Deep learning method for magnetoencephalography spike localization model

Abstract

The present invention provides a deep learning method for a magnetoencephalogram spike wave localization model. The deep learning method comprises: performing data preprocessing on multi-channel magnetoencephalogram signals; performing preliminary screening of spike waves, and performing regional multi-channel weighted fusion of the screened multi-channel magnetoencephalogram signals to obtain band data; calculating manual feature information of the band data; and, inputting the band data and the corresponding manual feature information as a whole into a neural network to train the magnetoencephalogram spike wave localization model. With such a configuration, the intrinsic correlation between multi-channel magnetoencephalogram signals and the dynamic characteristics of the spike wave signal time series are taken into account, thereby overcoming the problems existing in the prior art. In one embodiment of the present invention, the effect of the scheme is further improved by partitioning the magnetoencephalogram channels, adopting a multi-channel weighted fusion strategy in each partition, and constructing a neural network architecture combining a gated recurrent unit and an attention mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetoencephalogram signal processing, and in particular to a deep learning method for a magnetoencephalogram spike wave localization model. Background Art

[0002] Epilepsy is a neurological disease characterized by abnormal excessive discharge of neurons in the epileptic focus of the brain, which leads to imbalance in the function of the central nervous system, causing symptoms such as involuntary movements and loss of consciousness in patients, seriously affecting the quality of life of patients and even endangering their lives and health. In the process of epilepsy diagnosis and treatment, some patients can suppress epileptic seizures through drug treatment. However, there are still a large number of patients who cannot control their condition through drugs. Such patients often need to undergo surgery to remove the epileptic focus in order to achieve effective treatment of the disease.

[0003] The premise of surgical treatment is to accurately locate the specific location of the epileptic focus in the brain. Traditional methods usually use electroencephalography for source tracing analysis. However, due to the volume conduction effect of EEG signals in propagation media such as cerebrospinal fluid and skull, as well as the inconsistency of the conductivity of these media, the spatial resolution of EEG signals is low, which limits its tracing accuracy. Magnetoencephalography can get rid of the above shortcomings, but traditional magnetoencephalography measurement devices are expensive, which brings a high economic burden to patients who need to be examined. In recent years, magnetoencephalography systems based on spin-exchange relaxation-free atomic magnetometers have developed rapidly. Compared with traditional magnetoencephalography systems based on superconducting quantum interferometers, atomic magnetometer magnetoencephalography systems have the advantages of flexible sensor configuration and no need for cryogenic cooling devices, and are expected to become the next generation of brain functional imaging technology.

[0004] Magnetoencephalography (MEG) is a non-invasive brain functional imaging technology with significant value in the fields of neuroscience and clinical applications. Since the EEG signal does not significantly attenuate when passing through media such as cerebrospinal fluid and skull, MEG has a high spatial resolution and can accurately locate epileptic foci. In the process of epilepsy signal source tracing and analysis, it is first necessary to clarify the specific location of the spike wave in the EEG signal, then extract the EEG data of all channels in this period, and derive and reconstruct the spatial distribution of the corresponding abnormal discharge source of epilepsy through the inverse solution algorithm. Abnormal discharge of epilepsy exhibits characteristics such as randomness, transientness and recurrence. Therefore, in clinical examinations, the MEG acquisition of epilepsy patients usually requires long-term recording to fully capture the relevant EEG activity. In addition, the detection of spike waves is mostly completed by clinical experts through manual annotation. Due to the high time course and multi-channel characteristics of MEG, the manual annotation process is time-consuming and laborious, and it is highly dependent on the professional experience of doctors, which increases the complexity and subjectivity of diagnosis. Under the combined effect of long-term EEG data acquisition and complex clinical diagnosis requirements, the development of automated and high-precision spike wave detection methods has become a key issue in neuroscience research. The research on these methods aims to meet the actual needs of clinical diagnosis by reducing human intervention and improving detection efficiency and accuracy.

[0005] At present, the algorithms for spike detection and localization are mainly divided into three categories: signal analysis, traditional machine learning, and deep learning. Signal analysis methods extract time domain or frequency domain features and make judgments based on preset thresholds. However, such methods have high computational complexity, require artificial design of specific standards, and are easily interfered by artifact signals, resulting in a high false detection rate. Spike detection methods based on traditional machine learning are mainly focused on optimizing feature space and improving classifier performance, and have shown high accuracy in the detection of epileptic spikes. However, due to the significant individual differences in the magnetoencephalogram patterns of epileptic seizures in different patients, even the discharge patterns of the same patient in different brain regions may show significant differences, which leads to the high dependence of traditional machine learning algorithms on the feature extraction of individual patients. In addition, the feature engineering process of such methods is complex and difficult to implement. At the same time, their generalization ability is obviously insufficient in cross-patient scenarios, which limits their potential for clinical promotion and application. Deep learning algorithms have shown excellent performance in spike detection with their powerful nonlinear modeling capabilities and automated feature extraction advantages. However, current deep learning methods generally fail to fully utilize the intrinsic correlations between multi-channel EEG signals, and lack comprehensive modeling of the dynamic characteristics of spike wave signal time series, which to some extent limits their performance and generalization capabilities in complex epileptic signal scenarios.

[0006] In summary, the practicality of the existing spike wave detection and localization methods based on signal analysis and traditional machine learning is not ideal, while the spike wave detection and localization methods based on deep learning algorithms fail to fully utilize the intrinsic correlation between multi-channel EEG signals, and lack comprehensive modeling of the dynamic characteristics of the spike wave signal time series. Summary of the invention

[0007] The purpose of the present invention is to provide a deep learning method for a magnetoencephalogram spike wave localization model to solve the problem that the spike wave detection and localization method based on deep learning algorithm in the prior art fails to fully utilize the intrinsic correlation between multi-channel magnetoencephalogram signals and lacks comprehensive modeling of the dynamic characteristics of the spike wave signal time series.

[0008] In order to solve the above technical problems, the present invention provides a deep learning method for a magnetoencephalogram (MEG) spike wave localization model, which is used to train a MEG spike wave localization model. The MEG spike wave localization model is used to determine whether a spike wave currently exists based on input MEG data.

[0009] The deep learning method includes: performing data preprocessing on multi-channel magnetoencephalogram signals; performing preliminary screening of spike waves, and performing regional multi-channel weighted fusion of the screened multi-channel magnetoencephalogram signals to obtain band data; calculating manual feature information of the band data; and inputting the band data and the corresponding manual feature information as a whole into a neural network to train and obtain the magnetoencephalogram spike wave localization model.

[0010] Optionally, the step of performing data preprocessing on the multi-channel magnetoencephalogram signals includes: performing bandpass filtering and notch processing on the collected multi-channel magnetoencephalogram signals; removing noise peaks by notch filtering or frequency domain interpolation filtering based on the signal spectrum of the multi-channel magnetoencephalogram signals; eliminating noise information caused by empty rooms in the multi-channel magnetoencephalogram signals based on a signal space projection algorithm; and removing artifact signals in the multi-channel magnetoencephalogram signals based on an independent component analysis algorithm.

[0011] Optionally, the step of performing preliminary screening of spike waves includes: dividing the multi-channel EEG signals into n1 areas based on the acquisition area of ​​the multi-channel EEG signals; for the data of each area, dividing them into n2 channels based on the acquisition channel of the multi-channel EEG signals; for the data of each channel, performing preliminary screening based on the relative amplitude of each band, and eliminating the bands whose relative amplitude exceeds the preset range; performing further screening based on characteristic parameters, and eliminating any band whose characteristic parameters exceed the corresponding screening threshold; wherein the characteristic parameters include: left half-wave slope, right half-wave slope and duration.

[0012] Optionally, the screening threshold corresponding to the characteristic parameter is determined by the upper and lower limits of the confidence interval calculated by the Monte Carlo method using the annotated spike database.

[0013] Optionally, the step of performing regional multi-channel weighted fusion of the screened multi-channel magnetoencephalogram signals to obtain band data includes: performing statistical analysis of time information on the screened multi-channel magnetoencephalogram signals, if the time difference between any two candidate samples is less than a preset difference, they are regarded as source discharge events of the same spike discharge; otherwise, they are regarded as different source discharge events; and, in each region, based on the weight information corresponding to each channel, weighted summing the bands corresponding to n2 channels to obtain single-channel synthetic data of the candidate samples.

[0014] Optionally, the manual feature information includes at least one of the following information: variance, average amplitude, energy, root mean square value, kurtosis factor, peak factor and shape factor.

[0015] Optionally, the step of inputting the band data and the corresponding manual feature information as a whole into the neural network includes: each band data is downsampled, sampled over a preset time span, and a preset number of sampling points are obtained; and the sampling points and the manual feature information are input into the neural network as a whole.

[0016] Optionally, the structure of the neural network is, from input to output, input layer, first fully connected layer, first gated recurrent unit layer, attention calculation layer, second fully connected layer, second gated recurrent unit layer, third fully connected layer, and output layer.

[0017] Optionally, in the first gated recurrent unit layer, each gated neural unit includes: a query, a key, and a value, and the i-th output x of the attention calculation layer i Calculated by the following formula:

[0018] in, α j represents the jth attention weight, v j represents the value of the jth gated neural unit.

[0019] Among them, e i and e j Represent the output features of the i-th gated neural unit and the j-th gated neural unit respectively; W k and W q is a trainable matrix used to transform e i and e j Mapped to a vector of keys and a vector of queries respectively.

[0020] The result x of the attention calculation layer i After being connected to the output of the first gated recurrent unit layer, it serves as the input of the subsequent fully connected layer.

[0021] Optionally, the loss function during training is calculated based on the following formula:

[0022] in, y i For sample i The label of the positive class is 1 and the negative class is 0; p i For sample i The probability of being predicted as the positive class, N is the number of samples.

[0023] Compared with the prior art, the present invention provides a deep learning method for a magnetoencephalogram spike wave localization model, wherein the deep learning method comprises: preprocessing data of multi-channel magnetoencephalogram signals; performing preliminary screening of spike waves, and performing regional multi-channel weighted fusion of the screened multi-channel magnetoencephalogram signals to obtain band data; calculating manual feature information of the band data; and, inputting the band data and the corresponding manual feature information as a whole into a neural network to train the magnetoencephalogram spike wave localization model. With such a configuration, the intrinsic correlation between multi-channel magnetoencephalogram signals and the dynamic characteristics of the spike wave signal time series are taken into account, thereby overcoming the problems existing in the prior art. In one embodiment of the present invention, the magnetoencephalogram channels are first partitioned according to different brain regions, thereby enhancing the localization accuracy of epilepsy signals in different brain regions. Then, a multi-channel weighted fusion strategy is adopted in each partition to effectively improve the computational efficiency and the robustness of the model. Finally, a neural network architecture combining gated recurrent units and attention mechanisms was constructed, in which the gated neural units are responsible for capturing the dynamic changes of previous and subsequent temporal information, while the attention mechanism is used to analyze the interactive relationship between band signals in different regions at the same time, thereby achieving high-precision detection of spike signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Those skilled in the art will appreciate that the accompanying drawings are provided for a better understanding of the present invention and do not constitute any limitation on the scope of the present invention. Figure 1 It is a flowchart of a deep learning method according to an embodiment of the present invention.

[0025] Figure 2 It is a structural diagram of a neural network model according to an embodiment of the present invention.

[0026] Figure 3 4 is a logic diagram of a spike wave positioning method according to an embodiment of the present invention.

[0027] in: 1-sampling point; 2-the manual feature information; 3-the first fully connected layer; 4-the first gated recurrent unit layer; 5-attention calculation layer; 6-the second fully connected layer; 7-the second gated recurrent unit layer; 8-the third fully connected layer; 9-output layer. DETAILED DESCRIPTION

[0028] In order to make the purpose, advantages and features of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the drawings are all in a very simplified form and are not drawn to scale, and are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In addition, the structure shown in the drawings is often a part of the actual structure. In particular, the emphasis of each drawing is different, and sometimes different scales are used.

[0029] As used in the present invention, the singular forms "one", "an" and "the" include plural objects, the term "or" is generally used to include the meaning of "and / or", the term "several" is generally used to include the meaning of "at least one", and the term "at least two" is generally used to include the meaning of "two or more". In addition, the terms "first", "second" and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first", "second" and "third" may explicitly or implicitly include one or at least two of the features, "one end" and "the other end" and "the proximal end" and "the distal end" generally refer to two corresponding parts, which include not only the endpoints, and the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral body; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. In addition, as used in the present invention, an element disposed on another element generally only indicates that there is a connection, coupling, cooperation or transmission relationship between the two elements, and the connection, coupling, cooperation or transmission between the two elements may be direct or indirect through an intermediate element, and cannot be understood as indicating or implying the spatial position relationship between the two elements, that is, one element may be in any orientation such as inside, outside, above, below or on one side of another element, unless otherwise clearly indicated in the content. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0030] The core idea of ​​the present invention is to provide a deep learning method for the magnetoencephalogram spike wave localization model to solve the problem that the spike wave detection and localization method based on deep learning algorithm in the prior art fails to fully utilize the intrinsic correlation between multi-channel magnetoencephalogram signals and lacks comprehensive modeling of the dynamic characteristics of the spike wave signal time series.

[0031] The following description is given with reference to the accompanying drawings.

[0032] This embodiment provides a deep learning method for a magnetoencephalogram (MEG) spike wave localization model, which is used to train a MEG spike wave localization model. The MEG spike wave localization model is used to determine whether there is a spike wave currently based on input MEG data.

[0033] Please refer to Figure 1 , the deep learning method comprises: S10, performing data preprocessing on multi-channel magnetoencephalogram signals.

[0034] S20, performing preliminary spike wave screening, and performing regional multi-channel weighted fusion on the screened multi-channel EEG signals to obtain band data.

[0035] S30, calculating manual feature information of the band data.

[0036] And, S40, the band data and the corresponding manual feature information are input as a whole into a neural network to train and obtain the magnetoencephalogram spike wave localization model.

[0037] Finally, the trained magnetoencephalogram spike wave localization model is used for spike wave discrimination and localization.

[0038] Specifically, step S10, the step of performing data preprocessing on the multi-channel magnetoencephalogram signals includes: S11, performing bandpass filtering and notch processing on the collected multi-channel EEG signals; performing basic bandpass filtering and 50 Hz notch processing on the collected EEG data to preliminarily eliminate the interference of geomagnetic field and power frequency noise in the environment.

[0039] S12, performing notch filtering or frequency domain interpolation filtering to remove noise peaks based on the signal spectrum of the multi-channel magnetoencephalogram signal; S13, based on a signal space projection algorithm, eliminating the noise information brought by the empty room in the multi-channel electroencephalographic signal.

[0040] And, S14, based on the independent component analysis algorithm, remove the artifact signals in the multi-channel EEG signal; for example, artifact signals such as slow drift, eye movement, heartbeat, etc.

[0041] In step S20, the step of performing preliminary spike wave screening includes: S21, dividing the multi-channel EEG signal into n1 regions based on the acquisition region of the multi-channel EEG signal, for example, into 8 regions of the frontal lobe, parietal lobe, temporal lobe and occipital lobe of the left and right hemispheres of the brain.

[0042] S22, for each region of data, based on the acquisition channel of the multi-channel EEG signal, each is divided into n2 channels. In a specific embodiment, the number of n2 is 8. That is, the total amount of data collected in a single time is 8×8.

[0043] S23, for the data of each channel, preliminary screening is performed based on the relative amplitude of each band, and the bands whose relative amplitude exceeds the preset range are eliminated. This step is intended to quickly eliminate abnormal data and reduce the interference of invalid samples.

[0044] And, S24, further screening is performed based on the characteristic parameters to eliminate any band whose characteristic parameters exceed the corresponding screening threshold; wherein the characteristic parameters include: left half-wave slope, right half-wave slope and duration.

[0045] The screening thresholds corresponding to the characteristic parameters are determined by the upper and lower limits of the confidence interval calculated by the Monte Carlo method using the annotated spike database. These screening thresholds serve as hyperparameters and provide a reliable reference for the screening process.

[0046] In step S20, the step of performing regional multi-channel weighted fusion of the screened multi-channel EEG signals to obtain band data includes: S25, performing time information statistical analysis on the screened multi-channel EEG signals, if the time difference between any two candidate samples is less than a preset difference (for example, the preset difference is 0.02 seconds), they are regarded as source discharge events of the same spike discharge; otherwise, they are regarded as different source discharge events. At this time, data of 8×8×n dimensions can be obtained, where n is the number of data corresponding to each channel that is initially judged to be suitable for training.

[0047] And, S26, in each region, based on the weight information corresponding to each channel, the bands corresponding to the n2 channels are weighted and summed to obtain the single-channel synthetic data of the candidate sample. The weighted weight is determined by the importance and relevance of the channel features to ensure that the fusion result can better represent the overall characteristics of the multi-channel signal. The dimension of the final synthetic data is 8×n.

[0048] For the obtained regional weighted fusion band signals, the main features are extracted and redundant information is eliminated to assist the subsequent deep neural network in the spike discrimination task. Specifically, the waveform is analyzed and quantified in the time domain to extract the following core indicators. That is, the manual feature information includes at least one of the following information: variance, average amplitude, energy, root mean square value, kurtosis factor, peak factor and shape factor.

[0049] 1). Variance: used to measure the degree of signal fluctuation and reflect the discreteness of signal amplitude distribution. The calculation formula is as follows:

[0050] Among them, X i Represents the signal value at the i-th sampling point, N represents the total number of sampling points contained in the sample, and μ represents the mean value of the band.

[0051] 2) Average amplitude: represents the overall average level of signal amplitude and provides an intuitive measure of signal energy strength. The calculation formula is as follows:

[0052] 3). Energy: The total energy characteristic is reflected by calculating the square sum of the signal. The calculation formula is as follows:

[0053] 4). Root mean square value: It comprehensively reflects the overall strength and fluctuation level of the signal. The calculation formula is as follows:

[0054] 5). Kurtosis factor: It is used to describe the sharpness of the signal amplitude distribution and can characterize the non-Gaussian characteristics of the signal. The calculation formula is as follows: E[(X i -μ) 4 ] / σ 4 Here, σ represents the standard deviation and E(·) represents the expectation.

[0055] 6). Peak factor: It indicates the ratio of the maximum value of the signal to the root mean square value, reflecting the local transient characteristics of the signal. The calculation formula is as follows:

[0056] 7). Shape factor: It is defined as the ratio of the signal root mean square value to the absolute average value, reflecting the complexity of the signal waveform. The calculation formula is as follows:

[0057] Through the extraction and analysis of the above features, redundant information can be effectively reduced and important features related to spikes can be highlighted. This process provides a solid foundation for subsequent classification and decision-making of deep neural networks, while ensuring that the feature input during model training is efficient and discriminative.

[0058] Please refer to Figure 2 The structure of the neural network from input to output is: input layer 10, first fully connected layer 3, first gated recurrent unit layer 4, attention calculation layer 5, second fully connected layer 6, second gated recurrent unit layer 7, third fully connected layer 8, and output layer 9.

[0059] In step S40, optionally, the step of inputting the band data and the corresponding manual feature information as a whole into the neural network includes: each band data is downsampled, sampled over a preset time span, and a preset number of sampling points 1 are obtained; and the sampling points 1 and the manual feature information 2 are input into the neural network as a whole.

[0060] In a specific embodiment, the sampling rate is uniformly adjusted to 250 Hz. The data of each band includes 50 sampling points, and the corresponding time span is 0.2 seconds.

[0061] The structure of the neural network is specifically described as follows: Input layer: The input layer dimension of the network is 8×57, where 8 represents 8 brain regions. The input of each brain region consists of signal bands (50 data in total) and manual feature information (7 data in total), totaling 57 features. This feature selection is designed to ensure multi-level representation of the signal and provide sufficient information support for subsequent network modules.

[0062] First fully connected layer: The output dimension of the first fully connected layer is 8×n3, where n3>=64, and a high-level abstraction of features is achieved through linear transformation and nonlinear activation function.

[0063] Gated recurrent unit layer: The second layer is designed as a gated recurrent unit, whose core function is to capture the temporal dependencies between different bands within a single brain region. By efficiently modeling the dynamic changes of time series, the gated recurrent unit layer can extract key temporal features and then pass them to the subsequent attention mechanism layer to achieve further processing and aggregation of cross-brain information.

[0064] The attention layer aims to model the interaction between different brain regions. In the first gated recurrent unit layer, each gated neural unit includes: query, key and value. The i-th output xi of the attention calculation layer is calculated by the following formula:

[0065] in, α j represents the jth attention weight, v j represents the value of the jth gated neural unit;

[0066] Among them, e i and e j Represent the output features of the i-th gated neural unit and the j-th gated neural unit respectively; W k and W q is a trainable matrix used to transform e i and e j Mapped to a vector of keys and a vector of queries respectively.

[0067] The result x of the attention calculation layer i After being connected to the output of the first gated recurrent unit layer, it serves as the input of the subsequent fully connected layer.

[0068] Subsequent fully connected layers and output layers: The output feature dimension of the attention layer is 8×n4, where n4>=128. After further processing by the second gated recurrent neural unit layer, these features are passed to the final fully connected layer, whose output dimension is 8×n5, where n5>=64. Finally, the output layer completes the classification and discrimination task of spikes based on the extracted high-dimensional features, ensuring accurate classification and interpretation of the input signal.

[0069] To ensure the efficiency of the training process and the discrimination accuracy of the model, the network's 8 brain region signal inputs need to be synchronized. This means that at any time, the network input is composed of the EEG signals of the 8 brain regions at that time. During the training process, in order to effectively capture the specific characteristic performance of each brain region, the sub-network of each brain region is trained independently, so it is necessary to define a corresponding loss function for each sub-network. In order to improve the training speed and model convergence effect, this paper adopts the cross entropy loss function, which is defined as follows:

[0070] in, y i For sample i The label of the positive class is 1 and the negative class is 0; p i For sample i The probability of being predicted as the positive class, N is the number of samples.

[0071] Write a spike position determination method, such as Figure 3, input the data to be judged into the trained spike discriminant network. At the same time, the peak point coordinates corresponding to the band and the output results of the spike discriminant network are used as the spike position judgment method. When the output results of the discriminant network contain a probability value greater than 0.9, the peak point coordinates are output, otherwise 0 is output.

[0072] This embodiment performs a preliminary screening of each frequency band of the magnetoencephalogram data and performs regional weighted fusion based on the preliminary screening results, which can significantly reduce the complexity of subsequent processing while retaining the differences between different regions. Subsequently, the main features of the weighted fused data are extracted to provide strong auxiliary support for spike wave discrimination.

[0073] This embodiment constructs a special deep neural network architecture, which can effectively make joint judgments on bands in different regions at the same time, and comprehensively consider the impact of the precursor band on the current band, so as to make more accurate auxiliary judgments on spikes.

[0074] In summary, this embodiment provides a deep learning method for a magnetoencephalogram spike wave localization model. The deep learning method includes: preprocessing data of multi-channel magnetoencephalogram signals; performing preliminary screening of spike waves, and performing regional multi-channel weighted fusion of the screened multi-channel magnetoencephalogram signals to obtain band data; calculating manual feature information of the band data; and, inputting the band data and the corresponding manual feature information as a whole into a neural network to train the magnetoencephalogram spike wave localization model. Such a configuration takes into account the intrinsic correlation between multi-channel magnetoencephalogram signals and the dynamic characteristics of the spike wave signal time series, thereby overcoming the problems existing in the prior art. In one embodiment of the present invention, the magnetoencephalogram channels are first partitioned according to different brain regions, thereby enhancing the localization accuracy of epilepsy signals in different brain regions. Then, a multi-channel weighted fusion strategy is adopted in each partition to effectively improve the computational efficiency and the robustness of the model. Finally, a neural network architecture combining gated recurrent units and attention mechanisms was constructed, in which the gated neural units are responsible for capturing the dynamic changes of previous and subsequent temporal information, while the attention mechanism is used to analyze the interactive relationship between band signals in different regions at the same time, thereby achieving high-precision detection of spike signals.

[0075] The above description is only a description of the preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any changes or modifications made by a person skilled in the art in the field of the present invention based on the above disclosure shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A deep learning method for magnetoencephalogram spike wave localization model, characterized in that: Used to train a magnetoencephalogram spike wave localization model, the magnetoencephalogram spike wave localization model is used to determine whether there is a spike wave based on the input magnetoencephalogram data; the deep learning method includes: Perform data preprocessing on multi-channel EEG signals; Performing a preliminary screening of spike waves, and performing regional multi-channel weighted fusion of the screened multi-channel EEG signals to obtain band data; Calculating manual feature information of the band data; and, The band data and the corresponding manual feature information are input as a whole into a neural network to train and obtain the magnetoencephalogram spike wave localization model.

2. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 1, characterized in that: The step of performing data preprocessing on the multi-channel magnetoencephalogram signals comprises: Performing bandpass filtering and notch processing on the collected multi-channel EEG signals; Based on the signal spectrum of the multi-channel EEG signal, performing notch filtering or frequency domain interpolation filtering to remove noise peaks; Based on the signal space projection algorithm, the noise information brought by the empty room in the multi-channel electroencephalographic signal is eliminated; and Based on the independent component analysis algorithm, the artifact signal in the multi-channel brain magnetic signal is removed.

3. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 1, characterized in that: The steps of performing preliminary spike wave screening include: Based on the acquisition area of ​​the multi-channel EEG signal, the multi-channel EEG signal is divided into n1 areas; The data of each region are divided into n2 channels based on the acquisition channels of the multi-channel EEG signals; For the data of each channel, preliminary screening is performed based on the relative amplitude of each band, and the bands whose relative amplitude exceeds the preset range are eliminated; and, Further screening is performed based on the characteristic parameters to eliminate any band whose characteristic parameters exceed the corresponding screening threshold; wherein the characteristic parameters include: left half-wave slope, right half-wave slope and duration.

4. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 3, characterized in that: The screening threshold corresponding to the characteristic parameter is determined by the upper and lower limits of the confidence interval calculated by the Monte Carlo method using the annotated spike database.

5. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 3, characterized in that: The step of performing regional multi-channel weighted fusion on the screened multi-channel EEG signals to obtain band data comprises: Performing a statistical analysis of the time information of the screened multi-channel EEG signals, if the time difference between any two candidate samples is less than a preset difference, they are regarded as source discharge events of the same spike discharge; otherwise, they are regarded as different source discharge events; and, In each region, based on the weight information corresponding to each channel, the bands corresponding to the n2 channels are weighted and summed to obtain the single-channel synthetic data of the candidate sample.

6. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 1, characterized in that: The manual feature information includes at least one of the following information: variance, average amplitude, energy, root mean square value, kurtosis factor, peak factor and shape factor.

7. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 1, characterized in that: The step of inputting the band data and the corresponding manual feature information as a whole into the neural network comprises: Each of the band data is downsampled, sampled over a preset time span, and a preset number of sampling points are obtained; and, The sampling points and the manual feature information are input as a whole into the neural network.

8. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 1, characterized in that: The structure of the neural network from input to output is: input layer, first fully connected layer, first gated recurrent unit layer, attention calculation layer, second fully connected layer, second gated recurrent unit layer, third fully connected layer, and output layer.

9. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 1, characterized in that: In the first gated recurrent unit layer, each gated neural unit includes: query, key and value. The i-th output xi of the attention calculation layer is calculated by the following formula: in, α j represents the jth attention weight, v j represents the value of the jth gated neural unit; Among them, e i and e j Represent the output features of the i-th gated neural unit and the j-th gated neural unit respectively; W k and W q is a trainable matrix used to transform e i and e j Mapped to key vector and query vector respectively; The result x of the attention calculation layer i After being connected to the output of the first gated recurrent unit layer, it serves as the input of the subsequent fully connected layer.

10. The deep learning method of the magnetoencephalogram spike wave localization model according to claim 1, characterized in that: The loss function during training is calculated based on the following formula: in, y i For sample i The label of the positive class is 1 and the negative class is 0; p i For sample i The probability of being predicted as the positive class, N is the number of samples.

Images