Automatic labeling system for abnormal bands in epileptic EEG images

Abstract

The present invention discloses an automatic labeling system for abnormal bands in epileptic EEG images, which relates to the technical field of epileptic EEG. The system performs frequency domain analysis and time domain feature extraction on EEG signals through Fourier transform and convolution operations, and can accurately identify abnormal waveforms such as sharp waves, spike waves and slow waves, thereby enhancing the system's sensitivity to subtle abnormal bands and merging highly similar abnormal bands into a complete abnormal event, avoiding redundant labeling and repeated event labeling. By calculating spectral entropy and spectral entropy difference, setting a reasonable threshold, and further refining it in combination with ratio difference, the system can more accurately identify and label different epileptic stages. The system optimizes labeling through multi-dimensional feature differences such as spectral entropy difference and ratio difference of different bands, and can more accurately distinguish between onset, precursors and normal in different stages of epilepsy. This multi-level labeling method effectively improves the predictive ability of epilepsy.

Description

Technical Field

[0001] The present invention relates to the technical field of epileptic electroencephalograms, and in particular to an automatic marking system for abnormal wavebands in epileptic electroencephalograms. Background Art

[0002] Epilepsy is a neurological disease caused by abnormal discharges of brain neurons, often accompanied by recurrent seizures. Monitoring the electrical activity of epilepsy patients using an EEG can help doctors identify the location of the lesion, diagnose the type of epilepsy, assess the severity of the condition, and provide important evidence for treatment decisions. Spatial and temporal analysis of the EEG allows doctors to identify the location of the epileptic lesion. For example, some epilepsy patients may experience abnormal discharges in specific brain regions, such as the frontal and temporal lobes. EEG analysis can pinpoint the source of these abnormal discharges, which is crucial for determining interventions such as surgical treatment. Some epilepsy patients experience premonitory signs before a seizure, manifesting as specific changes in the EEG, such as changes in localized electrical activity. By monitoring the EEG for 24 hours or longer, doctors can detect potential premonitory signs and provide a basis for seizure prevention and intervention. Furthermore, the EEG can also reflect the patient's brain function, especially in the absence of an epileptic seizure. For example, the EEG can display basic brain electrical activity, such as alpha and theta waves, which helps assess whether the patient's neurological function is normal.

[0003] For EEG annotation of epilepsy patients, the existing automatic annotation system has relatively mixed sample features and lacks age stratification, making it impossible to perform targeted annotation of EEGs of patients of different ages. The epileptiform waveforms of different ages vary greatly, such as the spike wave amplitude ranging from 50uV to 500uV, resulting in low annotation accuracy. Summary of the Invention

[0004] In view of the above problems existing in the prior art, the present application provides an automatic labeling system for abnormal bands in epileptic EEG images.

[0005] In one aspect, the present invention provides a system for automatically labeling abnormal bands in epileptic EEG images, comprising:

[0006] The data acquisition module is used to collect EEG data sets, including original EEG signals and corresponding annotation labels, to obtain EEG sample data sets;

[0007] The signal preprocessing module is used to preprocess the input EEG signal to be labeled, perform noise removal, filtering, and standardization;

[0008] The abnormality identification module is used to randomly segment the pre-processed EEG signal to be labeled, identify abnormal bands in the input EEG signal to be labeled based on a normal EEG data set and multiple EEG signals, retain the EEG segments containing abnormal bands, generate a new interval set, and calculate the similarity of adjacent abnormal bands in the interval set. Based on the similarity of adjacent abnormal bands, the EEG segments containing the same abnormal band are merged;

[0009] The labeling module is used to automatically label the identified abnormal bands, with specific labeling status including normal, preictal aura and epileptic seizure;

[0010] The correction module is used to correct the anomaly bands that have been marked and establish a correction mechanism.

[0011] Optionally, the data acquisition module includes:

[0012] The age classification unit is used to preliminarily classify the collected EEG dataset according to age. For each EEG sample, the age classification unit uses age as an additional label and divides the EEG dataset into four main categories according to the patient's age, namely newborn, child, adolescent and adult datasets;

[0013] The state division unit is used to perform secondary division on the EEG data sets of normal, preictal, and epileptic seizure states in each age category to obtain an EEG sample data set, and store each EEG sample data set together with its corresponding age group and state classification label.

[0014] Optionally, the anomaly identification module includes:

[0015] The segmentation unit is used to segment the pre-processed EEG signal to be labeled, split it according to the time window, determine the time window size of each segment, and segment the EEG signal according to the window sliding step size and overlap method. During the sliding window process, the start and end time of the EEG signal contained in each window needs to be accurately recorded and synchronized to ensure that the signal timestamp remains consistent;

[0016] The feature extraction unit is used to extract EEG segments containing abnormal bands. Through multiple EEG signals, each EEG segment containing abnormal bands is extracted and stored as a new interval set. Each interval set should include the start time, end time, type of abnormal band and extracted features.

[0017] Optionally, the anomaly identification module further includes:

[0018] The difference calculation unit is used to calculate the similarity of adjacent abnormal bands in the interval set, and normalize the power features of each frequency band extracted from the EEG signal. The purpose of power feature normalization is to convert the power features of different frequency bands into a dimensionless standardized value. The power features after normalization are represented as a feature vector. Each element in the feature vector represents the power features of different frequency bands. By normalizing the power features of these different frequency bands, the abnormal feature vector of the abnormal band is obtained, the Euclidean distance between the abnormal feature vectors of two adjacent abnormal bands is calculated, and the difference between the two abnormal bands is calculated. The similarity is determined by the Euclidean distance. According to the similarity in the previous time period, the mean-standard deviation method is used to set the similarity threshold. When the similarity of two adjacent abnormal bands is greater than or equal to the set similarity threshold, it indicates that the two adjacent abnormal bands currently belong to the same abnormal activity. At this time, the abnormal bands belonging to the same abnormal activity are merged into one abnormal event and a new data set is generated. When the similarity of two adjacent abnormal bands is lower than the set similarity threshold, it indicates that the two adjacent abnormal bands currently belong to different abnormal events. Each abnormal band is regarded as an independent abnormal event and is labeled and analyzed separately.

[0019] Optionally, the annotation module includes:

[0020] The sample extraction unit is used to extract the power feature data of each frequency band in the EEG signal under different states in the EEG sample data set. According to the age of the patient to be annotated, the unit automatically extracts EEG samples of normal, preictal, and epileptic seizure states of the patient of the corresponding age from the EEG sample data set. The patient's age information is used as a screening criterion to ensure that the extracted samples are consistent with the typical EEG activity pattern of the age group. The power features of each frequency band in the EEG signal under different states are extracted and normalized to obtain the sample feature vector.

[0021] The preliminary judgment unit is used to preliminarily judge the marking range and input the vector data after feature extraction, which contains the feature information of the abnormal feature vector and the sample feature vector. The feature difference is obtained by subtracting the features of each dimension item by item. The feature difference is used to represent the deviation between the power characteristics of each frequency band in the abnormal band and the power characteristics of each frequency band in the normal sample. According to the feature difference in the previous time period, the mean-standard deviation method is used to pre-set the feature threshold. When the feature difference is less than the feature threshold, it indicates that the power characteristics of each frequency band in the abnormal band are highly overlapped with the power characteristics of each frequency band in the normal sample. At this time, the abnormal band is marked as the corresponding state in the EEG data set sample. If the feature difference is greater than or equal to the feature threshold, it indicates that the power characteristics of each frequency band in the abnormal band are not highly overlapped with the power characteristics of each frequency band in the normal sample, and the abnormal band is not marked.

[0022] Optionally, the annotation module also includes:

[0023] The constraint unit is used to further optimize the preliminary annotation range and calculate the spectral entropy of the abnormal band EEG. The spectral entropy difference is obtained by subtracting the spectral entropy value of the abnormal band EEG from the spectral entropy under normal conditions. According to the spectral entropy difference in previous time periods, the mean-standard deviation method is used to set the threshold of the spectral entropy difference, calculate the ratio difference of some data in the abnormal feature vector, and further optimize the annotation range based on the preliminary annotation.

[0024] Optionally, the correction module includes:

[0025] The verification unit is used to verify the annotation of the abnormal bands that have been marked, use the wavelet transform method to perform time-frequency analysis on the EEG signal, generate a time-frequency diagram, and obtain the frequency distribution corresponding to each time period. According to the analysis results of the time-frequency diagram, if there is an error in the time length of the marked abnormal band, the marked time range can be dynamically adjusted based on the instantaneous frequency change, and by analyzing the frequency band characteristics of the time-frequency diagram, it is checked whether there is an incorrect frequency band allocation, and then the marked frequency range is adjusted.

[0026] Optionally, the correction module also includes:

[0027] The adjustment unit is used to adjust the annotated range of the abnormal band and compare the automatically annotated abnormal band with the actual time-frequency characteristics. If the annotated range is found to be inaccurate, the boundary is fine-tuned and expanded.

[0028] Beneficial effects of the present invention: The data acquisition module constructs a structured, label-rich sample library for the EEG data set by introducing age division units and state division units, which greatly improves the efficiency and accuracy of subsequent model training and analysis. It performs frequency domain analysis and time domain feature extraction on the EEG signal through Fourier transform and convolution operations, and can accurately identify abnormal waveforms such as sharp waves, spike waves and slow waves, thereby enhancing the system's sensitivity to subtle abnormal bands, thereby improving the recognition accuracy of abnormal bands. By calculating the similarity of adjacent abnormal bands and merging them, highly similar abnormal bands can be merged into a complete abnormal event, avoiding redundant annotations and repeated event annotations. The segmentation unit accurately segments the EEG through a sliding window mechanism and records the start and end time of each window to ensure signal timestamps. Accurate synchronization not only ensures the temporal consistency of data processing, but also provides an accurate time reference for the subsequent abnormal band positioning. The initial judgment unit can quickly determine the possible state of the abnormal band by calculating the feature difference between the abnormal band and the normal sample. The constraint unit further optimizes the annotation results by calculating the spectral entropy and the spectral entropy difference. By setting a reasonable threshold and further refining it in combination with the ratio difference, the system can more accurately identify and annotate different epilepsy stages. The system optimizes the annotation through multi-dimensional feature differences such as spectral entropy difference and ratio difference of different bands, and can more accurately distinguish between different stages of epilepsy, onset, precursors and normal. This multi-level annotation method effectively improves the predictive ability of epilepsy, helps to identify the risk of epileptic seizures early, and thus provides a reliable basis for clinical treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application.

[0030] Figure 1 This is a block diagram of the system for automatically labeling abnormal bands in epileptic EEG images of the present invention. DETAILED DESCRIPTION

[0031] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0032] like Figure 1 As shown, the present invention proposes an automatic labeling system for abnormal bands in epileptic EEG images, comprising:

[0033] The data acquisition module is used to collect EEG data sets, including original EEG signals and corresponding annotation labels, to obtain EEG sample data sets;

[0034] The data acquisition module includes: an age classification unit, which is used to preliminarily divide the collected EEG dataset according to age. For each EEG sample, the age classification unit uses age as an additional label and divides the EEG dataset into four main categories according to the patient's age, namely newborn, child, adolescent and adult datasets.

[0035] The specific classification criteria are: newborns (0-1 years old), EEG rhythm is unstable, containing a large number of delta waves and theta waves, EEG rhythm is irregular, "burst-suppression" pattern frequently appears, epileptic waveform characteristics are mostly intermittent sharp waves or multiple spike waves, not necessarily accompanied by obvious movement, seizure activity may show "myoclonic seizure" style, and is more dependent on EEG; children (1-12 years old), EEG has relatively clear rhythm, such as alpha waves (8-13Hz) that begin to appear when eyes are closed, and frequent sleep-related sharp waves may be misdiagnosed as epilepsy. Common benign central-temporal epilepsy in children, EEG characteristics: central-temporal sharp waves appear during sleep, frequency It is about 3-4Hz and absence epilepsy, with a 3Hz spike-slow wave syndrome, which is common in 6-10 years old; in adolescents (12-18 years old), the EEG alpha waves are active, and the background rhythm tends to be adult-like, which is common in juvenile myoclonic epilepsy, with frequent attacks in the early morning, and the EEG shows a 4-6Hz multi-spike-slow wave syndrome, which is sensitive to flash stimulation. The causes of attacks may include lack of sleep, visual stimulation, and mental stress; in adults and the elderly (>18 years old), the EEG alpha waves dominate the waking state (8-13Hz), and enhanced beta waves may indicate an anxious state. The background slow wave activity in the elderly is enhanced, which may be related to brain atrophy. The epileptic waveform characteristics are common in focal epilepsy.

[0036] It can be seen that the EEG characteristics of different age groups vary greatly. The EEG characteristics of newborns and children are more dependent on the development stage. The background rhythm and seizure manifestations of the EEG often show relatively "fuzzy" characteristics. For example, sleep-related sharp waves or irregular activities often appear, which can easily be confused with epileptic seizures. The EEG background of adults is relatively stable, and the manifestations of seizures are more obvious, usually manifested as regular abnormal activities such as sharp waves, spike waves, and spike-slow waves. With age, the EEG of the elderly may show more background slow waves, which may be caused by brain atrophy, aging, etc. The epilepsy manifestations of this group are also often more hidden, especially in complex focal epilepsy such as temporal lobe epilepsy. Therefore, classification according to different age groups can effectively improve the accuracy of EEG abnormality detection.

[0037] The state division unit is used to perform secondary division on the EEG data sets of normal, preictal, and epileptic seizure states in each age category to obtain an EEG sample data set, and store each EEG sample data set together with its corresponding age group and state classification label.

[0038] Under normal circumstances, there are no sharp waves or spike waves in the normal state, the background rhythm is stable, symmetrical on the left and right, and there is no sudden activity; before the attack, intermittent high-amplitude slow waves appear in the local area, accompanied by slight sharp waves, the frequency slows down, and regional synchronization is enhanced, which usually occurs a few seconds before the attack; during an epileptic seizure, high-frequency sharp waves or multiple spike waves suddenly appear, followed by rhythmic spike-slow waves, the signal amplitude increases, the frequency accelerates, and it is explosive. The electroencephalogram shows typical epileptic-like discharges. Further division of different label states based on each age group can significantly improve the accuracy of EEG abnormality judgment, improve the richness of EEG sample data sets, and reduce the impact of age differences on diagnostic results.

[0039] The signal preprocessing module is used to preprocess the input EEG signal to be labeled, and perform noise removal, filtering, and standardization. There are often multiple noise sources in EEG signals, such as power supply noise, myoelectric noise, eye movement artifacts, artifacts, etc. These noises will not only interfere with signal analysis, but also cause errors in model training. Therefore, noise removal is the primary goal of EEG signal preprocessing. By smoothing and denoising the EEG signal, low-quality signals or abnormal waveforms can be eliminated to ensure that the analysis model is trained based on real neural activity. The amplitude, frequency and other characteristics of the EEG signal may vary greatly between different devices and individuals. Standardization can help eliminate such individual differences, enable the model to adapt to different data sources, improve data quality, and thus improve the accuracy and reliability of the analysis results.

[0040] In an embodiment of the present invention, the data acquisition module constructs a structured, label-rich sample library for the EEG dataset by introducing an age division unit and a state division unit, which greatly improves the efficiency and accuracy of subsequent model training and analysis. First, the age division unit automatically classifies the original EEG data into four categories: newborns, children, adolescents and adults according to the patient's age, so that the model can fully learn the characteristic differences of EEG signals in different age groups and improve the model's adaptability to signals on different physiological bases; secondly, the state division unit further subdivides the EEG state of each sample into normal, pre-ictal, and epileptic seizure based on the age division, so that the data not only has the time dimension and individual difference dimension, but also covers the pathological state dimension, providing multi-dimensional label support for abnormality detection and seizure prediction models. This systematic data organization method helps to improve the representativeness and label quality of training data, effectively reduce the misjudgment rate of the model, and provide a basis for clinical individualized EEG analysis, enhancing the practicality and reliability of artificial intelligence-assisted diagnosis.

[0041] Furthermore, the abnormality recognition module is used to randomly segment the pre-processed EEG signal to be labeled, identify abnormal bands in the input EEG signal to be labeled based on a normal EEG data set through multiple EEG signals, retain the EEG segments containing abnormal bands, generate a new interval set, and calculate the similarity of adjacent abnormal bands in the interval set, and merge the EEG segments containing the same abnormal bands based on the similarity of adjacent abnormal bands; the abnormality recognition module includes: a segmentation unit, used to segment the pre-processed EEG signal to be labeled, split it according to the time window, determine the time window size of each segment, and segment the EEG signal according to the window sliding step and overlapping method. During the sliding window process, the start and end time of the EEG signal contained in each window needs to be accurately recorded and synchronized to ensure that the timestamp of the signal remains consistent.

[0042] EEG signals are continuous time series data with strong variability on the time axis. Abnormal bands may appear in a specific time period or moment. Therefore, by segmenting the EEG signals, the amount of data can be effectively reduced while more accurately capturing abnormal fluctuations in a short period of time. In addition, segmentation can also increase the robustness of the model and prevent inaccurate calculations due to long signals. For the segmentation of EEG signals, a common practice is to split them according to the time window. For example, first determine the time window size of each segment. Generally speaking, the window size should take into account the duration of abnormal activities such as epilepsy. It is generally selected as a time window of 1-5 seconds, and each window contains 1000 sampling points. Assuming the sampling frequency is 200Hz, the window sliding step determines the time interval of each segment. You can choose 50% overlap, that is, the step size is half of the window size, or no overlap. The overlapping method can improve the model's detection ability for small-amplitude abnormal fluctuations. Assume that the length of the EEG signal we collected is 20 seconds and the sampling frequency is 200Hz. The entire EEG signal contains 4000 data points. If we choose a 1-second window and a 50% overlap: each segment of data will contain 200 sampling points, and the window step is 100 sampling points. Therefore, each segment is segmented by sliding 100 sampling points each time. For a 20-second EEG signal, the segmentation results will be: 0-1 second (200 data points), 0.5-1.5 seconds (200 data points), 1-2 seconds (200 data points)... and so on.

[0043] The feature extraction unit is used to extract EEG segments containing abnormal bands. Specifically, the feature extraction unit uses multiple EEG signals to convert time domain signals into frequency domain signals through Fourier transform, thereby obtaining power characteristics of different frequency bands. Abnormal activities such as epilepsy are usually manifested as changes in specific frequency bands. For example, delta waves (0.5-4Hz) are usually associated with deep sleep or severe abnormalities, theta waves (4-8Hz) are increased in anxiety, mild abnormalities, and sleep, alpha waves (8-13Hz) are associated with relaxation and rest, and beta waves (13-30Hz) are associated with activity or concentration. Features such as the mean, standard deviation, peak value, and waveform morphology of the signal are extracted from the time domain. Through real-time detection of the segmented EEG signals, convolution operations are used to identify abnormal waveforms such as sharp waves, spike waves, and slow waves. Each EEG segment containing an abnormal band is extracted and stored as a new interval set. Each interval set should include the start time, end time, type of abnormal band, and extracted features.

[0044] The abnormality identification module also includes: a difference calculation unit, which is used to calculate the similarity of adjacent abnormal bands in the interval set, and normalize the power features of each frequency band extracted from the electroencephalogram signal. The purpose of power feature normalization is to convert the power features of different frequency bands into a dimensionless standardized value. The power features after normalization are represented as a feature vector. Each element in the feature vector represents the power feature of a different frequency band. By normalizing the power features of these different frequency bands, the abnormal feature vector of the abnormal band is obtained, and the Euclidean distance between the abnormal feature vectors of two adjacent abnormal bands is calculated. Assuming that the feature vector of the first abnormal band is [f1, f2, f3, ..., fn], and the feature vector of the second abnormal band is [f1', f2', f3', ..., fn'], the Euclidean distance calculation formula is:

[0045] d= , where d is the Euclidean distance between two abnormal bands and n is the total number of features in the feature vector;

[0046] The similarity of the two abnormal bands is judged by the Euclidean distance between them. According to the similarity in the previous time period, the mean-standard deviation method is used to set the similarity threshold. When the similarity of two adjacent abnormal bands is greater than or equal to the set similarity threshold, it indicates that the two adjacent abnormal bands belong to the same abnormal activity. At this time, the abnormal bands belonging to the same abnormal activity are merged into one abnormal event, and a new data set is generated to reduce redundant annotations. When the similarity of two adjacent abnormal bands is lower than the set similarity threshold, it indicates that the two adjacent abnormal bands belong to different abnormal events. Each abnormal band is regarded as an independent abnormal event and is annotated and analyzed separately.

[0047] By calculating the similarity between abnormal bands, we can intelligently determine whether different abnormal bands belong to the same abnormal event, thereby reducing redundant annotations and improving the quality of the data set. Suppose we have an EEG signal with a sampling frequency of 256Hz and a data length of 10 minutes. After preprocessing, the power features of each frequency band have been extracted. The power feature values ​​of each frequency band are normalized to obtain the abnormal feature vector of each band. Then, the Euclidean distance between two adjacent abnormal bands is calculated. The Euclidean distance between the intervals (15-20 seconds) and (21-25 seconds) is 0.022, and the Euclidean distance between the intervals (40-45 seconds) and (80-85 seconds) is 0.381. The Euclidean distance between the intervals (100-105 seconds) and (120-125 seconds) is 0.441. If the similarity threshold is set to 0.3, the Euclidean distance between 15-20 seconds and 21-25 seconds is 0.022, which is lower than the threshold of 0.3, indicating that the two bands are not similar and are considered independent abnormal events. The Euclidean distance between 40-45 seconds and 80-85 seconds is 0.381, which is higher than the threshold of 0.3, indicating that the two bands are similar and belong to the same abnormal activity, so they are merged into one abnormal event. The Euclidean distance between 100-105 seconds and 120-125 seconds is 0.441, which is higher than the threshold of 0.3, indicating that the two bands are similar and belong to the same abnormal activity, so they are merged into one abnormal event.

[0048] In the embodiment of the present invention, the frequency domain analysis and time domain feature extraction of the electroencephalogram signal are performed through Fourier transform and convolution operations, which can accurately identify abnormal waveforms such as sharp waves, spike waves and slow waves, which are usually closely related to epileptic seizures or other abnormal neural activities. The feature extraction unit combines multiple electroencephalogram signal parameters to enhance the system's sensitivity to subtle abnormal bands, thereby improving the recognition accuracy of abnormal bands. By calculating the similarity of adjacent abnormal bands and merging them, highly similar abnormal bands can be merged into a complete abnormal event, avoiding redundant annotations and repeated event annotations. This method reduces manual intervention in the annotation process, making annotation more efficient and the data set more concise. By setting a similarity threshold, the module can automatically distinguish similar bands and merge them into one abnormal event. When the similarity of adjacent abnormal bands is lower than When the threshold is reached, they are regarded as independent abnormal events and are labeled and analyzed separately. This mechanism improves the detection ability of complex abnormal patterns such as epileptic seizures, making the analysis process more accurate and reliable. The segmentation unit accurately segments the EEG through a sliding window mechanism and records the start and end time of each window to ensure the accurate synchronization of signal timestamps. This not only ensures the temporal consistency of data processing, but also provides an accurate time reference for the subsequent positioning of abnormal bands, so that the occurrence and duration of abnormal activities can be accurately recorded. By automatically identifying and analyzing abnormal bands in the EEG, this module can provide clinicians with more detailed and accurate diagnostic support for epilepsy or other neurological diseases. The accurate labeling of the data set and the accurate identification of abnormal events can provide strong data support for subsequent clinical research, treatment plan design and patient monitoring.

[0049] Furthermore, the labeling module is used to automatically label the identified abnormal bands, and the specific labeling states include normal, pre-ictal and epileptic seizure; the labeling module includes: a sample extraction unit, used to extract the power feature data of each frequency band in the EEG signal under different states in the EEG sample data set, and automatically extract EEG samples of different states such as normal, pre-ictal, epileptic seizure, etc. of patients of the corresponding age from the EEG sample data set according to the age of the patient to be labeled. The patient's age information is used as a screening criterion to ensure that the extracted samples are consistent with the typical EEG activity pattern of the age group, and the power features of each frequency band in the EEG signal under different states are extracted and normalized to obtain the sample feature vector.

[0050] First, the system receives the input EEG data to be labeled and the patient's age information. The system will automatically extract the sample data corresponding to the age group from the EEG dataset according to the patient's age, classify and label it, and judge the age group to which the patient belongs based on the input patient age information, including newborns, children, adolescents, adults and the elderly. The system will automatically extract the EEG samples corresponding to the age group from the dataset based on the input patient age, and the EEG samples have been labeled with different sample types according to age groups, such as "normal", "pre-ictal aura", "epileptic seizure", etc., to ensure that the extracted samples are consistent with the typical EEG activity pattern of the age group, and extract the power characteristics of each frequency band in the EEG signal under different states for normalization to obtain the sample feature vector.

[0051] The preliminary judgment unit is used to preliminarily judge the marking range and input the vector data after feature extraction, which contains the feature information of the abnormal feature vector and the sample feature vector. The feature difference is obtained by subtracting the features of each dimension item by item. The feature difference is used to represent the deviation between the power features of each frequency band in the abnormal band and the power features of each frequency band in the normal sample. The smaller the difference, the higher the state overlap. According to the feature difference in the previous time period, the mean-standard deviation method is used to pre-set the feature threshold. When the feature difference is less than the feature threshold, it indicates that the power features of each frequency band in the abnormal band are highly overlapped with the power features of each frequency band in the normal sample. At this time, the abnormal band is marked as the corresponding state in the EEG data set sample. If the feature difference is greater than or equal to the feature threshold, it indicates that the power features of each frequency band in the abnormal band are not highly overlapped with the power features of each frequency band in the normal sample, and the abnormal band is not marked.

[0052] By calculating the characteristic difference of the power characteristics of each frequency band between the abnormal band and the sample, the degree of deviation between the two can be accurately reflected. If the characteristic difference is small, it means that the power characteristics of the abnormal band are highly consistent with the sample. The system can more accurately match the abnormal band with the sample state, mark the abnormal band as the state corresponding to the sample, and improve the accuracy of the marking. After setting the characteristic threshold, when the characteristic difference exceeds the threshold, the system can automatically ignore those bands that are too different from the sample state. This processing method can avoid incorrectly marking irrelevant abnormal bands as a certain state, thereby enhancing the system's robustness to various complex situations and reducing the possibility of mislabeling.

[0053] The annotation module also includes: a constraint unit for further optimizing the initial annotation range and calculating the spectral entropy of the abnormal band EEG. The calculation process of the spectral entropy can be divided into: using Fourier transform to perform spectral analysis on the signal to obtain the power spectrum density of the signal, that is, the energy distribution of each frequency component, and normalizing the power spectrum density value so that the total energy of all frequency components is 1;

[0054] , where m is the number of frequencies in the spectrum, is the power value corresponding to frequency f in the spectrum, is the normalized power spectral density.

[0055] The formula for spectral entropy is as follows:

[0056] , where m is the number of all frequencies in the spectrum, H is the spectrum entropy, is the normalized power spectral density, log is the logarithmic function, usually the natural logarithm.

[0057] The higher the spectral entropy, the more complex the spectral components of the signal, that is, it contains more different frequency components; the lower the spectral entropy, the more concentrated and regular the spectral components in the signal are, which is usually seen in normal and stable EEG activity; in the premonitory stage of an epileptic seizure, the EEG shows a more chaotic high-frequency component, which causes the spectral entropy to gradually increase. In the epileptic seizure stage, the electrical activity of the brain will show strong irregularity, which leads to a further increase in the spectral entropy. The spectral entropy difference is obtained by subtracting the spectral entropy value of the abnormal band EEG from the spectral entropy under normal conditions. According to the spectral entropy difference in previous time periods, the mean-standard deviation method is used to set the frequency The threshold of the spectral entropy difference is 0.2. The ratio difference of some data in the abnormal feature vector is calculated, including the ratio difference of α wave and β wave, and the ratio difference of θ wave and α wave. Specifically, the ratio difference of α wave and β wave in the EEG sample data set under normal circumstances is calculated and the ratio of α wave and β wave in the abnormal feature vector data is taken as the difference, and the difference is taken to obtain the ratio difference of α wave and β wave. Similarly, the ratio difference of θ wave and α wave is calculated, and the annotation range is further optimized based on the preliminary annotation. Specifically, when the α wave / β wave ratio difference is greater than 20% and the spectral entropy difference is greater than 0.2, it is confirmed that the signal belongs to an epileptic seizure state and is labeled accordingly. When the θ wave / α wave ratio difference is greater than 15% and the spectral entropy difference is greater than 0.1, it is confirmed that the signal belongs to a pre-epileptic aura and the annotation range is further refined.

[0058] In the embodiment of the present invention, the sample extraction unit screens relevant EEG samples according to the patient's age to ensure that the extracted samples can reflect the typical EEG activity pattern of the age group. This approach greatly improves the accuracy of annotation. Because there are significant differences in EEG activity patterns between different age groups, the introduction of age information makes the annotation of abnormal bands more accurate. The initial judgment unit can quickly judge the possible state of the abnormal band by calculating the feature difference between the abnormal band and the normal sample. This method sets a feature threshold and only annotates the bands that meet the requirements, thereby improving the efficiency of annotation, reducing the need for manual intervention, and saving a lot of time and resources. The constraint unit further optimizes the annotation result by calculating the spectral entropy and the spectral entropy difference. The spectral entropy reflects the complexity of the EEG signal and is very useful in the pre-epileptic stage. and seizure stages, the signal shows a higher spectral entropy, which helps to accurately distinguish the normal state, premonitory state and epileptic seizure state. By setting reasonable thresholds and further refining them in combination with ratio differences, the system can more accurately identify and label different epilepsy stages. The system optimizes the labeling through multi-dimensional feature differences such as spectral entropy difference and ratio differences in different bands, and can more accurately distinguish different stages of epilepsy, seizures, premonitory and normal. This multi-level labeling method effectively improves the predictive ability of epilepsy, helps to identify the risk of epileptic seizures early, and thus provides a reliable basis for clinical treatment. By multiple optimizations of the labeling results, the high quality of the labeled data is ensured. Accurate labeling makes the EEG dataset more representative and reliable, thereby improving the predictive ability and practical application value of the analysis model.

[0059] Furthermore, a correction module is used to correct the annotated abnormal bands and establish a correction mechanism; the correction module includes: a verification unit, used to verify the annotated abnormal bands, use the wavelet transform method to perform time-frequency analysis on the EEG signal, generate a time-frequency graph, and obtain the frequency distribution corresponding to each time period. According to the analysis results of the time-frequency graph, if the annotated abnormal band has a time duration error, the annotated time range can be dynamically adjusted based on the instantaneous frequency change, and by analyzing the frequency band characteristics of the time-frequency graph, it is checked whether there is an erroneous frequency band allocation, and then the annotated frequency range is adjusted.

[0060] By introducing the verification unit and wavelet transform time-frequency analysis method, the annotated abnormal bands can be dynamically checked and adjusted to ensure the accuracy of the annotation results. Specifically, the correction module identifies and corrects potential errors in the annotation process through the analysis of the time-frequency diagram, thereby improving the reliability and practical application value of the system. First, the EEG signal is analyzed by wavelet transform to generate a time-frequency diagram. This process combines the time and frequency information of the signal and provides the frequency distribution of the signal in different time periods, so that the system can intuitively identify the detailed characteristics of the signal in the time and frequency domain. Based on the analysis of the time-frequency diagram, the verification unit can The time length in the annotation is dynamically adjusted. If the anomaly band is marked with an error of being too long or too short in duration, the system can optimize the annotated time range through instantaneous frequency changes to ensure that the time window of each band is consistent with the actual EEG activity. During the verification process, the verification unit will further analyze the frequency band characteristics of the time-frequency graph to check whether there is an incorrect frequency band allocation. For example, the abnormal band may be mistakenly marked in an inappropriate frequency band. The system identifies these errors by analyzing the frequency distribution in the time-frequency graph and automatically adjusts the annotated frequency range to make it consistent with the actual frequency characteristics of the abnormal band.

[0061] The correction module also includes: an adjustment unit, which is used to adjust the marking range of the abnormal band, compare the automatically marked abnormal band with the actual time-frequency characteristics, and if it is found that the marking range is inaccurate, fine-tune and expand the boundary. For example, for the marking of epileptic seizures, when it is found that the abnormal frequency fluctuation occurs in a short period of time, but the automatic marking is too long, the adjustment unit can narrow the time window.

[0062] In the embodiment of the present invention, the correction module can accurately verify and correct the annotation of abnormal bands in the electroencephalogram signal by combining time-frequency analysis with wavelet transform. This method allows the time range and frequency range of the abnormal band to be accurately adjusted, reducing the errors that may be caused by manual annotation. Through real-time analysis of the time-frequency graph, the module can dynamically identify and correct the duration and frequency band allocation problems of the annotation to ensure the accuracy and reliability of the annotation data. The verification unit and adjustment unit in the correction module enable the system to have a powerful dynamic adjustment capability. When it is detected that the time range in the annotation is too long or the frequency range is wrong, the system will automatically adjust according to the instantaneous frequency change of the time-frequency graph to ensure that the annotation time of the abnormal band is consistent with the actual brain For example, when labeling epileptic seizures, if the frequency of abnormal fluctuations changes suddenly in a short period of time and the labeled time window is too long, the adjustment unit will shrink or expand it to ensure that the characteristics of the seizure are accurately reflected. This dynamic adjustment capability greatly improves the precision of the labeling. Through detailed analysis of the frequency band characteristics in the time-frequency diagram, the correction module can identify and correct the frequency band allocation errors that may occur during the labeling process. The abnormal band may be mistakenly labeled in an inappropriate frequency band, and the correction module will automatically adjust the labeled frequency range based on the frequency distribution analysis to ensure that the labeling of each band is consistent with the actual EEG activity. This optimization process reduces the possibility of mislabeling and improves the reliability of the overall labeling.

[0063] Furthermore, the automatic labeling system for abnormal bands in epileptic EEG images also includes: a manual verification module for manually checking the labeled EEG, and the manual verification module includes: a correction unit for manually correcting the automatic labeling results, in which experts verify the automatic labeling results and modify erroneous labels, and provide misjudgment type options for manual selection, including artifact interference (such as electromyography, blinking, etc.), atypical abnormal morphology, misjudgment of abnormal time length (too short / too long), unprocessed baseline drift, and others, to provide structured data for error analysis;

[0064] The feedback learning unit is used to use the manually modified EEG as feedback sample data to retrain the annotation module to obtain an updated annotation module.

[0065] In the embodiment of the present invention, firstly, the introduction of the correction unit provides a professional manual correction mechanism for the annotation results. In the process of automatic annotation, the system may make misjudgments, such as artifact interference (such as electromyography, blinking, etc.), atypical abnormal morphology, misjudgment of abnormal time length (too short or too long), unprocessed baseline drift, etc. Through expert verification and modification, manual correction can correct these errors and improve the accuracy of the annotation results. This process not only solves the defects of automatic annotation, but also provides misjudgment type options, further structures the error data, lays the foundation for subsequent error analysis, and helps the annotation system discover its shortcomings and perform targeted optimization; secondly, the feedback learning unit uses the manually modified EEG as a feedback sample Data promotes the retraining and updating of the annotation module. As experts continuously provide feedback samples, the annotation module can gradually improve its accuracy and ability to recognize complex situations after each training, forming a virtuous feedback loop, which not only accelerates the self-optimization of the annotation system, but also enables the system to better perform automatic annotation when facing new or complex EEG data. Overall, the manual verification module provides strong support for the improvement of the automated annotation system while ensuring the quality of annotation through a combination of manual and machine methods. Its main effects include improving annotation accuracy, reducing the misjudgment rate, enhancing the adaptability and scalability of the system, and promoting the continuous optimization of the annotation module, ultimately improving the efficiency and reliability of EEG analysis.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. Automatic labeling system for abnormal bands in epileptic EEG images, characterized by: include: The data acquisition module is used to collect EEG data sets, including original EEG signals and corresponding annotation labels, and obtain EEG sample data sets; The signal preprocessing module is used to preprocess the input EEG signal to be labeled, perform noise removal, filtering, and standardization; The abnormality identification module is used to randomly segment the pre-processed EEG signal to be labeled, identify abnormal bands in the input EEG signal to be labeled based on a normal EEG data set and multiple EEG signals, retain the EEG segments containing abnormal bands, generate a new interval set, and calculate the similarity of adjacent abnormal bands in the interval set. Based on the similarity of adjacent abnormal bands, the EEG segments containing the same abnormal band are merged; The labeling module is used to automatically label the identified abnormal bands, with specific labeling status including normal, preictal aura and epileptic seizure; The correction module is used to correct the anomaly bands that have been marked and establish a correction mechanism; The annotation module includes: The sample extraction unit is used to extract the power feature data of each frequency band in the EEG signal under different states in the EEG sample data set. According to the age of the patient to be annotated, the EEG samples of normal, preictal, and epileptic seizure states of the patient of the corresponding age are automatically extracted from the EEG sample data set. The patient's age information is used as a screening criterion. The power features of each frequency band in the EEG signal under different states are extracted and normalized to obtain the sample feature vector. The preliminary judgment unit is used to preliminarily judge the marking range and input the vector data after feature extraction, which contains the feature information of the abnormal feature vector and the sample feature vector. By subtracting the features of each dimension item by item, the feature difference is obtained. The feature difference is used to represent the deviation between the power characteristics of each frequency band in the abnormal band and the power characteristics of each frequency band in the normal sample, and the feature threshold is set in advance. When the feature difference is less than the feature threshold, it indicates that the power characteristics of each frequency band in the abnormal band are highly overlapped with the power characteristics of each frequency band in the normal sample. At this time, the abnormal band is marked as the corresponding state in the EEG data set sample. If the feature difference is greater than or equal to the feature threshold, it indicates that the power characteristics of each frequency band in the abnormal band are not highly overlapped with the power characteristics of each frequency band in the normal sample, and the abnormal band is not marked.

2. The automatic labeling system for abnormal bands in epileptic EEG images according to claim 1 is characterized by: The data acquisition module includes: The age classification unit is used to preliminarily classify the collected EEG dataset according to age. For each EEG sample, the age classification unit uses age as an additional label and divides the EEG dataset into four main categories according to the patient's age, namely newborn, child, adolescent and adult datasets; A state classification unit is used to perform secondary division on the EEG data sets of normal, preictal, and epileptic seizure states in each age category to obtain an EEG sample data set, and store each EEG sample data set together with its corresponding age group and state classification label; The signal preprocessing module is used to smooth and denoise the EEG signal and eliminate low-quality signals and abnormal waveforms.

3. The automatic marking system for abnormal bands in epileptic EEG images according to claim 1 is characterized by: The anomaly recognition module includes: The segmentation unit is used to segment the pre-processed EEG signal to be labeled, split it according to the time window, determine the time window size of each segment, and segment the EEG signal according to the window sliding step size and overlap method. During the sliding window process, the start and end time of the EEG signal contained in each window needs to be accurately recorded and synchronized to ensure that the signal timestamp remains consistent; The feature extraction unit is used to extract EEG segments containing abnormal bands. Through multiple EEG signals, each EEG segment containing abnormal bands is extracted and stored as a new interval set. Each interval set contains the start time, end time, type of abnormal band and extracted features.

4. The automatic marking system for abnormal bands in epileptic EEG images according to claim 3 is characterized by: The anomaly recognition module also includes: The difference calculation unit is used to calculate the similarity of adjacent abnormal bands in the interval set, extract the power features of each frequency band from the electroencephalogram signal and perform normalization processing to obtain the abnormal feature vector of the abnormal band, calculate the Euclidean distance between the abnormal feature vectors of two adjacent abnormal bands, and judge their similarity through the Euclidean distance of the two abnormal bands, set a similarity threshold, and when the similarity of two adjacent abnormal bands is greater than or equal to the set similarity threshold, it indicates that the two adjacent abnormal bands currently belong to the same abnormal activity. At this time, the abnormal bands belonging to the same abnormal activity are merged into one abnormal event and a new data set is generated; when the similarity of two adjacent abnormal bands is lower than the set similarity threshold, it indicates that the two adjacent abnormal bands currently belong to different abnormal events. Each abnormal band is regarded as an independent abnormal event and is labeled and analyzed separately.

5. The automatic marking system for abnormal bands in epileptic EEG images according to claim 1 is characterized by: The annotation module also includes: The constraint unit is used to further optimize the preliminary annotation range, calculate the spectral entropy of the abnormal band EEG, obtain the spectral entropy difference by subtracting the spectral entropy value of the abnormal band EEG from the spectral entropy under normal conditions, set the threshold of the spectral entropy difference, calculate the ratio difference of some data in the abnormal feature vector, and further optimize the annotation range based on the preliminary annotation.

6. The automatic marking system for abnormal bands in epileptic EEG images according to claim 1 is characterized by: The correction module includes: The verification unit is used to verify the annotation of the abnormal bands that have been marked, use the wavelet transform method to perform time-frequency analysis on the EEG signal, generate a time-frequency diagram, and obtain the frequency distribution corresponding to each time period. According to the analysis results of the time-frequency diagram, if there is an error in the time length of the marked abnormal band, the marked time range will be dynamically adjusted based on the instantaneous frequency change, and by analyzing the frequency band characteristics of the time-frequency diagram, it will be checked whether there is an incorrect frequency band allocation, and then the marked frequency range will be adjusted.

7. The automatic marking system for abnormal bands in epileptic EEG images according to claim 6, characterized in that: The correction module also includes: The adjustment unit is used to adjust the annotated range of the abnormal band and compare the automatically annotated abnormal band with the actual time-frequency characteristics. If the annotated range is found to be inaccurate, the boundary is fine-tuned and expanded.

Images