A multi-perspective and collaborative learning-based epilepsy detection method based on multi-scale feature fusion.

Abstract

This invention belongs to the field of intelligent epilepsy detection, and relates to an epilepsy detection method based on multi-scale feature fusion to construct a multi-viewpoint and collaborative learning approach. The method first performs deep learning to obtain a deep multi-viewpoint feature set based on three perspectives: time domain, frequency domain, and time-frequency domain. This invention further constructs a global multi-scale fusion network based on the deep multi-viewpoint feature set to learn a global fusion perspective of the three deep perspectives, as well as a global fusion multi-scale fusion network that locally enhances each perspective. Based on the above multi-scale fusion network, four perspectives are extracted: one generalized global fusion perspective and three locally enhanced global fusion perspectives, resulting in a total of seven deep feature perspectives along with the aforementioned three-viewpoint deep feature set. Compared to traditional classifiers, this invention provides a multi-viewpoint classifier, ultimately yielding an epilepsy detection model with high generalization and good transparency, which can more effectively improve detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent epilepsy detection, and relates to an epilepsy detection method based on multi-scale feature fusion to construct multi-view and collaborative learning. Background Technology

[0002] Epilepsy is a chronic, non-communicable brain disorder that affects people of all ages. There are approximately 50 million people with epilepsy worldwide, making it one of the most common neurological disorders globally. People with epilepsy have a three times higher risk of premature death compared to the general population. With proper and timely diagnosis and treatment, up to 70% of people with epilepsy will not experience a recurrence of seizures.

[0003] EEG is the most important auxiliary examination method for diagnosing epilepsy. It is valuable for diagnosing paroxysmal symptoms and helps to clarify the diagnosis and classification of the disease. EEG signals can be used to diagnose whether epilepsy is present and can also assess whether the patient is in a seizure state. Theoretically, any type of epileptic seizure can be recorded using EEG signals to capture the seizure or interictal epileptiform discharges. However, manual detection has always been a time-consuming and laborious process, placing a burden on neurologists. Therefore, it is necessary to develop automated detection algorithms to analyze EEG signals, thereby converting the information contained in the signals into a state output that can determine whether an epileptic seizure has occurred.

[0004] With the development of computer technology, scholars both domestically and internationally have deepened their research on EEG signals, and machine learning methods such as support vector machines, deep learning, ensemble learning, and fuzzy systems have been applied to this field. Based on these methods, the accuracy of automated epilepsy detection has been greatly improved. For existing intelligent EEG recognition methods, the effectiveness of feature extraction from EEG signals plays a crucial role in the quality of the recognition results. Although many EEG feature extraction methods have been proposed, further in-depth research is still needed on how to more effectively improve feature representation capabilities and reasonably increase the comprehensiveness of features. Summary of the Invention

[0005] This invention realizes a novel method for epilepsy detection based on multi-scale feature fusion to construct a global multi-view and collaborative learning method. The method includes two parts: a training phase and an application phase. The training phase includes four parts: an initial multi-view feature extraction module, a deep multi-view feature extraction module, a multi-view generation module based on multi-scale feature fusion, and a multi-view classifier module based on a fuzzy system.

[0006] Training Phase: The extraction of shallow and deep EEG features begins with the construction of frequency domain and time-frequency data using Fast Fourier Transform and wavelet packet decomposition. This data, combined with the original time-domain data, forms a three-view EEG dataset. Further, the data from each viewpoint is fed into a convolutional neural network to learn and obtain the deep features of each viewpoint. The construction of the global multi-view set requires the introduction of a multi-scale feature fusion network model to integrate information from the time domain, frequency domain, and time-frequency domain. Based on the selection of deep viewpoint information, a generalized global viewpoint with global information and multiple globally fused views that enhance different local information are constructed. Using multi-view learning techniques, a rule-based multi-view classifier is trained using the deep multi-view feature set and the globally fused multi-view feature set. This multi-view classifier, with its good transparency, is used to determine whether an epileptic seizure has occurred.

[0007] Usage phase: Acquire the EEG data to be tested, construct shallow features using Fast Fourier Transform and Wavelet Packet Decomposition, and then extract the deep features of the three shallow features as a deep multi-view feature set through three trained convolutional neural networks; then connect the deep multi-view feature set with CONCAT and put it into a multi-scale feature fusion network to generate a generalized global view and three locally enhanced global fusion views; finally, use the trained multi-view Takagi-Sugeuo-Kang (TSK) fuzzy classifier to classify the seven combined multi-view deep features to obtain the final classification result.

[0008] The novel epilepsy detection method based on multi-scale feature fusion and global multi-view collaborative learning integrates deep learning, multi-scale feature fusion, and multi-view learning techniques. The deep structural optimization representation of deep learning and the high representational power and comprehensive information of multi-scale feature fusion effectively combine these feature extraction methods with multi-view learning to fully extract effective information from EEG signals and improve the generalization ability of the classifier.

[0009] EEG signals are important signals recording the activity of neurons in the brain. In biomedical research on electrical signals, during an epileptic seizure, the abnormal discharge of neurons causes the EEG signal to differ from that in a normal state. Therefore, identifying EEG signals to detect epileptic seizures is an effective method for epilepsy detection. EEG signals possess various shallow features, such as frequency domain features, time-frequency domain features, and nonlinear features. While these features are helpful for epilepsy detection, they have a drawback: limited information content. Deep learning learns the weights of each layer by obtaining the desired output. Each layer in the hierarchical structure further adjusts the features to obtain inventive features that are more likely to produce the desired output. In other words, each layer optimizes the input features, thereby obtaining increasingly discriminative features. This paper's method uses convolutional neural networks to perform deep feature extraction on the time domain, frequency domain, and time-frequency data of EEG signals. A convolutional neural network consists of one or more convolutional layers and a fully connected layer at the top. The fully connected layer maps the feature space calculated by the previous layers (convolutional, pooling, etc.) to the sample label space, reducing the influence of feature location on the classification results and improving the robustness of the entire network. Therefore, this invention selects the output of this layer as the input of the learned features to the lower-level feature fusion network. The features obtained through automatic learning by the convolutional neural network have a smaller dimension than the original features, and the obtained features are non-linearly combined features with better discriminative ability, which can enable the subsequent classification model to have better generalization performance.

[0010] Figure 5 The process of extracting deep features from three perspectives is illustrated. For time-domain and frequency-domain data, this invention uses a one-dimensional convolutional neural network for feature extraction, while for the time-frequency domain, which is related to both time and frequency variations, a two-dimensional convolutional neural network is used for modeling. The corresponding feature sizes are represented as c@h and c@h*w, where c represents the number of feature channels in the layer and h*w represents the feature size. The specific settings are as follows:

[0011] (1) The input data size in the time domain is 22@256, which means there are 22 signal channels, and we set the sampling frequency to 256Hz, with one sample per second.

[0012] (2) The input data size in the frequency domain is 22@27, which means that the number of signal channels is 22 and the number of sampling points in the frequency is 27;

[0013] (3) The time-frequency domain is 22@256*14, which means that the number of signal channels is 22, the number of sampling points in time is 256, and the number of sampling points in frequency is 14.

[0014] (4) The front-layer features extracted from the three perspectives in the fully connected layer are uniformly mapped to 1@1024. This helps to reduce the impact of feature location on classification while standardizing the feature vectors to facilitate the training of the lower-layer feature fusion network.

[0015] (5) The three front-layer deep features are connected into a global deep feature by the CONCAT method, which is used to construct a new global fusion perspective and a global multi-view with local enhancement.

[0016] This invention uses the ReLU function as the activation function to achieve non-linear transformation of features when extracting deep features. The ReLU function is expressed as follows:

[0017] R(x) = max(0,x)

[0018] The cross-entropy loss function was used in network optimization, and its definition is as follows:

[0019]

[0020] For the multi-view depth features already obtained, the receptive regions emphasized by the feature vectors for different details of the data vary greatly. Effective fusion can refine the model and reduce the impact of noise. Therefore, how to effectively fuse features is a problem that needs further exploration. Many inventions at home and abroad have improved detection performance by fusing multi-layer networks. Here, we apply multi-scale feature fusion to the field of epilepsy detection to explore an efficient feature network model.

[0021] Each sample's deep features contain rich information, yet lack a unified effective information range and amount. Simply increasing network complexity to recognize large sample data only leads to verbose models and overfitting; conversely, simple models lack learning ability and cannot simultaneously handle diverse information ranges and amounts. Multi-scale feature fusion effectively solves this problem. For signal data containing multiple information ranges and information of different sizes, it utilizes shallow features to locate simple, limited information, while simultaneously leveraging deep network features to focus on complex and abundant information.

[0022] like Figure 6As shown, we construct an architecture with lateral connectivity and both bottom-up and top-down characteristics to build high-level semantic feature maps at all scales. On the one hand, although the feature maps obtained by the bottom-up path have weaker semantic information representation capabilities, they have stronger spatial information representation capabilities, which is more conducive to the extraction of simple information. On the other hand, the other top-down path generates high-resolution features by sampling spatially coarser but semantically stronger feature maps from the upper-layer network, which is beneficial for extracting features from complex samples. Most importantly, the two paths can be laterally connected to ultimately obtain features that enhance spatial information representation capabilities while preserving high semantic and high-resolution characteristics.

[0023] Figure 6 The method includes two multi-scale fusion extraction approaches. The first approach extracts a generalized global network architecture consisting of a bottom-up path, a top-down path, and a lateral connection structure. In the bottom-up path, the first layer of one-dimensional convolutional downsampling uses 8 convolutional kernels with a kernel size of 3, then passes through a one-dimensional pooling layer with a stride of 2 to obtain an 8*512 feature map, which is then convolved with 1*1 to retain the feature map. The second layer of one-dimensional convolutional downsampling uses 16 convolutional kernels with a kernel size of 3, then passes through a one-dimensional pooling layer with a stride of 2 to obtain a 16*256 feature map, which is then convolved with 1*1 to retain the feature map. The third layer of one-dimensional convolutional downsampling uses 32 convolutional kernels with a kernel size of 3, then passes through a one-dimensional pooling layer with a stride of 2 to obtain a 32*128 feature map. In the top-down path, the first layer performs a 1*1 convolution to obtain a 1*128 feature map; the second layer uses an upsampling convolution with a stride of 2 to obtain a 1*258 feature map, which is then fused laterally with the feature map retained from the second layer in the bottom-up path using CONCAT, and then subjected to a 1*1 convolution; the third layer uses an upsampling convolution with a stride of 2 to obtain a 1*512 feature map, which is then fused laterally with the feature map retained from the first layer in the bottom-up path using CONCAT, and then subjected to a 1*1 convolution to obtain a 1*512 feature map; the fourth layer performs an upsampling convolution with a stride of 2 to obtain a global view fused depth feature map of size 1*1024.

[0024] The second network architecture, which extracts local enhancements from a global fusion perspective, includes a bottom-up and top-down path and a lateral connection structure. In the bottom-up path, the first layer of one-dimensional convolutional downsampling uses 8 convolutional kernels with a kernel size of 3, and then passes through a one-dimensional pooling layer with a stride of 2 to obtain an 8*512 feature map. This feature map is then convolved with 1*1 and retained. The second layer of one-dimensional convolutional downsampling uses 16 convolutional kernels with a kernel size of 3, and then passes through a one-dimensional pooling layer with a stride of 2 to obtain a 16*256 feature map. This feature map is then convolved with 1*1 and retained. The third layer of one-dimensional convolutional downsampling uses 32 convolutional kernels with a kernel size of 3, and then passes through a one-dimensional pooling layer with a stride of 2 to obtain a 32*128 feature map. In the top-down path, the first layer performs a 1*1 convolution to obtain a 1*128 feature map; the second layer uses an upsampling convolution with a stride of 2 to obtain a 1*258 feature map, which is then fused laterally with the feature map retained from the second layer in the bottom-up path using CONCAT, and then performs a 1*1 convolution; the third layer uses an upsampling convolution with a stride of 2 to obtain a 1*512 feature map, which is then fused laterally with the feature map retained from the first layer in the bottom-up path using CONCAT, and then performs a 1*1 convolution to obtain a 1*512 feature map; the fourth layer uses an upsampling convolution with a stride of 2, and then performs lateral CONCAT fusion with the time domain, frequency domain, and time-frequency domain depth view feature maps respectively, resulting in three locally enhanced global view depth feature maps of size 1*1024 after a 1*1 convolution.

[0025] Finally, based on the introduced multi-scale feature fusion network, this invention obtains feature sets from four perspectives. These four global perspectives encompass feature information from all directions, while enhancing their respective advantages while preserving commonalities, thus possessing both high feature representation capabilities and comprehensiveness, which is beneficial for subsequent multi-view learning.

[0026] The specific steps for constructing multiple perspectives are as follows:

[0027] Step 1: Feed the raw EEG signal (time-domain signal) into a one-dimensional convolutional neural network to extract time-domain depth features.

[0028] Step 2: Input the shallow frequency domain features into a one-dimensional convolutional neural network to extract deep frequency domain features.

[0029] Step 3: Input the shallow time-frequency domain features into a two-dimensional convolutional neural network to extract the deep time-domain features.

[0030] Step 4: Use the CONCAT method to extract the deep multi-view feature set {X} 1 ,X2 ,X 3} converted to

[0031] Step 5: Construct a generalized global perspective fused deep feature network using a multi-scale feature fusion network.

[0032] Step 6: Construct global perspective deep features by using a multi-scale feature fusion network to enhance local temporal, frequency, and time-frequency information respectively.

[0033] Step 7: Combine the deep multi-view deep feature set and the global view deep feature set to construct a seven-view feature set for training.

[0034] How to efficiently apply the constructed seven-view feature set to epilepsy detection is a question that needs in-depth exploration. Here, we introduce multi-view learning for discussion. Multi-view classifiers can effectively utilize the collaboration of multiple views to construct classifiers with better generalization ability. MV-TSK-FS is a multi-view classifier constructed based on the TSK fuzzy system. It inherits the interpretability of the TSK fuzzy system and can label the weights of each view, thereby quantifying the importance of each view and ultimately obtaining the effect of collaborative decision-making. This method first trains a fuzzy system model for each view feature, obtaining high-dimensional features mapped from the original features to a high-dimensional space. Then, based on the weighting mechanism of the multi-view fuzzy system and the multi-view collaborative learning mechanism, the objective function of MV-TSK-FS is constructed as follows:

[0035]

[0036]

[0037] in,

[0038]

[0039]

[0040]

[0041] here The weight of the k-th viewpoint; y i Let C represent the C-dimensional label vector of the i-th sample; The expected rule for the k-th viewpoint; w represents the fixed prior knowledge of the l-th perspective; k Let C be the weight of the k-th viewpoint; C represents the number of sample categories; N represents the total number of samples; and... λ co and λw These are all regularization parameters. The first equation introduces a multi-view weighting mechanism to control the loss term of each view sub-model; the second equation is a multi-view collaborative learning mechanism, which is used to learn view independence information from each view by using the classic TSK-FS; the third equation is a view weighting mechanism based on Shannon's sentropy, used to adjust the importance of different views, and the optimal weights can be further used for the final decision.

[0042] Finally, according to and weight w k The global decision value can be calculated from the decision values ​​of each perspective:

[0043]

[0044] The specific steps of training the MV-TSK-FS classifier are as follows:

[0045] Step 1: Train a TSK model for each viewpoint to obtain the antecedent parameters for each viewpoint, and calculate the original model using the antecedent parameters. Mapped to higher-dimensional space

[0046] Step 2: Transform high-dimensional features By inputting the objective function into the classifier of the multi-view fuzzy system, the consequent parameters of each viewpoint can be obtained. and weight w k ;

[0047] Step 3: According to and w k The final decision value Y is obtained by calculating the decision function of the multi-view fuzzy system classifier. Test The algorithm terminates.

[0048] The beneficial effects of this invention are:

[0049] A key innovation of this method is the introduction of a global multi-scale fusion network to construct different global perspectives. To evaluate the effectiveness of this module, experiments were conducted to subtract and retain portions of the global multi-views constructed by multi-scale feature fusion, and the performance of each combination was experimentally obtained. This invention introduces a rule-based multi-view fuzzy system as a classifier. Compared to traditional classifiers, this classifier possesses multiple perspectives, ultimately resulting in an epilepsy detection model with high generalization and good transparency, which can more effectively improve detection accuracy. Attached Figure Description

[0050] Figure 1 This is a flowchart of the algorithm method framework of the present invention.

[0051] Figure 2 It is EEG time-domain signal data.

[0052] Figure 3 It is EEG frequency domain signal data.

[0053] Figure 4 It is EEG time-frequency domain signal data.

[0054] Figure 5 This is a diagram of the deep feature extraction network structure for time-domain, frequency-domain, and time-frequency data according to the present invention.

[0055] Figure 6 This is the structure diagram of the global fusion, multi-view, multi-scale fusion feature network of the present invention. Detailed Implementation

[0056] The present invention will now be described in detail with reference to the accompanying drawings and embodiments:

[0057] like Figures 1-5 As shown, this invention implements a novel epilepsy detection method based on multi-scale feature fusion to construct a global multi-view and collaborative learning approach. This method comprises four parts: an initial multi-view feature extraction module, a deep multi-view feature extraction module, a multi-view generation module based on multi-scale feature fusion, and a multi-view classifier module based on a fuzzy system. The core tasks of feature extraction are the construction of the initial view and the extraction of deep features from the initial view. The multi-view generation module, based on multi-scale feature fusion, mainly implements the construction process of global and deep view deep feature sets. The classification part mainly implements the construction of a classifier based on multi-view collaborative ensemble learning.

[0058] The specific steps of the training phase are as follows: The deep multi-view deep features of this method are constructed by three shallow features in the time domain, frequency domain, and time-frequency domain; the global multi-view deep features are obtained by multi-scale feature fusion of the deep multi-view deep feature set.

[0059] Example 1

[0060] Following the implementation method of the training phase, an example was completed using data from the standard dataset CHB-MIT. This dataset collected 24 sets of EEG signals from 23 patients at Boston Children's Hospital in the United States (sets 1 and 21 are records from different periods of the same patient). These signals were processed using a 10-20 international standard lead system with electrodes placed, a sampling frequency of 256 Hz, and 16-bit resolution. Lead information is shown in Table 1 (Channel 23 and Channel 15 are duplicates and need to be removed in actual use).

[0061]

[0062] Table 2 presents the performance metrics of this algorithm. The average accuracy, sensitivity, and specificity of the 23 data sets are 98.76%, 98.44%, and 98.90%, respectively. Except for group 12, the performance metrics of this method are all above 96%. In particular, group 9 achieves 100% accuracy, sensitivity, and specificity. From the perspective of accuracy and sensitivity, the performance of this method is especially outstanding.

[0063] Table 3 summarizes the research results on EEG-based epilepsy detection in recent years. These comparative methods were tested using the same CHB-MIT dataset for epilepsy detection. The table shows the main experimental methods, experimental settings, and classification performance. It is worth noting that Khan, Tian X, et al. achieved good specificity results at the expense of high sensitivity, enabling better exclusion of non-epileptic seizure segments. However, compared to this, in epilepsy detection, high sensitivity allows for more timely treatment and reduces missed diagnoses. This invention achieves high performance in both sensitivity and specificity. It can be seen that the method of this invention is superior in overall performance.

[0064] Table 3 Experimental results of different methods

[0065]

[0066] Example 2

[0067] A key innovation of this paper's method is the introduction of a global multi-scale fusion network to construct different global perspectives. To evaluate the effectiveness of this module, experiments were conducted to combine the global multi-perspectives constructed by subtracting and retaining the multi-scale feature fusion components, and the performance of each combination was experimentally obtained.

[0068] As shown in Table 4, each global fusion perspective plays a role and complements the others in different combinations. The collaboration between local and global perspectives yields better results in the epilepsy recognition task. In particular, global fusion perspective 1 and global fusion perspective 3 are more helpful for the epilepsy recognition task.

[0069] Table 4 Performance Comparison of Different Combined Viewpoints

[0070]

[0071] As shown in Table 4, the global fusion perspective of locally enhanced frequency domain information performs superiorly in every combined perspective, and the global fusion perspective of each locally enhanced feature shows better performance than its depth perspective. Furthermore, Table 5 compares the global fusion features and local features of locally enhanced features, demonstrating that the multi-scale fusion feature method indeed enhances the representational ability of features and can be more effectively applied in the field of epilepsy detection.

[0072] Table 5 Comparison of characteristic properties before and after local enhancement

[0073]

[0074]

Claims

1. A multi-perspective and collaborative learning-based epilepsy detection method based on multi-scale feature fusion is characterized by the following steps: Down: Phase 1, the steps for constructing shallow features of EEG signals are as follows: Step 1: Obtain shallow frequency domain features through Discrete Fourier Transform; Step 2: Obtain shallow time-frequency features through wavelet packet decomposition; Phase Two, the steps for constructing local multi-view depth features based on the initial viewpoint are as follows: Step 1: Feed the raw EEG signal (time-domain signal) into a one-dimensional convolutional neural network to extract time-domain depth features. Step 2: Input the shallow frequency domain features into a one-dimensional convolutional neural network to extract deep frequency domain features. ; Step 3: Input the shallow time-frequency domain features into a two-dimensional convolutional neural network to extract the deep time-domain features. Phase 3, the steps for constructing global multi-view features based on multi-scale feature fusion are as follows: Step 1: Use the CONCAT method to extract deep multi-view feature sets Convert to ; Step 2: Construct a generalized global-view fused deep feature network using a multi-scale feature fusion network. ; The multi-scale feature fusion network architecture of Stage 3 includes a bottom-up and top-down path and a lateral connection structure. In the bottom-up path, the first layer of one-dimensional convolutional downsampling uses 8 convolutional kernels with a kernel size of 3, and then passes through a one-dimensional pooling layer with a stride of 2 to obtain an 8*512 feature map. This feature map is then convolved with 1*1 and the feature map is retained. The second layer of one-dimensional convolutional downsampling uses 16 convolutional kernels with a kernel size of 3, and then passes through a one-dimensional pooling layer with a stride of 2 to obtain a 16*256 feature map. This feature map is then convolved with 1*1 and the feature map is retained. The third layer of one-dimensional convolutional downsampling uses 32 kernels. A convolutional kernel of size 3 is used, followed by a one-dimensional pooling layer with a stride of 2 to obtain a 32*128 feature map. In the top-down path, the first layer performs a 1*1 convolution to obtain a 1*128 feature map. The second layer uses an upsampling convolution with a stride of 2 to obtain a 1*258 feature map, which is then fused laterally with the feature map preserved in the second layer of the bottom-up path using CONCAT, followed by another 1*1 convolution. The third layer uses an upsampling convolution with a stride of 2 to obtain a 1*512 feature map, which is then fused laterally with the feature map preserved in the first layer of the bottom-up path using CONCAT, followed by another 1*1 convolution to obtain a 1*512 feature map. Step 3: Construct a global view using a multi-scale feature fusion network to enhance local temporal, frequency, and time-frequency information respectively. Angular depth features ; Step 4: Combine the local multi-view deep feature sets from Phase 3 to construct a seven-view feature set for training. ； Phase 4: Based on MV TSK The steps for training multi-view deep features using FS are as follows: MV TSK FS is a multi-view classifier constructed based on the TSK fuzzy system; Step 1: Train a TSK model for each viewpoint to obtain the antecedent parameters for each viewpoint, and then calculate using the antecedent parameters. The original mapping to the high-dimensional space can be calculated. ; Step 2: Transform high-dimensional features The objective function of the multi-view fuzzy system classifier is input, and the result of solving for each viewpoint is obtained. Part parameters and weight wk; Step 3: According to And wk, under the calculation of the decision function of the multi-view fuzzy system classifier, finally obtain the decision value. YTest, the algorithm terminates.

2. The epilepsy detection method based on multi-scale feature fusion to construct multiple perspectives and collaborative learning as described in claim 1. Its features are: The CNN network architecture for extracting temporal depth features in the first step of stage two includes four one-dimensional convolutional layers, four batch normalization (BN) layers, one flattening layer, one dropout layer, and one fully connected layer. The first convolutional layer consists of 64 kernels of size 64, resulting in a 64*193 feature map, which then passes through the first BN layer. The second convolutional layer consists of 32 kernels of size 32, resulting in a 32*162 feature map, which then passes through the second BN layer. The third convolutional layer consists of 16 kernels of size 16, resulting in a 16*147 feature map, which then passes through the third BN layer. The fourth convolutional layer consists of 8 kernels of size 8, resulting in an 8*140 feature map, which then passes through the fourth BN layer. Following this, the flattening layer, dropout layer, and fully connected layer are applied, ultimately yielding a 1*1024 feature map.

3. The epilepsy detection method based on multi-scale feature fusion to construct multi-view and collaborative learning as described in claim 1 or 2, characterized in that: The CNN network architecture for extracting frequency domain depth features in the second step of stage two includes two one-dimensional convolutional layers, two batch normalization (BN) layers, one flattening layer, one dropout layer, and one fully connected layer. The first convolutional layer of the CNN network architecture has 32 convolutional kernels with a kernel size of 4, resulting in a 32*24 feature map, which then passes through the first BN layer. The second convolutional layer has 16 convolutional kernels with a kernel size of 2, resulting in a 23*16 feature map, which then passes through the second BN layer. Next, the flattening layer, dropout layer, and fully connected layer are applied, finally obtaining a 1*1024 feature map.

4. The epilepsy detection method based on multi-scale feature fusion to construct multiple perspectives and collaborative learning as described in claim 1, characterized in that: The CNN network architecture for extracting time-frequency depth features in the third step of stage two includes four 2D convolutional layers, four batch normalization (BN) layers, one flattening layer, one dropout layer, and one fully connected layer. The first convolutional layer of the CNN network architecture has 64 kernels of size 1*128, resulting in a 64*14*129 feature map, which then passes through the first BN layer. The second convolutional layer has 32 kernels of size 2*64, resulting in a 32*13*66 feature map, which then passes through the second BN layer. The third convolutional layer has 16 kernels of size 2*32, resulting in a 16*12*35 feature map, which then passes through the third BN layer. The fourth convolutional layer has 8 kernels of size 4*16, resulting in an 8*9*20 feature map, which then passes through the fourth BN layer. Following this, a flattening layer, a dropout layer, and a fully connected layer are applied, ultimately yielding a 1*1024 feature map.

5. The epilepsy detection method based on multi-scale feature fusion to construct multiple perspectives and collaborative learning as described in claim 1, characterized in that: The generalized global view fusion deep feature construction in the second step of stage three involves performing an upsampling convolution with a stride of 2 after obtaining the 1*512 feature map of stage three, ultimately resulting in a global view fusion deep feature map of size 1*1024.

6. The epilepsy detection method based on multi-scale feature fusion to construct multi-view and collaborative learning as described in claim 1, characterized in that: The third step of the third stage involves constructing a local enhanced global view depth feature map. After obtaining the 1*512 feature map from the third stage, it is again subjected to an upsampling convolution with a stride of 2. Then, it is laterally CONCAT fused with the local view feature maps in the time domain, frequency domain, and time-frequency domain, respectively. After a 1*1 convolution, three local enhanced global view depth feature maps of size 1*1024 are obtained.

Images