consecutive frames (M3 consecutive images, where M3 can be 10 / 20, etc.) must be identified as epileptic movements, and the model confidence level for all M3 consecutive images must exceed 0.95.

[0090] Step S1.4.2: The model3 obtained in S1.4.1 performs feature extraction on M images, flattens the last feature layer of the model to obtain M3 image features, then performs the add() operation on the M3 image features, and finally randomly extracts features after add() to obtain and ensure that the feature length is L+1.

[0091] Step S1.5: Epilepsy identification, 、 、 、 Find the mean respectively, and get 、 、 、 , through machine learning algorithms (decision tree, random forest, etc.), 、 、 、 Perform weighted fitting for two categories: epilepsy or non-epileptic, and obtain 、 、 、 The corresponding weights 、 、 、 .

[0092] Step S2: Pathogenic focus type determination unit, functional modules such as Figure 5 As shown, it includes two functional modules: feature fusion and splicing module and PET-CT feature extraction module.

[0093] Step S2.1: Feature fusion and splicing module, which combines the features obtained from steps S1.1 to S1.4 、 、 、 The four features are concatenated () to obtain a feature concatenation A of length 144. Then, each 12 lengths are combined into a row to generate a matrix, and then each element in the matrix is ​​scaled to a grayscale image of [0,255]. .

[0094] Step S2.2: Obtain PET-CT tomographic scan images of patients confirmed to be epileptic in step S1, and train a deep learning model model4 (Resnet, VGG, etc.) for identifying abnormalities. Abnormalities include diffuse lesions and those without obvious lesions in the functional areas involved; abnormalities include focal lesions without clear boundaries and focal lesions with clear boundaries (these two situations cannot be distinguished here and can be determined in the subsequent step S2.3).

[0095] Step S2.3: For the abnormal PET-CT images in step S2.2, train an abnormal region segmentation model (such as Unet, D-link, Mask-Rcnn). If the boundary can be segmented, it is a focal lesion with clear boundaries (such as tumors, blood clots, etc., which can be directly surgically removed without subsequent SEEG signal acquisition). If the boundary cannot be segmented, it is a focal lesion with no clear boundaries.

[0096] Step S2.4: For the PET-CT images without abnormalities in step S2.2, when model 4 identifies a PET-CT image without abnormalities, the features of the last convolutional layer are flattened and concatenated with the features obtained in step S2.1 (A) for a concatenation operation. The concatenation is then input into the BP neural network for binary classification. The labels are diffuse and no obvious lesions in the involved functional areas (for example, if the temporal lobe or occipital lobe is involved, but no obvious abnormal areas can be observed in the PET-CT image).

[0097] Step S3: The three-dimensional reconstruction unit of the brain skull uses the PET-CT tomography image obtained in step S2 to load the trained brain functional area and anatomical structure image segmentation model, such as Unet, D-link, Mask-Rcnn and other deep learning models, to segment the brain functional areas and anatomical structures and mark their boundaries, such as the temporal lobe, occipital lobe, frontal lobe, hippocampus and hypothalamus, and then uses VTK and other three-dimensional reconstruction software to perform three-dimensional reconstruction of the brain skull.

[0098] Step S4: Electrode implantation range: Combined with the location of the focal lesion without clear boundaries or the target functional area in step S2.3, the stereo electroencephalography (SEEG) electrode implantation range is locked. When the SEEG electrode locks the precise boundary of the epileptogenic focus, multiple SEEG electrodes are usually implanted around the target area. The effect is shown in the figure below. Figure 6 The embodiment of the present invention is described using an abnormality in a functional area of ​​the cerebral cortex as an example. K SEEG electrodes are implanted at equal intervals around the boundary line of this functional area in the three-dimensional reconstructed image of the skull obtained in step S3. The value of K can be 6, 8, 10, etc., without limitation.

[0099] Step S5: SEEG stimulation and signal acquisition unit, functional modules such as Figure 7 As shown, it includes four functional modules: current stimulation module, neural signal acquisition module, neural signal pattern recognition module, and typical symptom recognition module.

[0100] Step S5.1: Current stimulation module. Each electrode requires a current of no more than 5 mA. The present invention uses 0.5 mA, 3 mA, and 5 mA as examples, but the present invention is not limited to these three current values. When one SEEG electrode provides current stimulation, the other SEEG electrodes only receive neural signals and do not provide stimulation.

[0101] Step S5.2: Take a SEEG electrode as number 1, and number the other electrodes sequentially. According to the current stimulation method of step S5.1, store the collected electrode numbers in a two-dimensional matrix (each matrix has K rows), for a total of Two-dimensional matrices, which are .

[0102] Step S5.3: In a two-dimensional matrix , respectively input into the trained epilepsy recognition model, such as 2DCNN / 2DCNN+RNN convolutional neural network model,

[0103] Step S5.3.1: Eliminate k two-dimensional matrices that fail to identify epilepsy, and the remaining A two-dimensional matrix is ​​formed (it should be noted that if all signals collected by applying 5mA current to an electrode fail to identify epilepsy, then applying 3mA or 0.5mA current will not identify epilepsy either), and the corresponding SEEG electrode is silenced and no longer used. At this time, only SEEG electrodes to be analyzed, forming a set ;

[0104] Step S5.3.2: The remaining electrodes are divided into three categories: only 0.5 mA current is needed to identify epilepsy, 3 mA current is needed to identify epilepsy, and 5 mA current is needed to identify epilepsy. Two-dimensional matrices give the current label attributes One of them.

[0105] Step S5.3.3: The two-dimensional matrices are input into the 2DCNN / 2DCNN+RNN SEEG neural signal epilepsy recognition model trained in step S5.3, and the last feature layer of the model is flattened and the features are randomly extracted. , ensure that the characteristic length is L, and and Perform feature splicing to obtain new features .

[0106] Step S5.4: Decompose the video of the patient at the time of epilepsy detection in step S5.3 into frames (e.g., 25 frames / s)

[0107] Then input the model1 model of step S1.2, select the image frames with the highest confidence level, m11 consecutive images before and m12 consecutive images after, and ensure that m11+m12+1=M1, extract features from the M1 images at this time, flatten the last feature layer of the model to obtain M1 image features, then perform the add() operation on the M1 image features, and finally randomly extract features from the features after add() to obtain And ensure that the characteristic length is L.

[0108] Step S5.5: Input the patient's sound signal when epilepsy can be identified in step S5.3 into the model2 model of step S1.3, select the first m21 and the last m22 consecutive cropping windows with the highest confidence, and ensure that m21+m22+1=M2, extract features from the M2 windows at this time, flatten the last feature layer of the model to obtain M2 sound features, then perform add() operation on the M2 sound features, and finally randomly extract features from the features after add() to obtain And ensure that the characteristic length is L.

[0109] Step S5.6: When epilepsy can be identified in step S5.3, the patient's video is disassembled into frames (such as 25 frames / s) and input into the model3 model of step S1.4. The continuous m31 and m32 frames before and after the image frame with the highest confidence are selected, and m31+m32+1=M3 are ensured. The features of the M3 images at this time are extracted, and the last feature layer of the model is flattened to obtain M3 image features. Then, the M3 image features are added () and the features after add () are randomly extracted to obtain features. And ensure that the characteristic length is L.

[0110] Step S5.7: Feature fusion and splicing module, which combines the features obtained from steps S5.3 to SS5.6 、 、 、 The four features are concatenated and the feature concatenation B with a length of 144 is obtained. . Each line is 12 long, generating The matrix, then Each element in the matrix is ​​scaled to a grayscale image of [0,255]. .

[0111] Step S6: mapping relationship.

[0112] Step S6.1: is the input image, To create a training set for decoding images, train the adversarial generative network model5 (such as GAN / DCGAN, etc.), and establish a mapping relationship between feature A extracted based on EEG+clinical manifestations and feature B extracted based on SEEG+clinical manifestations.

[0113] Step S6.2: Splice the feature image corresponding to the patient feature A Generate model5 obtained by step S6.1 , and compare Feature image corresponding to feature splicing B Similarity, methods for comparing image similarity, such as histogram method, SSIM (structural similarity) structural similarity method, hash similarity method, etc. Similarity results , sort them from large to small according to the similarity of 0.5mA / 3mA / 5mA, and store the corresponding electrode numbers in the corresponding order.

[0114] Step S7: epilepsy boundary determination.

[0115] Step S7.1: 0.5 mA epilepsy boundary determination, 0.5 mA similarity in step S6.2 , It can be 60%, 80%, etc. There is no specific requirement; for SEEG electrodes that meet the similarity conditions in step S7.1, the single electrode pattern recognition accuracy is trained by 1DCNN / RNN / 1DCNN+RNN and other methods , retain such electrodes, It can be 88%, 92%, etc. There is no specific requirement. The electrode set at this time is The area enclosed by these electrodes is the initial epileptic boundary, e.g. Figure 8 ABC area in the calculation of the combined centroid of ABC area .

[0116] Step S7.2: 3mA epilepsy boundary determination, determined in step S7.1 When the applied current is 3mA, , such electrode signals are trained by 1DCNN / RNN / 1DCNN+RNN and other methods to achieve single electrode pattern recognition accuracy. , It can be determined according to the actual situation, and no specific requirements are given here. All electrodes and points Maximum Euclidean distance At this time, the epilepsy boundary contour line determined in step S7.1 is expanded proportionally ,in It can be 10 / 20 / 100, depending on the actual situation. No specific requirements are given here. Calculate the centroid of the area enclosed by the epileptic boundary at this time .

[0117] Step S7.3: 5mA epilepsy boundary determination, determined in step S7.1 When the applied current is 5mA, , such electrode signals are trained by 1DCNN / RNN / 1DCNN+RNN and other methods to achieve single electrode pattern recognition accuracy. , It can be determined according to the actual situation, and no specific requirements are given here. All electrodes and points Maximum Euclidean distance ,like , expand the epilepsy boundary contour determined in step S7.2 proportionally At this time, the expanded epilepsy boundary line is the final epilepsy boundary line.

[0118] Step S8: If the abnormal region segmentation model trained in step S2.3 cannot identify the abnormal region or the abnormal region boundary, the epilepsy boundary determined in step S7 is recalibrated on the PET-CT tomography scan, and the mask boundary of the segmentation model is retrained to achieve the purpose of model optimization.

[0119] The optimized model has the following capabilities: 1) The optimized segmentation model can automatically identify epileptogenic lesions and their regions in PET-CT scans that are difficult for even experienced physicians to visually identify; 2) The optimized model can locally segment the boundaries of epileptogenic lesions in PET-CT scans that were previously difficult to segment (some lesions have infiltrated edges).

[0120] The following describes the epileptic focus determination system provided by the present invention. The epileptic focus determination system described below and the epileptic focus determination method described above can be referenced to each other.

[0121] Figure 9 FIG. 1 is a schematic diagram of the structure of the epileptogenic focus determination system provided by an embodiment of the present invention. Figure 9 As shown, it includes: an acquisition module 91, a type determination module 92, a reconstruction module 93, a range determination module 94, a mapping module 95, a boundary determination module 96 and a boundary optimization module 97, wherein:

[0122] The acquisition module 91 is used to collect EEG, limb movement characteristics, sound characteristics and emotional characteristics of epilepsy patients, and extract epilepsy identification features; the type determination module 92 is used to fuse and splice the epilepsy identification features, extract positron emission tomography (PET-CT) tomography images, train abnormal area segmentation models, and determine the type of lesions; the reconstruction module 93 is used to obtain a three-dimensional reconstructed image of the brain based on the PET-CT tomography images; the range determination module 94 is used to determine the implantation range of the stereoscopic brain wave signal SEEG electrode based on the three-dimensional reconstructed image of the brain; the mapping module 95 is used to perform SEEG stimulation and signal acquisition within the SEEG electrode implantation range, identify typical epilepsy symptoms, and perform feature fusion and splicing on the typical epilepsy symptoms to obtain mapped SEEG signal features; the boundary determination module 96 is used to determine the boundary of the epileptic focus area based on the mapped SEEG signal features; the boundary optimization module 97 is used to optimize the boundary of the epileptic focus area based on the abnormal area segmentation model.

[0123] Figure 10 An example of a physical structure diagram of an electronic device is shown below. Figure 10As shown, the electronic device may include: a processor (processor) 1010, a communication interface (Communications Interface) 1020, a memory (memory) 1030 and a communication bus 1040, wherein the processor 1010, the communication interface 1020, and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 can call the logic instructions in the memory 1030 to execute the epileptic focus determination method, which includes: collecting EEG, limb movement characteristics, sound characteristics and emotional characteristics of epilepsy patients to extract epilepsy identification features; performing feature fusion and splicing on the epilepsy identification features, extracting PET-CT tomographic scan images, training abnormal area segmentation models, and determining the type of lesion focus; obtaining a three-dimensional reconstructed image of the brain based on the PET-CT tomographic scan images; determining the implantation range of the stereoscopic brain wave signal SEEG electrode according to the three-dimensional reconstructed image of the brain; performing SEEG stimulation and signal acquisition within the SEEG electrode implantation range to identify typical epileptic symptoms, performing feature fusion and splicing on the typical epileptic symptoms to obtain mapped SEEG signal features; determining the boundary of the epileptic focus area by the mapped SEEG signal features; and optimizing the boundary of the epileptic focus area based on the abnormal area segmentation model.

[0124] Furthermore, the logic instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0125] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the epileptic focus determination method provided by the above-mentioned methods, the method comprising: collecting EEG, limb movement characteristics, sound characteristics and emotional characteristics of epilepsy patients to extract epilepsy identification features; performing feature fusion and splicing on the epilepsy identification features, extracting PET-CT tomographic scan images, training an abnormal area segmentation model, and determining the type of lesion focus; obtaining a three-dimensional reconstructed image of the brain based on the PET-CT tomographic scan images; determining the implantation range of the stereoscopic brain wave signal SEEG electrode based on the three-dimensional reconstructed image of the brain; performing SEEG stimulation and signal acquisition within the SEEG electrode implantation range to identify typical epileptic symptoms, performing feature fusion and splicing on the typical epileptic symptoms to obtain mapped SEEG signal features; determining the boundary of the epileptic focus area by the mapped SEEG signal features; and optimizing the boundary of the epileptic focus area based on the abnormal area segmentation model.

[0126] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0127] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for determining an epileptogenic focus, characterized in that: include: Collect EEG signals, body movement features, voice features, and emotional features of epilepsy patients to extract epilepsy recognition features; The epilepsy recognition features are fused and spliced, positron emission tomography (PET-CT) tomography images are extracted, an abnormal region segmentation model is trained, and the type of lesion is determined; Obtaining a three-dimensional reconstructed image of the brain based on the PET-CT tomographic image; Determining the implantation range of the stereoscopic electroencephalogram (SEEG) electrodes according to the three-dimensional reconstructed brain image; Performing SEEG stimulation and signal acquisition within the implantation range of the SEEG electrode to identify typical epileptic symptoms, and performing feature fusion and splicing on the typical epileptic symptoms to obtain mapped SEEG signal features; Determining the epileptic focus area boundary based on the mapped SEEG signal features; Optimizing the boundary of the epileptic lesion area based on the abnormal area segmentation model; The epilepsy recognition features are fused and spliced, PET-CT tomography images are extracted, an abnormal area segmentation model is trained, and the type of lesion is determined, including: After concatenating the signal feature, the action feature, the sound feature, and the emotion feature, a feature square matrix is ​​formed, and each element in the feature square matrix is ​​scaled to a grayscale image within a preset grayscale range; Obtain PET-CT tomographic images of patients confirmed to have epilepsy, train a deep learning model for normal / abnormal identification, and output abnormal PET-CT images and normal PET-CT images; The abnormal PET-CT image is used to train an abnormal region segmentation model. If a segmentation boundary is determined to exist in the image, the abnormal region is considered to be focal with a clear boundary; otherwise, the abnormal region is considered to be focal with no clear boundary. Flattening the last convolutional layer features of the normal PET-CT image and concatenating them with the concatenated features, inputting them into a BP neural network for binary classification recognition to obtain the type of the lesion; Within the SEEG electrode implantation range, SEEG stimulation and signal acquisition are performed to identify typical epileptic symptoms, and feature fusion and splicing of the typical epileptic symptoms are performed to obtain mapped SEEG signal features, including: One SEEG electrode is provided with current stimulation at any one of several typical SEEG current values, and the remaining SEEG electrodes are responsible for receiving neural signals; The collected electrode numbers are stored in a two-dimensional matrix to obtain several two-dimensional matrices; Input the electrode signals in the plurality of two-dimensional matrices into the trained epilepsy recognition model respectively, remove the two-dimensional matrices that fail to recognize epilepsy signals, silence the corresponding SEEG electrodes and no longer use them, and take the remaining SEEG electrode sets; Inputting the remaining SEEG electrode set into the trained epilepsy recognition model, flattening the last feature layer of the model and randomly extracting features, which are then concatenated with the features corresponding to the remaining SEEG electrode set to obtain new features; The new features are processed in sequence according to the steps of signal features, action features, sound features and emotion features to obtain the mapped SEEG signal features.

2. The method for determining the epileptogenic focus according to claim 1, wherein: Collect EEG, body movement features, voice features, and emotional features of epilepsy patients to extract epilepsy recognition features, including: Collecting multiple channels of brain nerve signals from an epileptic patient through an electroencephalogram (EEG) cap, and decomposing the multiple channels of brain nerve signals into multiple frequency band signals to obtain signal features; Decomposing the action video into consecutive frames, inputting the trained first convolutional neural network, identifying multiple consecutive images exceeding a preset confidence level, flattening the last feature layer of the first convolutional neural network, adding the multiple consecutive images, and randomly extracting features to obtain action features, where the length of the action feature is 1 greater than the length of the signal feature; The sound signal is window-cropped and input into the trained second convolutional neural network to identify multiple consecutive window segments exceeding a preset confidence level. The last feature layer of the second convolutional neural network is flattened and then the multiple consecutive window segments are added together to randomly extract features to obtain sound features, where the length of the sound features is 1 greater than the length of the signal features. Decomposing the emotional video into consecutive frames, inputting the frames into a trained third convolutional neural network, identifying multiple consecutive images exceeding a preset confidence level, flattening the last feature layer of the third convolutional neural network, adding the multiple consecutive images, and randomly extracting features to obtain emotional features, wherein the length of the emotional features is 1 greater than the length of the signal features; The signal features, the motion features, the sound features, and the emotion features are averaged respectively to obtain the signal feature average, the motion feature average, the sound feature average, and the emotion feature average. The signal feature average, the motion feature average, the sound feature average, and the emotion feature average are weightedly fitted by a machine learning algorithm to obtain the epilepsy recognition feature.

3. The method for determining epileptogenic focus according to claim 2, wherein: The electroencephalogram (EEG) cap is used to collect brain nerve signals from multiple channels of an epileptic patient, and the brain nerve signals from the multiple channels of the patient are decomposed into multiple frequency band signals to obtain signal features, including: Based on the deep learning classification model, each frequency band is trained to identify epilepsy models and obtain the frequency band with the highest accuracy; Based on the highest accuracy frequency band, each channel is trained to identify an epilepsy model, the accuracy of each model is ranked, and several channels with the highest accuracy are selected; Randomly extracting the plane coordinates of 4 channels from the plurality of channels each time and the model recognition accuracy corresponding to the 4 channels; performing equivalent coordinate processing on a combination of the plane coordinates and the model recognition accuracy with the highest accuracy, and determining attribution label attributes of the equivalent coordinates, wherein the attribution label attributes include frontal lobe, parietal lobe, occipital lobe, and temporal lobe; The last feature layer of the model corresponding to the obtained best combination is flattened and features are randomly extracted, and all features are spliced ​​together to obtain the signal features.

4. The method for determining epileptogenic focus according to claim 1, wherein: Obtaining a three-dimensional reconstructed image of the brain based on the PET-CT tomographic image, including: A trained brain functional area and anatomical structure image segmentation model is obtained, and combined with the PET-CT tomographic scan image, the brain functional area and anatomical structure are segmented and the boundary lines are marked, and the three-dimensional reconstruction image of the brain is obtained using three-dimensional reconstruction software.

5. The method for determining epileptogenic focus according to claim 1, wherein: Determining the epileptic focus area boundary based on the mapped SEEG signal features includes: Determining the centroids of electrode arrangement regions corresponding to different current magnitudes based on the similarity of the same current in the mapped SEEG signal features; Based on the preset similarity ratio and centroid, the boundaries of the epileptic lesion area with different current sizes are determined.

6. A system for determining an epileptic focus, based on the method for determining an epileptic focus according to any one of claims 1 to 5, characterized in that: include: The acquisition module is used to collect EEG, body movement characteristics, voice characteristics and emotional characteristics of epilepsy patients and extract epilepsy recognition features; A type determination module is used to fuse and splice the epilepsy identification features, extract positron emission tomography (PET-CT) tomography images, train an abnormal region segmentation model, and determine the type of lesion; A reconstruction module, configured to obtain a three-dimensional reconstructed image of the brain based on the PET-CT tomographic image; A range determination module is used to determine the implantation range of the stereoscopic brain wave signal SEEG electrode according to the three-dimensional reconstructed brain image; A mapping module is used to perform SEEG stimulation and signal acquisition within the implantation range of the SEEG electrode, identify typical epileptic symptoms, perform feature fusion and splicing on the typical epileptic symptoms, and obtain mapped SEEG signal features; a boundary determination module, configured to determine the boundary of the epileptic focus area based on the mapped SEEG signal features; A boundary optimization module is used to optimize the boundary of the epileptic focus area based on the abnormal area segmentation model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for determining the epileptic focus according to any one of claims 1 to 5 is implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for determining an epileptic focus according to any one of claims 1 to 5 is implemented.

Images