A Method and System for Detecting and Localizing EEG Abnormalities Based on Grouped Cosine Convolutional Networks

Abstract

This invention discloses a method and system for detecting and locating EEG abnormalities based on a grouped cosine convolutional network, belonging to the field of EEG abnormality detection technology. The method includes: acquiring multi-channel EEG signals and preprocessing them; inputting the multi-channel EEG signals into a grouped cosine convolutional network, grouping the multi-channel EEG signals by channel, and performing cosine convolution operations on each group of EEG signals; extracting multiple groups of EEG features after multiple grouped cosine convolution operations; concatenating the multiple groups of EEG features; and outputting the classification result and classification probability distribution of EEG abnormality detection based on the concatenated features; triggering an alarm for abnormal EEG based on the classification probability distribution of EEG abnormality detection and a preset threshold; calculating the contribution score of each channel of EEG signal based on the classification result of EEG abnormality detection, determining the channel of the abnormal EEG signal, and completing the abnormal EEG localization. This invention can achieve end-to-end EEG abnormality detection and real-time channel localization, with effective accuracy and real-time performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electroencephalogram anomaly detection, and particularly relates to a method and system for electroencephalogram anomaly detection and positioning based on a group cosine convolutional neural network. BACKGROUND

[0002] Electroencephalogram (EEG) is an important tool for recording brain electrical activity. It detects and records brain electrical activity by placing electrodes on the scalp, and is widely used in clinical diagnosis, neuroscience research and other brain activities. Electroencephalogram is not only an important tool for diagnosing certain brain diseases, but also provides effective treatment for certain brain diseases, and plays an important role in the detection of diseases caused by abnormal brain discharges.

[0003] Traditional electroencephalogram anomaly detection mainly relies on manual visual detection, which is not only time-consuming and laborious, but also low in efficiency and easily affected by human factors, resulting in strong subjectivity and low detection accuracy. With the development of deep learning technology, deep learning-based electroencephalogram anomaly detection models show high detection accuracy and automation level. However, the existing deep learning detection model still has certain drawbacks. On the one hand, the existing model is complex in structure and has a large number of model parameters, which is slow in training efficiency, large in calculation amount and high in calculation cost, and the detection accuracy of this model needs to be further improved. On the other hand, the existing model usually lacks analysis and processing of spatial information related to the electroencephalogram channel, which further limits its application effect in electroencephalogram channel positioning. SUMMARY

[0004] To solve the above problems of the prior art, the present application provides a method and system for electroencephalogram anomaly detection and positioning based on a group cosine convolutional neural network, which adopts a hardware-friendly and memory-efficient cosine convolution operator to build an electroencephalogram anomaly detection model based on a group cosine convolutional neural network (Group CosCNN). The model realizes end-to-end electroencephalogram anomaly detection and real-time channel positioning, which can effectively improve the accuracy and real-time performance of electroencephalogram anomaly detection and positioning.

[0005] In a first aspect, the present application provides a method for electroencephalogram anomaly detection and positioning based on a group cosine convolutional neural network.

[0006] A method for electroencephalogram anomaly detection and positioning based on a group cosine convolutional neural network, comprising:

[0007] Collecting multi-channel electroencephalogram signals and pre-processing the collected signals;

[0008] The preprocessed multi-channel electroencephalogram signal is input into a grouped cosine convolution network, the multi-channel electroencephalogram signal is grouped according to channels, and a cosine convolution operation is performed on each group of electroencephalogram signals, a plurality of groups of electroencephalogram features are extracted after a plurality of times of grouped cosine convolution operation, the plurality of groups of electroencephalogram features are spliced, and a classification result and a classification probability distribution of electroencephalogram anomaly detection are output based on the spliced features;

[0009] According to the classification probability distribution of the electroencephalogram anomaly detection and the preset threshold, an abnormal electroencephalogram trigger alarm is performed.

[0010] According to the classification result of the electroencephalogram anomaly detection, the contribution score of each channel electroencephalogram signal is calculated, the channel of the abnormal electroencephalogram signal is determined, and the abnormal electroencephalogram positioning is completed.

[0011] In a second aspect, the present application provides an electroencephalogram anomaly detection and positioning system based on a grouped cosine convolution network.

[0012] An electroencephalogram anomaly detection and positioning system based on a grouped cosine convolution network comprises:

[0013] A data acquisition module is configured to acquire multi-channel electroencephalogram signals and pre-process the acquired signals.

[0014] An abnormal electroencephalogram detection module is configured to input the pre-processed multi-channel electroencephalogram signal into a grouped cosine convolution network, group the multi-channel electroencephalogram signal according to channels, and perform a cosine convolution operation on each group of electroencephalogram signals, extract a plurality of groups of electroencephalogram features after a plurality of times of grouped cosine convolution operation, splice the plurality of groups of electroencephalogram features, and output a classification result and a classification probability distribution of electroencephalogram anomaly detection based on the spliced features.

[0015] An abnormal electroencephalogram alarm module is configured to perform an abnormal electroencephalogram trigger alarm according to the classification probability distribution of the electroencephalogram anomaly detection and the preset threshold.

[0016] An abnormal electroencephalogram channel positioning module is configured to calculate the contribution score of each channel electroencephalogram signal according to the classification result of the electroencephalogram anomaly detection, determine the channel of the abnormal electroencephalogram signal, and complete the abnormal electroencephalogram positioning.

[0017] In a third aspect, the present application further provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein when the computer instructions are run by the processor, the steps of the method of the first aspect are completed.

[0018] In a fourth aspect, the present application further provides a computer readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the steps of the method of the first aspect are completed.

[0019] The above one or more technical solutions have the following beneficial effects:

[0020] 1. The application provides a brain electrical abnormality detection and positioning method and system based on a grouped cosine convolutional network, which adopts a hardware-friendly and memory-efficient cosine convolution operator to build a brain electrical abnormality detection model based on a grouped cosine convolutional neural network, wherein a channel convolution layer is used to perform a cosine convolution operation on each channel of the input multi-channel electroencephalogram signal, extract the electroencephalogram features of each channel, and perform multiple grouped convolution operations through a multi-layer grouped convolution layer, i.e., grouping the input multi-channel electroencephalogram features by channel and performing a cosine convolution operation on each group of electroencephalogram features, so as to extract deep-level electroencephalogram features while retaining the channel-specific information in each layer of the network, thereby achieving brain electrical abnormality detection and end-to-end real-time channel positioning.

[0021] 2. In the application, the convolution kernels in the multi-layer convolutional layers in the grouped cosine convolutional neural network are in the form of a cosine function, only having two learnable parameters, i.e., amplitude and frequency, which can effectively reduce the number of parameters, thereby improving the compactness and hardware friendliness of the model, improving the training efficiency of the model, and effectively improving the accuracy and real-time performance of brain electrical abnormality detection and positioning. BRIEF DESCRIPTION OF DRAWINGS

[0022] The drawings accompanying the specification of the application form part of the application and serve to provide a further understanding of the application, the illustrative embodiments thereof, and their description serve to explain the application without imposing undue limitations on the application.

[0023] Figure 1 The overall flowchart of the brain electrical abnormality detection and positioning method based on a grouped cosine convolutional network described in the embodiments of the application;

[0024] Figure 2 The structural schematic diagram of the grouped cosine convolutional network described in the embodiments of the application;

[0025] Figure 3 The topographic schematic diagram of abnormal electroencephalogram channel positioning in the embodiments of the application;

[0026] Figure 4 The schematic diagram of the brain electrical abnormality detection and positioning system based on a grouped cosine convolutional network described in the embodiments of the application. DETAILED DESCRIPTION

[0027] It should be noted that the following detailed description is exemplary only and is intended to provide further description of the exemplary embodiments according to the present application, and is not intended to limit the exemplary embodiments according to the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. Furthermore, it should be understood that the use of the terms "including", "comprising", "having" and "involving" herein are used to mean that the specified features, steps, operations, devices, components and / or combinations thereof are included, but not limited to, the features, steps, operations, devices, components and / or combinations thereof.

[0028] Embodiment One

[0029] The present embodiment provides a method for detecting and locating abnormal electroencephalogram (EEG) based on a grouped cosine convolution network, which can efficiently detect abnormal activities in EEG signals and accurately locate abnormal EEG channels, thereby realizing real-time abnormal EEG alarm and positioning. Figure 1 As shown in the figure, the method specifically includes the following steps:

[0030] Step S1, collecting multi-channel EEG signals and pre-processing the collected signals;

[0031] Step S2, inputting the pre-processed multi-channel EEG signals into a grouped cosine convolution network, grouping the multi-channel EEG signals by channel, and performing a cosine convolution operation on each group of EEG signals, extracting multiple groups of EEG features after multiple grouped cosine convolution operations, concatenating the multiple groups of EEG features, and outputting the classification results and classification probability distribution of EEG abnormality detection based on the concatenated features;

[0032] Step S3, performing abnormal EEG trigger alarm according to the classification probability distribution of EEG abnormality detection and a preset threshold;

[0033] Step S4, calculating the contribution score of each channel EEG signal according to the classification results of EEG abnormality detection, determining the channel of abnormal EEG signal, and completing abnormal EEG positioning.

[0034] The method for detecting and locating abnormal EEG based on a grouped cosine convolution network proposed in the present embodiment will be described in more detail in the following.

[0035] In step S1, a standard electroencephalogram (EEG) acquisition device is used, which includes an EEG electrode and an amplifier, to collect multi-channel EEG signals and pre-process the collected EEG signals. The EEG analog signal is obtained through the EEG electrode, and the EEG analog signal is converted into a digital EEG signal. The digital EEG signal is further enhanced in strength by a signal amplifier to make it suitable for computer processing.

[0036] Further, to realize the analysis and processing of spatial information related to the brain electrical channel, in the embodiment, a plurality of brain electrical electrodes are used to synchronously collect brain electrical signals of a plurality of different regions of the brain, each brain electrical electrode collects and records the brain electrical signals of one channel, thereby obtaining multi-channel brain electrical signals, which can provide electrical activity information of different regions of the brain; and then the obtained multi-channel brain electrical signals are digitized and amplified for subsequent processing.

[0037] In step S2, the preprocessed multi-channel brain electrical signals are input into the brain electrical abnormality detection model based on the grouped cosine convolution network (which is deployed on a computer device), the grouped cosine convolution network is used to process and analyze the multi-channel brain electrical signals, extract brain electrical signal features (referred to as brain electrical features) and detect abnormal brain electrical activity, and finally output the classification results and classification probability distribution of brain electrical abnormality detection.

[0038] The grouped cosine convolution network proposed in the embodiment is a neural network architecture, which includes a plurality of convolution layers, such as Figure 2 As shown in the figure, it specifically includes an input layer, a channel convolution layer, a plurality of grouped convolution layers and an output layer connected in sequence, and the output layer includes a fully connected layer and a Softmax layer; wherein each convolution operation in the channel convolution layer and the plurality of grouped convolution layers adopts a cosine convolution kernel, and the local features of the input signal are extracted through cosine convolution. Through the above grouped cosine convolution network, the input brain electrical features are divided into a plurality of groups, and then cosine convolution operation is performed on each group of features to retain channel-specific information and lay a foundation for subsequent channel positioning.

[0039] Further, the above-mentioned channel convolution layer includes a plurality of cosine convolution kernels and a pooling layer, which is used to perform cosine convolution operation on each channel of the input multi-channel brain electrical signals to extract the brain electrical features of each channel; wherein the brain electrical features are frequency features of the brain electrical signals. In this layer, the convolution kernel adopts a cosine convolution kernel (the cosine convolution kernel is a convolution kernel in the form of a cosine function), and the convolution kernel is a parameter matrix in the convolution layer, which is used to extract the local features of the input signal. In addition, the cosine convolution kernel only has two learnable parameters of amplitude and frequency, which can effectively reduce the number of parameters and improve the compactness and hardware friendliness of the model.

[0040] wherein the convolution operation formula of each convolution kernel in the channel convolution of the lth layer is:

[0041]

[0042] In the above formula, and respectively represent the amplitude and frequency parameters of the cth channel convolution kernel of the lth layer, K is the size of the convolution kernel, is the output feature of the (i+m)-th channel position of the c-th group in the l-th layer. In this embodiment, only one channel convolution layer is set, i.e., l = 1.

[0043] Further, each of the above-mentioned group convolution layers includes a plurality of cosine convolution kernels and a pooling layer, which are used to group the input multi-channel electroencephalogram features by channels and perform cosine convolution operation on each group of electroencephalogram features, so as to extract deep-level electroencephalogram signal features while retaining channel-specific information in each network layer. Similarly, in each group convolution layer, the convolution kernel also adopts a cosine convolution kernel.

[0044] Suppose the number of channels of the electroencephalogram signal is C (in this embodiment, the number of channels C = 18), each group convolution layer divides the input features into C groups, each group containing C' / G channel features, where C' is the output channel number of the previous layer, and G is the number of groups, which is equal to the number of channels C.

[0045] The multi-layer group cosine convolution is performed through the multi-layer group convolution layer, wherein the convolution operation formula of each convolution kernel in the group convolution of the l-th layer is as follows:

[0046]

[0047] In the above formula, represents the output of the i-th position of the g-th group in the l-th layer, and are the amplitude and frequency parameters of the c'-th channel convolution kernel of the g-th group in the l-th layer, and g' = c' + (g-1)(C' / G) is the position index of the c'-th channel of the g-th group in the previous layer.

[0048] Through the above design, the group cosine convolution network includes a plurality of convolution layers, the number of convolution kernels of each layer is N1, N2, N3, N4, and N5, respectively, and after the convolution operation of each layer, a shared batch normalization layer (BN layer) and a maximum pooling layer (pooling size is 2) are connected; the final layer splices the flattened features of each channel into a single feature vector and inputs it into a fully connected layer for classification.

[0049] Specifically, the multiple groups of electroencephalogram features output by the last group convolutional layer are input to the full connection layer and the Softmax layer. First, the input feature vector is mapped to the output vector through the weight matrix of the full connection layer (Dense Layer), i.e., the multiple groups of electroencephalogram features are spliced into a feature vector, and classification is performed based on the feature vector, and the classification result of electroencephalogram anomaly detection is output, i.e., whether the electroencephalogram signal is normal or abnormal category is output. Then, the output of the full connection layer is converted into a probability distribution through the Softmax layer (Softmax Layer), i.e., the classification result of abnormal electroencephalogram activity is converted into a probability distribution, and the classification probability distribution of electroencephalogram anomaly detection is output, i.e., the probability that the electroencephalogram signal is normal and abnormal categories is output, so as to determine whether there is abnormal electroencephalogram activity.

[0050] wherein it is assumed that is the flattened feature vector of the gth group, is the weight matrix of the full connection layer, is the bias vector, and the calculation formula of the abnormal electroencephalogram probability is:

[0051] P v =softmax(W · ·concat(Y g＝1 ,…,Y g＝G )+b)

[0052] In the above formula, concat(·) represents a feature splicing operation, and P v is the probability distribution of the classification output.

[0053] As another implementation, a ReLU activation function (Rectified Linear Unit, ReLU) can also be provided after each convolutional layer (including channel convolution kernel group convolution), and the contribution score of subsequent calculation is ensured to be a non-negative value through the ReLU function.

[0054] In this embodiment, taking the 18-channel electroencephalogram signal as an example, the specific structure of the designed group cosine convolutional network model is as follows:

[0055] (1) Input layer: the input size of the multi-channel electroencephalogram signal is 18x1024;

[0056] (2) 1st layer (channel convolutional layer): cosine convolution is performed on each channel of the 18 channels, and there are 16 convolution kernels in each channel, and the convolution kernel size K=23; the output size of the final layer is 288x512 (including a pooling layer, and the pooling size is 2), and the output channel is 16x18=288;

[0057] (3) The 2nd layer (grouped convolution layer): according to the output channel of the previous layer, it is divided into G = C = 18 groups, each group contains 16 channel features, cosine convolution is performed on each group, there are 16 convolution kernels in each group, and the convolution kernel size K = 23; the final output size of the layer is 288x256 (including a pooling layer, and the pooling size is 2), wherein for each group, the convolution input channel is 16, and the output channel is also 16, so the total output channel number of the final layer is actually unchanged, still 16x18 = 288;

[0058] (4) The 3rd layer (grouped convolution layer): using the above method, continue the grouped convolution operation, and cosine convolution is performed on each group, there are 16 convolution kernels in each group, and the convolution kernel size K = 23; the final output size of the layer is 288x128 (including a pooling layer, and the pooling size is 2);

[0059] (5) The 4th layer (grouped convolution layer): using the above method, continue the grouped convolution operation, and cosine convolution is performed on each group, there are 32 convolution kernels in each group, and the convolution kernel size K = 23; the final output size of the layer is 576x64 (including a pooling layer, and the pooling size is 2), wherein for each group, the convolution input channel is 16, and the output channel is 32, so the total output channel number of the final layer is 32x18 = 576;

[0060] (6) The 5th layer (grouped convolution layer): using the above method, continue the grouped convolution operation, and cosine convolution is performed on each group, there are 64 convolution kernels in each group, and the convolution kernel size K = 23; the final output size of the layer is 1152x64;

[0061] (7) The 6th layer (grouped convolution layer): using the above method, continue the grouped convolution operation, and cosine convolution is performed on each group, there are 64 convolution kernels in each group, and the convolution kernel size K = 23; the final output size of the layer is 1152x32 (including a pooling layer, and the pooling size is 2);

[0062] (8) The output layer: the features of each group are spliced to obtain a feature vector, and two neurons are output through a fully connected layer, respectively corresponding to an abnormal class and a normal class, and the final classification probability is output through a softmax layer.

[0063] As another implementation, the Adam optimization algorithm (Adam Optimizer) is used to train the above grouped cosine convolution network, so as to adaptively adjust the learning rate of each parameter, and improve the efficiency and stability of model training.

[0064] Further, in the network model training process, a loss function of the grouped cosine convolution network is constructed based on a cross-entropy loss and a regularization term, which can effectively improve the detection accuracy and prevent overfitting. The regularization term limits the excessive change of the parameters by imposing a penalty on the model parameters, thereby improving the generalization ability of the model. The specific formula of the constructed loss function is:

[0065]

[0066] In the above formula, M is the minimum batch size of training, Q=2 is the number of classification categories, γ i,j is an indicator function, σ is a regularization coefficient, and θ is a model parameter.

[0067] The iterative training of the grouped cosine convolution network is performed based on the above loss function until the training is completed. During the iterative training of the grouped cosine convolution network, the amplitude parameters and the frequency parameters of the cosine convolution kernel in each convolution layer of the network are iteratively updated using a back propagation algorithm, and the amplitude and frequency parameters of the grouped cosine convolution network are optimized using an Adam optimization algorithm. Specifically, the amplitude parameters and the frequency parameters of the convolution kernel are updated using the back propagation algorithm, wherein the gradient calculation formula of the amplitude parameters is:

[0068]

[0069] The gradient calculation formula of the frequency parameters is:

[0070]

[0071] In the above formula,

[0072] In step S3, an abnormal electroencephalogram triggering alarm is performed according to the classification probability distribution of the electroencephalogram anomaly detection and a preset threshold. Specifically, according to the abnormal category probability of the electroencephalogram anomaly detection obtained in step S2 and the preset threshold, an abnormal electroencephalogram triggering alarm is performed. When the abnormal category probability is greater than the preset threshold, the alarm mechanism is triggered to perform real-time alarm and reminder of the electroencephalogram anomaly. In this embodiment, the preset threshold is set to 0.5.

[0073] In step S4, the contribution score of each channel electroencephalogram signal is calculated according to the classification result of the electroencephalogram anomaly detection, and the channel of the abnormal electroencephalogram signal is determined to complete the abnormal electroencephalogram positioning.

[0074] Specifically, according to the weight matrix of the normal and abnormal categories in the classification result of the electroencephalogram anomaly detection, and in combination with the multiple groups of electroencephalogram features output by the last grouped convolution layer, the contribution score S cThe calculation formula is:

[0075]

[0076] In the above formula, ΔW c = W c,2 -W c,1 , W c,2 , W c,1 indicate the weight matrix corresponding to the output neurons of the abnormal and normal categories in the electroencephalogram anomaly detection model; Y c is the feature vector of the cth channel, S c is the contribution score of the cth channel.

[0077] Then, according to the size of the contribution score, the channel of the abnormal electroencephalogram signal is determined, that is, if S c > 0, the channel has a positive contribution to the detection of the abnormal electroencephalogram, and the greater the contribution score value, the greater the possibility of the corresponding channel electroencephalogram signal being abnormal.

[0078] In order to further verify the superiority of the method proposed in the embodiment, the patient data in the public CHB-MIT abnormal electroencephalogram dataset is used to comprehensively ablate different types of detection models, and the performance in the electroencephalogram anomaly detection is compared. Specifically, two traditional grouped convolutional neural network models are used, that is, a groupCNN-s model with a convolution kernel length of 3 and a group CNN-l model with a convolution kernel length of 23, and a grouped cosine convolution operation is used to replace the traditional grouped convolution operation in the groupCNN-l to construct the grouped cosine convolutional neural network (groupCosCNN) model proposed in the embodiment.

[0079] In the embodiment, the abnormal electroencephalogram data and the random normal electroencephalogram data of the same length in the public CHB-MIT database are selected for training and testing. Each data sample is an electroencephalogram signal collected from 18 channels, each channel has a sampling rate of 256 Hz, a recording time length of 4 seconds, and a total of 1024 data points; each model uses the Adam optimization algorithm to adaptively adjust the learning rate of each parameter, in the training process, the hyperparameters M and σ are set to 256 and 0.0005 respectively, and the learning rate is reduced from the initial 2×10 -4 to 2×10 -5 ; finally, five-fold cross-validation is used, and the average AUC (Area Under the Curve) is used as the performance indicator. The AUC is an indicator for evaluating the performance of a binary classification model, indicating the area under the receiver operating characteristic (ROC) curve. The AUC value is between 0.5 and 1, and the larger the value, the stronger the ability of the model to distinguish between positive and negative samples.

[0080] The final experimental results are shown in Table 1 below. The AUC of the group CNN-l model is 97.99%, which is higher than the AUC of the group CNN-s model, which is 95.67%, but the parameter amount is larger, which is 3.0M. The AUC of the grouping cosine convolutional neural network model proposed in the embodiment is 98.17%, which can further improve the AUC performance index, and the parameter amount is reduced to 0.34M.

[0081] Table 1: EEG abnormality detection performance indicators of different types of models

[0082] Model Convolution kernel length AUC (%) Parameter amount groupCNN-s 3 95.67±0.71 0.47M groupCNN-1 23 97.99±0.96 3.0M groupCosCNN 23 98.17±1.19 0.34M

[0083] Through the above examples, the grouping cosine convolutional neural network proposed in the embodiment can still achieve higher detection accuracy with reduced parameter amount. In addition, the method proposed in the embodiment can also determine the channel of the abnormal EEG signal and realize the positioning of the abnormal EEG. The contribution score of each channel calculated finally is converted into a topographic map for abnormal EEG channel positioning, as shown in Figure 3 The circle in the figure represents the brain, and the abnormal EEG region can be directly displayed, which is convenient for staff to view more intuitively. The embodiment has significant advantages and application potential in EEG abnormality detection and positioning.

[0084] Embodiment Two

[0085] The embodiment provides an EEG abnormality detection and positioning system based on a grouping cosine convolutional network, as shown in Figure 4 , and specifically includes:

[0086] A data acquisition module is configured to acquire multi-channel EEG signals and pre-process the acquired signals.

[0087] An abnormal EEG detection module is configured to input the pre-processed multi-channel EEG signals into the grouping cosine convolutional network, group the multi-channel EEG signals by channel, and perform cosine convolution operation on each group of EEG signals. After multiple grouping cosine convolution operations, multi-group EEG features are extracted, the multi-group EEG features are spliced, and the classification result and classification probability distribution of EEG abnormality detection are output based on the spliced features.

[0088] An abnormal EEG alarm module is configured to trigger an alarm according to the classification probability distribution of EEG abnormality detection and a preset threshold.

[0089] An abnormal EEG channel positioning module is configured to calculate the contribution score of each channel EEG signal according to the classification result of EEG abnormality detection, determine the channel of the abnormal EEG signal, and complete the positioning of the abnormal EEG.

[0090] Embodiment Three

[0091] The embodiment provides an electronic device, comprising a memory and a processor and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, steps in the electroencephalogram anomaly detection and positioning method based on a grouped cosine convolution network are completed.

[0092] Embodiment Four

[0093] The embodiment also provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, steps in the electroencephalogram anomaly detection and positioning method based on a grouped cosine convolution network are completed.

[0094] The steps and methods in the above embodiments two to four correspond to the method embodiment one, and the specific embodiments can be understood with reference to the related description part of the embodiment one. The term 'computer readable storage medium' should be understood as including a single medium or multiple media of one or more instruction sets; and should also be understood as including any medium capable of storing, encoding or carrying instruction sets for execution by a processor and causing the processor to perform any method in the present application.

[0095] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be respectively manufactured into each integrated circuit module, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0096] The above only describes the preferred embodiments of the present application, and the specific embodiments of the present application are described in combination with the drawings, but are not limited to the protection scope of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A method for electroencephalogram anomaly detection and localization based on grouped cosine convolutional network, characterized in that, The method comprises the following steps: Collecting multi-channel electroencephalogram signals and pre-processing the collected signals; Inputting the pre-processed multi-channel electroencephalogram signals into a grouped cosine convolution network, grouping the multi-channel electroencephalogram signals by channels, and performing cosine convolution operation on each group of electroencephalogram signals, extracting multiple groups of electroencephalogram features after multiple grouped cosine convolution operations, splicing the multiple groups of electroencephalogram features, and outputting the classification result and classification probability distribution of electroencephalogram anomaly detection based on the spliced features; According to the classification probability distribution of electroencephalogram anomaly detection and the preset threshold, an abnormal electroencephalogram trigger alarm is performed; According to the classification result of electroencephalogram anomaly detection, the contribution score of each channel electroencephalogram signal is calculated to determine the channel of the abnormal electroencephalogram signal and complete the abnormal electroencephalogram positioning.

2. The method of claim 1, wherein the method comprises: determining a plurality of feature maps based on the plurality of convolutional layers; and determining a plurality of feature maps based on the plurality of pooling layers. The method comprises the following steps: Synchronously collecting electroencephalogram signals from multiple different regions of the brain by using multiple electroencephalogram electrodes, collecting electroencephalogram signals of one channel by each electroencephalogram electrode, and obtaining multi-channel electroencephalogram signals in this way; Digitizing and amplifying the obtained multi-channel electroencephalogram signals.

3. The method for detecting and locating EEG abnormalities based on grouped cosine convolutional networks as described in claim 1, characterized in that, The grouped cosine convolution network comprises an input layer, a channel convolution layer, multiple layered grouped convolution layers and an output layer connected in sequence, and the output layer comprises a fully connected layer and a Softmax layer; wherein each convolution operation in the channel convolution layer and the multiple layered grouped convolution layers adopts a cosine convolution kernel to extract local features of the input signals by cosine convolution; The channel convolution layer comprises multiple cosine convolution kernels and a pooling layer, which is used for performing separate cosine convolution operation on each channel of the input multi-channel electroencephalogram signals to extract electroencephalogram features of each channel; the electroencephalogram features are frequency features of the electroencephalogram signals; Each layered grouped convolution layer comprises multiple cosine convolution kernels and a pooling layer, which is used for grouping the input multi-channel electroencephalogram features by channels and performing separate cosine convolution operation on each group of electroencephalogram features to extract deep features of the electroencephalogram signals.

4. The method of claim 3, wherein the method is characterized by, The multiple groups of electroencephalogram features output by the last layered grouped convolution layer are input into the fully connected layer and the Softmax layer, the multiple groups of electroencephalogram features are spliced into a feature vector by the fully connected layer, classification is performed based on the feature vector, and the classification result of electroencephalogram anomaly detection is output, i.e., whether the electroencephalogram signal is normal or abnormal; the classification result of abnormal electroencephalogram activity is converted into a probability distribution by the Softmax layer, and the classification probability distribution of electroencephalogram anomaly detection is output, i.e., the probability of the electroencephalogram signal being normal and abnormal.

5. The method of claim 3, wherein the method comprises: calculating a plurality of cosine similarity values between the plurality of cosine similarity values and the plurality of cosine similarity values of the plurality of cosine similarity values; and calculating a plurality of cosine similarity values between the plurality of cosine similarity values and the plurality of cosine similarity values of the plurality of cosine similarity values. The cosine convolution kernel adopts a cosine function form and has two learnable parameters of amplitude and frequency; the convolution operation formula of each convolution kernel in the channel convolution of the lth layer is as follows: wherein, and respectively represent the amplitude and frequency parameters of the lth layer cth channel convolution kernel, K is the size of the convolution kernel, is the output feature of the previous layer, i.e., the l-1th layer cth channel position i+m. Each layered grouped convolution layer divides the input features into C groups, and each group contains C' / G channel features, wherein C' is the output channel number of the previous layer, G is the number of groups, and is equal to the channel number C; the convolution operation formula of each convolution kernel in the grouped convolution of the lth layer is as follows: wherein, represents the output of the position i of the glayer ggroup, and are the amplitude and frequency parameters of the c' thchannel convolution kernel of the glayer ggroup, and g' = c' + (g - 1)(C' / G) is the position index of the c' thchannel of the g-1thgroup of the previous layer.

6. The method of claim 1, wherein the method is characterized by, According to the classification result of electroencephalogram anomaly detection, the contribution score of each channel electroencephalogram signal is calculated to determine the channel of the abnormal electroencephalogram signal and complete the abnormal electroencephalogram positioning, which comprises the following steps: According to the weight matrix of the normal and abnormal categories in the classification result of the electroencephalogram anomaly detection, and in combination with the multiple groups of electroencephalogram features output by the last layer of the grouped convolution layer, the contribution score of each channel electroencephalogram signal is calculated; According to the size of the contribution score, the channel of the abnormal electroencephalogram signal is determined, wherein the greater the contribution score value is, the greater the possibility of the abnormality of the corresponding channel electroencephalogram signal is.

7. The method of claim 1, wherein the method is characterized by, Based on the cross-entropy loss and the regularization term, a loss function of the grouped cosine convolution network is constructed, and the iterative training of the grouped cosine convolution network is performed based on the loss function until the training is completed. During the iterative training of the grouped cosine convolution network, the amplitude parameters and the frequency parameters of the cosine convolution kernel in each convolution layer of the network are iteratively updated by using the back propagation algorithm.

8. A system for electroencephalogram anomaly detection and localization based on grouped cosine convolutional network, characterized in that, The method comprises: The data acquisition module is configured to acquire the multi-channel electroencephalogram signals and pre-process the acquired signals. The abnormal electroencephalogram detection module is configured to input the pre-processed multi-channel electroencephalogram signals into the grouped cosine convolution network, group the multi-channel electroencephalogram signals by channel, and perform a cosine convolution operation on each group of electroencephalogram signals. The abnormal electroencephalogram alarm module is configured to trigger an alarm according to the classification probability distribution of the electroencephalogram anomaly detection and a preset threshold. The abnormal electroencephalogram channel positioning module is configured to calculate the contribution score of each channel electroencephalogram signal according to the classification result of the electroencephalogram anomaly detection, determine the channel of the abnormal electroencephalogram signal, and complete the positioning of the abnormal electroencephalogram.

9. An electronic device, characterized by comprising: The computer instructions are executed by the processor to complete the steps of the method for detecting and positioning electroencephalogram anomalies based on the grouped cosine convolution network.

10. A computer readable storage medium characterized by, The computer instructions are executed by the processor to complete the steps of the method for detecting and positioning electroencephalogram anomalies based on the grouped cosine convolution network.

Images