A biological tissue electric response signal data enhancement method based on content and style decoupling and statistical constraint
Patent Information
- Application Number
- CN202610926380.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-25
- Publication Date
- 2026-09-15
Smart Images
Figure CN122762004A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of biomedical signal processing, computer-aided analysis, and data augmentation, specifically to a method for data augmentation of biological tissue electrical response signals based on content and style decoupling and statistical constraints. This method is applicable to scenarios involving biological tissue electrical detection, medical signal analysis, intelligent diagnosis, and related machine learning modeling. Background Technology
[0002] By applying electrical stimulation or electric field excitation to biological tissues and collecting the voltage, current, impedance, or other electrical responses generated by the tissues, response information related to the tissue's internal structure, water content, cell arrangement, pathological state, and physiological state can be obtained. Different tissue types typically exhibit different response patterns under the same excitation conditions; therefore, the electrical response signals of biological tissues can be used for tissue classification, state assessment, and medical auxiliary analysis.
[0003] In practical research and engineering applications, the acquisition, preservation, and measurement of biological tissue samples are often costly, especially in in vivo experiments, ex vivo tissue experiments, or scenarios where rare pathological samples are scarce, limiting the number of effective samples available for training machine learning models. Furthermore, different samples within the same tissue category can exhibit intra-class variability due to individual differences, tissue state, sensor contact conditions, ambient temperature, and measurement noise. Insufficient sample size and intra-class variability together make the trained model prone to overfitting and reduce its generalization ability on new samples.
[0004] Existing data augmentation methods mainly include random noise addition, amplitude scaling, time or frequency axis transformation, sample interpolation, and data synthesis based on generative models. Random perturbation methods are simple to implement, but may disrupt the structural relationships of signals in the frequency and time dimensions; interpolation methods typically generate data limited to existing samples, making it difficult to reflect true intra-class differences; generative models often rely on a large amount of training data, and the stability and interpretability of the generated results are insufficient in small sample scenarios.
[0005] Existing methods typically lack explicit distinction between stable response patterns of organizational categories and within-class biases, and also lack mechanisms to limit the intensity of augmentation based on the statistical distribution of samples within the same class. Therefore, a data augmentation method is needed that can preserve stable response patterns of organizational categories, adjust for within-class differences, and limit augmented samples to the reasonable statistical distribution boundaries of the original samples. Summary of the Invention
[0006] To address the challenges of effectively expanding training data for existing biological tissue electrophysiological response signals in small-sample scenarios, and the tendency of traditional data augmentation methods to disrupt the inherent statistical structure of signals and lack physiological rationality, this patent proposes a style-content decoupling data augmentation method for biological tissue electrophysiological response signals. This method leverages the statistical characteristics of biological tissue electrophysiological responses, decomposing the electrophysiological response patterns of the same tissue type at specific excitation frequencies and time scales into stable content features characterizing the essential attributes of the tissue and variable style features reflecting differences between samples. While preserving the inherent electrophysiological characteristics of the tissue as much as possible, diverse augmented samples are generated through style transfer with applied statistical constraints, thereby improving data utilization efficiency and reducing the risk of overfitting in subsequent modeling processes.
[0007] The core idea of this patent is as follows: Under the same excitation conditions, the electrical response signals of biological tissues of the same type exhibit relatively stable statistical response characteristics in both frequency and time dimensions. These statistical response characteristics reflect the inherent electrical response attributes of the tissue and can be used as content features. However, different samples of the same type of tissue, influenced by individual differences, tissue state, measurement conditions, and environmental factors, show deviations in their electrical response signals relative to these content features. These deviations reflect the differences between samples of the same type and can be used as style features. By decoupling and modeling content features and style features, and achieving controllable adjustment of style features, efficient data augmentation can be achieved while reducing false discrimination information and maintaining physiological rationality.
[0008] This patent primarily discloses a method for enhancing the electrical response signal data of biological tissues based on content and style decoupling and statistical constraints. It acquires raw biological tissue electrical response data using an active electric field sensing system. The active electric field sensing system includes a sensor probe, a signal generation module, a signal acquisition module, and a data processing module. The sensor probe has a columnar insulated structure, and its detection end is equipped with a square-arranged four-electrode array. Two diagonal electrodes form a transmitting electrode pair, and the other two diagonal electrodes form a receiving electrode pair. The transmitting electrodes are connected to the signal generation module to apply a sinusoidal electric field excitation signal to the biological tissue under test. The receiving electrodes are connected to the signal acquisition module to acquire the electrical response signal generated by the biological tissue under the influence of the electric field. The specific steps are as follows:
[0009] S100: Acquire electrical response signals from multiple biological tissue samples of the same biological tissue category at multiple excitation frequency points and multiple sampling time points, collected by an electrical response signal acquisition device under a preset electric field excitation condition; preprocess and standardize the electrical response signals to obtain... A multidimensional matrix, where This indicates the number of frequency points contained in the data. This indicates the number of time points included in the data. This represents the number of samples. The standardized electrical response eigenvalue matrix is then... ; Indicates sample The standardized electrical response eigenvalue matrix, ; For the sample In the Frequency point The amplitude of the biological tissue feedback electrical response signal collected at each time point, , .
[0010] S200: Under the same frequency-time conditions, the standardized electrical response eigenvalues of the sample to be augmented are statistically calculated to obtain content features, which are defined as: And form a content feature matrix.
[0011] S300: Based on the aforementioned content features, the sample of the object being tested... At the corresponding frequency-time point, the deviation ratio of its electrical response value relative to the content feature is calculated, and this deviation ratio is defined as the style feature parameter, defined as: .
[0012] S400: While keeping the content features unchanged, introduce a style transfer coefficient. The style features are modulated to obtain candidate style features: To ensure the consistency of statistical distribution between the enhanced sample and the original sample, the style transfer coefficient is calculated. The range of values. For each frequency-time point. Calculate the sample variance among the electrical response values of all samples at that location. Calculate the samples separately. Frequency – Time Point The residual is The residual after style transfer is Then define the transferred sample The square of the Mahalanobis distance is: .in: Indicates sample The square of the Mahalanobis distance.
[0013] S500: To prevent generated samples from deviating from the original data distribution, a statistical safety threshold is set. The threshold is taken from the upper bound of the chi-square distribution at a preset confidence level. The square of the transferred Mahalanobis distance is required to satisfy: This yields the sample. At frequency-time point The upper bound of the style transfer coefficient: Calculate the corresponding frequency-time point for all samples. The minimum value among them is taken as the upper limit of the global style transfer coefficient: Thus, the style transfer coefficient is determined. The safe range of values.
[0014] S600: In the style transfer coefficient Under the condition of meeting the above safety range, the coefficients are selected using a preset sampling strategy. Then the sample The data augmentation samples are: ,in By repeating the selection and modulation process described above for all samples, the dataset can be expanded.
[0015] Compared with the prior art, the beneficial effects of the present invention are:
[0016] (1) The present invention decouples the content features of similar biological tissue samples from the style features of individual samples, and keeps the content features unchanged during data augmentation, which helps to avoid disrupting the stable response patterns of tissue categories.
[0017] (2) The present invention constrains the style transfer coefficient by Mahalanobis distance and statistical safety threshold, so that the generation intensity of the enhanced sample has a calculable statistical boundary, thereby improving the interpretability and controllability of the enhancement process.
[0018] (3) The present invention can effectively expand the scale of training data under small sample conditions, improve sample diversity, and reduce the risk of introducing false discrimination information. It helps to improve the generalization ability of subsequent modeling or classification and recognition tasks and reduce the risk of model overfitting.
[0019] (4) This invention is based on frequency-time structure modeling, which makes full use of the multidimensional structural information of biological tissue electrical response signals, which is beneficial to improving feature expression ability and downstream analysis accuracy.
[0020] (5) The method of the present invention does not rely on complex deep learning generation models. The calculation process is clear, the implementation cost is low, and it is easy to deploy in actual biomedical detection systems. It has good engineering application value and promotion prospects. Attached Figure Description
[0021] Figure 1 This is a flowchart of the data augmentation method of the present invention.
[0022] Figure 2 This is a schematic diagram illustrating the relationship between the content features, style features, and enhanced samples of this invention.
[0023] Figure 3 This is a schematic diagram of the sensor probe structure of the present invention.
[0024] Figure 4 This is a schematic diagram illustrating the implementation of the data acquisition scheme of the present invention. Detailed Implementation
[0025] The present invention will be further described in detail below with reference to embodiments. However, it should be understood that these descriptions are merely examples of the application of the present invention and are not intended to limit the scope of the above-mentioned subject matter of the present invention. All technologies implemented based on the content of this patent fall within the scope of this patent.
[0026] In this embodiment, taking a small sample of mouse tumor tissue voltage response signal data as an example, the active electric field sensing system used in this embodiment includes a sensor probe, a signal generator, a data acquisition card, and a host computer. The sensor probe has a columnar insulated structure, and the detection end is equipped with a square-arranged four-electrode array. The diagonal electrodes constitute the transmitting electrode pair and the receiving electrode pair, respectively. Its structural schematic diagram is shown below. Figure 3 As shown in the diagram. The probe is mounted on the Z-axis of the three-axis moving platform to control the contact position and depth between the probe and the tissue being measured. A schematic diagram of the data acquisition scheme is shown below. Figure 4 As shown, the sensor probe contacts the biological tissue for detection. The transmitting electrode is connected to a signal generator to apply an alternating electric field excitation to the tissue under test; the receiving electrode is connected to a data acquisition card to acquire the tissue response signal. The sampling rate of the data acquisition card is set to 20480Hz.
[0027] The experiment used 30 groups of tumor-bearing mouse tissue samples, each group including tumor tissue and normal tissue. The tissue samples were placed in a 0.9% sodium chloride solution for measurement, ensuring complete immersion. During the measurement, the excitation signal frequency range was set to 100–1500 Hz, with a frequency step size of 100 Hz, for a total of 15 frequency points. The response signal was continuously acquired for 1 second at each frequency point. Measurements were performed every 10 minutes starting from 0 min, and continued for 50 min to obtain tissue voltage data at 6 time points. Taking the tumor tissue electrical response information matrix as an example, 20 samples were randomly selected for data augmentation as the training set for subsequent classification model training, and the remaining 10 samples were used as the test set. The data augmentation steps are as follows:
[0028] S100: Using the above active electric field sensing system and acquisition method, voltage signal amplitudes at 15 frequencies and 6 time points after in vitro sampling were acquired from 30 tissue samples. The acquired electrical response signals were preprocessed, including filtering and denoising, to obtain 15×6×30 electrical response information matrices for tumor and normal tissues, respectively. Here, 15 represents the number of frequency points included in the data, 6 represents the number of time points included in the data, and 30 represents the number of samples. The samples... The 15×6 normalized electrical response eigenvalue matrix is ; Indicates sample The standardized electrical response eigenvalue matrix, ; For the sample In the Frequency point The amplitude of the biological tissue feedback voltage signal collected at each time point. , Twenty samples of tumor tissue were randomly selected for data augmentation, and the remaining 10 samples were used as the test set. The test set samples were not used for standardized parameter calculation, content feature extraction, style feature statistics, statistical safety threshold calculation, or augmentation sample generation.
[0029] S200: Under the same frequency-time point conditions, the standardized voltage characteristic values of 20 tumor tissues to be augmented are statistically calculated to obtain content features, which are defined as: And form a 15×6 content feature matrix. .
[0030] S300: Based on content feature matrix Samples of 20 tumor tissues treated with data augmentation At the corresponding frequency-time point, the deviation ratio of its voltage value relative to the content feature is calculated, and this deviation ratio is defined as the style feature parameter, defined as: This yields a style feature matrix of 15×6×20.
[0031] S400: While keeping the content features unchanged, introduce a style transfer coefficient. The style features are modulated to obtain candidate style features: To ensure the consistency of statistical distribution between the enhanced sample and the original sample, the style transfer coefficient is calculated. The range of values. For each frequency-time point. Calculate the variance among the electrical response values of all samples at that location. Calculate the samples separately. Frequency – Time Point The residual is The residual after style transfer is Then define the transferred sample The square of the Mahalanobis distance is: .in: Indicates sample The square of the Mahalanobis distance.
[0032] S500: To prevent generated samples from deviating from the original data distribution, a statistical safety threshold is set. The threshold is set to a chi-square distribution with 1 degree of freedom (in this embodiment, a frequency-time point independent constraint strategy is adopted, so each calculation of the Mahalanobis distance corresponds to only a single style feature variable, and its degree of freedom is 1), at a 99% confidence level. The transferred Mahalanobis distance is required to satisfy: This yields the sample. At frequency-time point The upper bound of the style transfer coefficient: Calculate the corresponding frequency-time point for all samples. The minimum value among them is taken as the upper limit of the global style transfer coefficient: In this embodiment Thus, the safe range of values for the style transfer coefficient is determined. It can also be reduced as needed. Range of values.
[0033] S600: In the style transfer coefficient Meets safety range Under the given conditions, coefficients are selected using a preset sampling strategy. Then the sample The data augmentation samples are: ,in The diagram illustrating the relationship between content features, style features, and enhanced samples is shown below. Figure 2 As shown, enhanced samples are obtained by uniformly and randomly sampling the style transfer coefficients 10 times within the safe range and then modulating them. The above selection and modulation process is repeated on 20 samples to be enhanced to generate 200 enhanced samples, which together with the original 20 samples constitute an expanded training set of 220 samples.
[0034] Similarly, normal tissue samples can be augmented using the above steps to obtain an expanded normal tissue dataset.
[0035] The support vector machine classification model was trained using the original and extended datasets from two different organizations, with the remaining 10 original samples used as an independent test set. To reduce the impact of randomness in sample partitioning on the experimental results, a random hold-out validation strategy was adopted. The above process was repeated 10 times, with each random partition consisting of 20 and 10 samples. The average of the results from each iteration was used as the final evaluation result.
[0036] Throughout the experiment, both the original and extended datasets used the same feature dimensions, classifier structure, and parameter settings, and maintained a consistent training and test set partitioning strategy to ensure the comparability of the experimental results. Tables 1 and 2 present the comparison results of the classification performance of the support vector machine model under the two dataset conditions. As can be seen from Tables 1 and 2, after introducing augmented samples generated based on the style-content decoupling method, the classification accuracy, recall, and generalization ability of the support vector machine model on the test set were significantly improved. The experimental results show that the data augmentation method proposed in this invention can effectively expand the small-sample biological tissue electrical response signal dataset, improve the training effect and generalization performance of the classification model, thus verifying the effectiveness and practical value of this method in the scenario of small-sample biological tissue electrical response signal analysis.
[0037] Table 1 Performance of SVM classification model on raw data
[0038] linear kernel 70.64% 71.53% 69.35% 0.70 0.79 polynomial kernel 73.66% 73.46% 76.07% 0.74 0.81 Radial base core 72.41% 69.96% 78.61% 0.74 0.81
[0039] Table 2 Performance of Data Augmented SVM Classification Model
[0040] linear kernel 73.59% 74.25% 72.69% 0.73 0.82 polynomial kernel 77.78% 77.19% 81.86% 0.78 0.86 Radial base core 75.93% 73.55% 81.16% 0.77 0.84
Claims
1. A method for enhancing biotissue electrical response signal data based on content and style decoupling and statistical constraints, characterized in that, Includes the following steps: S100: Acquire electrical response signals from multiple biological tissue samples of the same biological tissue category at multiple excitation frequency points and multiple sampling time points, collected by an electrical response signal acquisition device under a preset electric field excitation condition; preprocess and standardize the electrical response signals to obtain... A multidimensional matrix, where This indicates the number of frequency points contained in the data. This indicates the number of time points included in the data. This represents the number of samples. The standardized electrical response eigenvalue matrix is then... ; Indicates sample The standardized electrical response eigenvalue matrix, ; For the sample In the Frequency point The amplitude of the biological tissue feedback electrical response signal collected at each time point, , . S200: Under the same frequency-time conditions, the standardized electrical response eigenvalues of the sample to be augmented are statistically calculated to obtain content features, which are defined as: And form a content feature matrix. S300: Based on the aforementioned content features, the sample of the object being tested... At the corresponding frequency-time point, the deviation ratio of its electrical response value relative to the content feature is calculated, and this deviation ratio is defined as the style feature parameter, defined as: . S400: While keeping the content features unchanged, introduce a style transfer coefficient. The style features are modulated to obtain candidate style features: To ensure the consistency of statistical distribution between the enhanced sample and the original sample, the style transfer coefficient is calculated. The range of values. For each frequency-time point. Calculate the sample variance among the electrical response values of all samples at that location. Calculate the samples separately. Frequency – Time Point The residual is The residual after style transfer is Then define the transferred sample The square of the Mahalanobis distance is: .in: Indicates sample The square of the Mahalanobis distance. S500: To prevent generated samples from deviating from the original data distribution, a statistical safety threshold is set. The threshold is taken from the upper bound of the chi-square distribution at a preset confidence level. The square of the transferred Mahalanobis distance is required to satisfy: This yields the sample. At frequency-time point The upper bound of the style transfer coefficient: Calculate the corresponding frequency-time point for all samples. The minimum value among them is taken as the upper limit of the global style transfer coefficient: Thus, the style transfer coefficient is determined. The safe range of values. S600: In the style transfer coefficient Under the condition of meeting the above safety range, the coefficients are selected using a preset sampling strategy. Then the sample The data augmentation samples are: ,in By repeating the selection and modulation process described above for all samples, the dataset can be expanded.
2. The method according to claim 1, characterized in that, The electrical response signal is acquired through an active electric field sensing system. The active electric field sensing system includes a sensor probe, a signal generation module, a signal acquisition module, and a data processing module. The signal generation module is used to apply an alternating electric field excitation signal to the biological tissue under test through the sensor probe. The signal acquisition module is used to acquire the electrical response signal generated by the biological tissue under test under the action of the alternating electric field excitation signal through the sensor probe. The data processing module is used to preprocess and standardize the electrical response signal.
3. The method according to claim 2, characterized in that, The sensor probe has a columnar insulating structure. The detection end of the sensor probe is equipped with a square-arranged four-electrode array, in which two diagonal electrodes form a transmitting electrode pair and the other two diagonal electrodes form a receiving electrode pair. The transmitting electrode pair is connected to the signal generating module, and the receiving electrode pair is connected to the signal acquisition module.
4. The method according to claim 2, characterized in that, The AC electric field excitation signal is a sinusoidal excitation signal; the signal generation module outputs the sinusoidal excitation signal sequentially according to multiple preset excitation frequencies, and the signal acquisition module acquires the electrical response signal of the biological tissue to be tested at multiple preset sampling time points.
5. The method according to claim 4, characterized in that, The electrical response features include voltage, current, impedance, phase, or other electrical response signal features that can construct a two-dimensional feature matrix in both frequency and time dimensions.
6. The method according to claim 1, characterized in that, The preprocessing includes at least one of filtering, baseline correction, outlier removal, and denoising; the standardization includes at least one of maximum and minimum value standardization, baseline standardization, and amplitude normalization, and the standardization parameters used for the standardization are determined based on the original training samples.
7. The method according to claim 1, characterized in that, The content features are statistical benchmark values of standardized electrical response feature values of multiple samples under the same biological tissue category at corresponding excitation frequency points and sampling time points.
8. The method according to claim 1, characterized in that, The style feature is the relative deviation of the sample's standardized electrical response feature value relative to the content feature.
9. The method according to claim 1, characterized in that, The style transfer coefficient The statistical safety threshold is determined by the Mahalanobis distance of the data. and chi-square threshold Confirmed, satisfied The chi-square threshold The upper quantile of the chi-square distribution at a predefined confidence level is determined based on the style feature dimension used to calculate the Mahalanobis distance.
10. The method according to claim 1, characterized in that, The style transfer coefficient By employing uniform random sampling, truncated normal distribution sampling, pre-defined discrete step size sampling, or sampling methods based on the target augmentation quantity, within a safe range... Internal determination.