A label-free optical detection method for myocardial tissue electrophysiology based on deep learning

Abstract

The present invention relates to the field of medical health technology, and in particular to a label-free optical detection method for electrophysiological myocardial tissue based on deep learning. The present invention comprises the following steps: acquiring optical detection image data and electrophysiological signal data of myocardial tissue; performing image preprocessing on the optical detection image data and electrophysiological signal data to obtain optical-physiological signal mapping data, wherein the image preprocessing includes image registration, image cropping, periodic enhancement, time difference enhancement, spatial filtering, intensity normalization, and data smoothing; performing feature signal extraction based on mechanical contraction curves and electrophysiological curves according to the optical-physiological signal mapping data to obtain mechanical contraction curve data and electrophysiological curve data. The present invention adopts label-free optical detection, which avoids direct interference with myocardial tissue and better maintains the physiological state of the tissue.

Description

Technical Field

[0001] The present invention relates to the field of medical health technology, and in particular to a label-free optical detection method for myocardial tissue electrophysiology based on deep learning. Background Art

[0002] The electrocardiogram method is generally used to obtain myocardial tissue electrophysiological data clinically. This surface electrode detection method is contact-based and has low spatiotemporal resolution, and is not suitable for high-throughput, high-resolution applications such as cardiac drug screening and toxicity testing. Compared with electrode detection, optical electrophysiological detection methods have the advantages of non-contact and high spatiotemporal resolution. However, the traditional optical detection process relies on fluorescent markers (to mark cell membrane action potentials or intracellular calcium ion concentrations), which are highly toxic to the test samples and are not conducive to long-term observations. Label-free optical detection of myocardial tissue electrophysiology is a technology used for non-invasive monitoring and evaluation of myocardial electrophysiological activity, and is particularly suitable for the diagnosis, research and treatment evaluation of heart diseases. By combining optical imaging and deep learning models, this technology can infer electrophysiological signals from the optical properties of myocardial tissue without the need to implant electrodes or use chemical labels.

[0003] However, traditional electrophysiological optical detection of myocardial tissue often has the following problems: in traditional fluorescence detection, fluorescent dyes will gradually inactivate after long-term exposure to excitation light, affecting signal quality; traditional electrode detection methods cause physical damage to cells; and dyes may be internalized by cells, resulting in signal attenuation and inaccurate positioning. Summary of the Invention

[0004] Based on this, it is necessary for the present invention to provide a label-free optical detection method for myocardial tissue electrophysiology based on deep learning to solve at least one of the above technical problems.

[0005] To achieve the above objectives, a label-free optical detection method for myocardial tissue electrophysiology based on deep learning is provided, comprising the following steps:

[0006] Step S1: Acquire optical detection image data and electrophysiological signal data of myocardial tissue; perform image preprocessing on the optical detection image data and electrophysiological signal data to obtain optical-physiological signal mapping data, wherein the image preprocessing includes image registration, image cropping, periodic enhancement, time difference enhancement, spatial filtering, intensity normalization, and data smoothing;

[0007] Step S2: extracting characteristic signals based on the mechanical contraction curve and the electrophysiological curve according to the optical-physiological signal mapping data, thereby obtaining mechanical contraction curve data and electrophysiological curve data;

[0008] Step S3: Using the mechanical contraction curve data as an independent variable and the electrophysiological curve data as a dependent variable, a gated recurrent unit network model is trained to construct and train the model, thereby obtaining an electrophysiological signal prediction model, wherein the electrophysiological signal prediction model includes two hidden layers, and the width of each hidden layer is 64;

[0009] Step S4: using the electrophysiological signal prediction model to predict changes in cardiac electrophysiological activity based on the mechanical contraction curve data, thereby obtaining electrophysiological prediction data; performing a prediction performance evaluation on the electrophysiological prediction data based on relative error, mean square error, and root mean square error, thereby obtaining model performance evaluation data;

[0010] Step S5: Extract characteristic indicators based on DFD80 and APD80 for the mechanical contraction curve data and the electrophysiological curve data, and calculate the Pearson correlation coefficient of the characteristic indicators to obtain correlation data between label-free optical detection and electrophysiological signals; optimize the parameters of the electrophysiological signal prediction model based on the model performance evaluation data and the correlation data to achieve the model's prediction accuracy for the electrophysiological signals of myocardial tissue.

[0011] By acquiring optical detection image data and electrophysiological signal data of myocardial tissue, the present invention can simultaneously monitor the mechanical contraction and electrophysiological activity of the myocardium. This provides rich biological information for subsequent modeling and analysis. The process, including image registration, cropping, periodic enhancement, time-difference enhancement, spatial filtering, intensity normalization, and data smoothing, can remove noise and interference from the data, improving its accuracy and consistency. This process ensures that the optical image and electrophysiological signal data are strictly aligned in space and time, making the mapping between the two more precise and laying a solid foundation for subsequent modeling. By extracting key feature signals (such as mechanical contraction curves and electrophysiological curves) from the optical-physiological signal mapping, complex physiological data can be converted into actionable feature vectors. This allows the originally high-dimensional signals to be represented using low-dimensional features, simplifying the complexity of data analysis. The mechanical contraction curve and electrophysiological curve can reflect the core characteristics of the heart in contraction and electrical activity, providing strong support for subsequent model training and ensuring that the model can understand and capture the physiological changes of the heart. A model based on a gated recurrent unit network (GRU) was constructed by using the mechanical contraction curve as the independent variable and the electrophysiological curve as the dependent variable. The GRU is suitable for processing time series data and can effectively capture and predict dynamic changes in the heart. Its two-hidden layer design helps capture more complex spatiotemporal relationships, and a hidden layer width of 64 ensures the model has sufficient capacity to learn complex patterns without overfitting. This model can predict corresponding electrophysiological signals based on input mechanical contraction curves, providing a new approach for non-invasive prediction of cardiac physiological activity, particularly in label-free optical monitoring scenarios. The trained electrophysiological signal prediction model can predict cardiac electrophysiological activity from mechanical contraction curves. This prediction method reduces reliance on traditional invasive measurement methods. The model's prediction accuracy can be evaluated from multiple perspectives using relative error, mean squared error, and root mean square error. This comprehensive performance evaluation ensures the model's sufficient generalization ability and stable operation under diverse data conditions, ensuring the reliability of the results. DFD80 and APD80 are important indicators of myocardial contraction and electrical activity. Extracting these features accurately captures the cardiac physiological state and potential pathologies, providing strong data support for subsequent diagnosis and analysis. Analyzing the relationship between optical detection signals and electrophysiological signals using the Pearson correlation coefficient quantifies the degree of correspondence between label-free optical detection and electrophysiological activity. This analysis can reveal the underlying mechanisms linking the two, providing a basis for subsequent optimization. Optimizing model parameters by combining model performance evaluation data with correlation data can further improve the model's predictive accuracy. Through continuous optimization, a model with even higher accuracy in predicting myocardial tissue electrophysiological signals is ultimately achieved, providing reliable support for practical clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0013] Figure 1 Schematic diagram of the process flow of the label-free optical detection method of myocardial tissue electrophysiology based on deep learning of the present invention;

[0014] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0015] Figure 3 for Figure 1 Detailed step flow chart of step S2 in FIG. DETAILED DESCRIPTION

[0016] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0017] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0018] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0019] To achieve this, please refer to Figures 1 to 3 The present invention provides a label-free optical detection method for myocardial tissue electrophysiology based on deep learning, the method comprising the following steps:

[0020] Step S1: Acquire optical detection image data and electrophysiological signal data of myocardial tissue; perform image preprocessing on the optical detection image data and electrophysiological signal data to obtain optical-physiological signal mapping data, wherein the image preprocessing includes image registration, image cropping, periodic enhancement, time difference enhancement, spatial filtering, intensity normalization, and data smoothing;

[0021] Step S2: extracting characteristic signals based on the mechanical contraction curve and the electrophysiological curve according to the optical-physiological signal mapping data, thereby obtaining mechanical contraction curve data and electrophysiological curve data;

[0022] Step S3: Using the mechanical contraction curve data as an independent variable and the electrophysiological curve data as a dependent variable, a gated recurrent unit network model is trained to construct and train the model, thereby obtaining an electrophysiological signal prediction model, wherein the electrophysiological signal prediction model includes two hidden layers, and the width of each hidden layer is 64;

[0023] Step S4: using the electrophysiological signal prediction model to predict changes in cardiac electrophysiological activity based on the mechanical contraction curve data, thereby obtaining electrophysiological prediction data; performing a prediction performance evaluation on the electrophysiological prediction data based on relative error, mean square error, and root mean square error, thereby obtaining model performance evaluation data;

[0024] Step S5: Extract characteristic indicators based on DFD80 and APD80 for the mechanical contraction curve data and the electrophysiological curve data, and calculate the Pearson correlation coefficient of the characteristic indicators to obtain correlation data between label-free optical detection and electrophysiological signals; optimize the parameters of the electrophysiological signal prediction model based on the model performance evaluation data and the correlation data to achieve the model's prediction accuracy for the electrophysiological signals of myocardial tissue.

[0025] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of a method for label-free optical detection of myocardial tissue electrophysiology based on deep learning of the present invention. In this example, the method for label-free optical detection of myocardial tissue electrophysiology based on deep learning includes the following steps:

[0026] Step S1: Acquire optical detection image data and electrophysiological signal data of myocardial tissue; perform image preprocessing on the optical detection image data and electrophysiological signal data to obtain optical-physiological signal mapping data, wherein the image preprocessing includes image registration, image cropping, periodic enhancement, time difference enhancement, spatial filtering, intensity normalization, and data smoothing;

[0027] In this embodiment of the present invention, under laboratory conditions, optical detection image data of myocardial tissue is collected using fluorescent calcium imaging combined with a mechanical wave imaging system. Electrophysiological signal data of myocardial tissue, including calcium ion concentration change signals, is simultaneously recorded using a calcium imaging system. To ensure data consistency, the optical image and electrophysiological signal are first registered. This registration process is based on adjustments to the field of view size, position, and angle. Next, image cropping techniques are used to crop the rectangular image into a circular image, with the shorter side of the rectangle as the diameter and the centers aligned. Next, the cropped circular image is periodically enhanced. The periodic characteristics of myocardial contraction are calculated using a Fourier transform, the shift step size is determined, and multiple sets of cyclic shift data are generated. These data are then subjected to time-difference enhancement, spatial filtering, and intensity normalization. Spatial filtering is performed using a Bartlett filter, and pixel intensities are normalized to the [0, 1] range. Finally, data smoothing is used to optimize the image to generate optical-physiological signal mapping data, ensuring the temporal and spatial consistency of the signals.

[0028] Step S2: extracting characteristic signals based on the mechanical contraction curve and the electrophysiological curve according to the optical-physiological signal mapping data, thereby obtaining mechanical contraction curve data and electrophysiological curve data;

[0029] The embodiment of the present invention extracts the mechanical contraction curve and electrophysiological curve of the myocardial tissue from the obtained optical-physiological signal mapping data. First, the mechanical contraction curve is constructed by analyzing the local displacement information in the optical image and the spatial deformation trend of the myocardium. Then, the electrical signal intensity fluctuation of each pixel point in the optical-physiological signal mapping data is used to generate an electrophysiological curve reflecting the electrical activity of the myocardium using time as the horizontal coordinate and the electrical signal intensity as the vertical coordinate. Next, the cardiac contraction cycle is divided into multiple cycle segments, the start and end time of each cycle is determined, and the peak and valley values ​​of each cycle are calibrated to generate cardiac cycle segmentation data. Finally, based on the features such as contraction amplitude and maximum contraction rate, a feature vector is obtained, the scale differences between different samples are eliminated, and the data are uniformly mapped to the standard numerical interval to obtain mechanical contraction curve data and electrophysiological curve data.

[0030] Step S3: Using the mechanical contraction curve data as an independent variable and the electrophysiological curve data as a dependent variable, a gated recurrent unit network model is trained to construct and train the model, thereby obtaining an electrophysiological signal prediction model, wherein the electrophysiological signal prediction model includes two hidden layers, and the width of each hidden layer is 64;

[0031] The embodiment of the present invention uses mechanical contraction curve data as an independent variable and electrophysiological curve data as a dependent variable to construct and train a gated recurrent unit (GRU) network model. First, the training set, validation set, and test set are divided based on a ratio of 80:10:10. Then, the GRU model is initialized. The model includes an input layer, two hidden layers (each layer contains 64 GRU units), and an output layer. The input layer receives the time step information of the mechanical contraction curve data, and the hidden layer extracts timing features and integrates complex patterns. The hyperparameters are set as follows: learning rate is 0.001, optimizer type is Adam, batch size is 32, and number of training rounds is 100. The training data is input into the GRU model, the predicted value is calculated by forward propagation, and the weight is adjusted by back propagation to reduce the error. During the training process, the model is continuously iteratively optimized, and finally an electrophysiological signal prediction model is obtained with high spatiotemporal prediction accuracy.

[0032] Step S4: using the electrophysiological signal prediction model to predict changes in cardiac electrophysiological activity based on the mechanical contraction curve data, thereby obtaining electrophysiological prediction data; performing a prediction performance evaluation on the electrophysiological prediction data based on relative error, mean square error, and root mean square error, thereby obtaining model performance evaluation data;

[0033] The embodiment of the present invention uses a trained electrophysiological signal prediction model to predict cardiac electrophysiological activity for new mechanical contraction curve data. The model compares the predicted electrophysiological signal data with the actual measured electrophysiological signal through forward propagation calculation, and calculates the relative error, mean square error (MSE) and root mean square error (RMSE). For example, in one application, the relative error between the actual data and the predicted data was 3%, the MSE was 0.005, and the RMSE was 0.07. Based on these error indicators, model performance evaluation data is generated, and multi-scale performance testing is performed. The prediction performance of the model is excellent in different time scales and signal amplitude ranges. Finally, a comparison chart of the predicted data and the actual data is presented visually, and the distribution of the prediction error is analyzed to ensure the accuracy and stability of the model.

[0034] Step S5: Extract characteristic indicators based on DFD80 and APD80 for the mechanical contraction curve data and the electrophysiological curve data, and calculate the Pearson correlation coefficient of the characteristic indicators to obtain correlation data between label-free optical detection and electrophysiological signals; optimize the parameters of the electrophysiological signal prediction model based on the model performance evaluation data and the correlation data to achieve the model's prediction accuracy for the electrophysiological signals of myocardial tissue.

[0035] The embodiment of the present invention performs DFD80 feature extraction on the mechanical contraction curve data and APD80 feature extraction on the electrophysiological curve data to obtain corresponding feature data. Then, the Pearson correlation coefficient is used to calculate the correlation between the mechanical contraction feature data and the electrophysiological feature data. For example, in a set of experimental data, the correlation coefficient is 0.85, indicating that the two have a strong linear relationship. The mean, standard deviation and confidence interval of the correlation data are further calculated to obtain correlation statistics. Based on these data, the quantitative correlation between label-free optical detection and electrophysiological signals is evaluated. Finally, the parameters of the electrophysiological signal prediction model are adjusted in combination with the model performance evaluation data and the correlation data to optimize the prediction accuracy of the model and ensure its efficient prediction of the electrophysiological activity of myocardial tissue.

[0036] By acquiring optical detection image data and electrophysiological signal data of myocardial tissue, the present invention can simultaneously monitor the mechanical contraction and electrophysiological activity of the myocardium. This provides rich biological information for subsequent modeling and analysis. The process, including image registration, cropping, periodic enhancement, time-difference enhancement, spatial filtering, intensity normalization, and data smoothing, can remove noise and interference from the data, improving its accuracy and consistency. This process ensures that the optical image and electrophysiological signal data are strictly aligned in space and time, making the mapping between the two more precise and laying a solid foundation for subsequent modeling. By extracting key feature signals (such as mechanical contraction curves and electrophysiological curves) from the optical-physiological signal mapping, complex physiological data can be converted into actionable feature vectors. This allows the originally high-dimensional signals to be represented using low-dimensional features, simplifying the complexity of data analysis. The mechanical contraction curve and electrophysiological curve can reflect the core characteristics of the heart in contraction and electrical activity, providing strong support for subsequent model training and ensuring that the model can understand and capture the physiological changes of the heart. A model based on a gated recurrent unit network (GRU) was constructed by using the mechanical contraction curve as the independent variable and the electrophysiological curve as the dependent variable. The GRU is suitable for processing time series data and can effectively capture and predict dynamic changes in the heart. Its two-hidden layer design helps capture more complex spatiotemporal relationships, and a hidden layer width of 64 ensures the model has sufficient capacity to learn complex patterns without overfitting. This model can predict corresponding electrophysiological signals based on input mechanical contraction curves, providing a new approach for non-invasive prediction of cardiac physiological activity, particularly in label-free optical monitoring scenarios. The trained electrophysiological signal prediction model can predict cardiac electrophysiological activity from mechanical contraction curves. This prediction method reduces reliance on traditional invasive measurement methods. The model's prediction accuracy can be evaluated from multiple perspectives using relative error, mean squared error, and root mean square error. This comprehensive performance evaluation ensures the model's sufficient generalization ability and stable operation under diverse data conditions, ensuring the reliability of the results. DFD80 and APD80 are important indicators of myocardial contraction and electrical activity. Extracting these features accurately captures the cardiac physiological state and potential pathologies, providing strong data support for subsequent diagnosis and analysis. Analyzing the relationship between optical detection signals and electrophysiological signals using the Pearson correlation coefficient quantifies the degree of correspondence between label-free optical detection and electrophysiological activity. This analysis can reveal the underlying mechanisms linking the two, providing a basis for subsequent optimization. Optimizing model parameters by combining model performance evaluation data with correlation data can further improve the model's predictive accuracy. Through continuous optimization, a model with even higher accuracy in predicting myocardial tissue electrophysiological signals is ultimately achieved, providing reliable support for practical clinical applications.

[0037] Preferably, step S1 includes the following steps:

[0038] Step S11: collecting optical detection image data of myocardial tissue under experimental conditions through an optical imaging system, wherein the optical imaging system is specifically a fluorescent calcium imaging combined with a mechanical wave imaging system;

[0039] Step S12: synchronously collecting electrophysiological signal data of myocardial tissue using a calcium imaging system, wherein the electrophysiological signal data includes a calcium ion concentration change signal;

[0040] Step S13: performing image registration on the optical detection image data and the electrophysiological signal data based on the field of view size, position and angle, thereby obtaining image consistency data;

[0041] Step S14: performing image cropping processing on the image consistency data to obtain circular consistent image data, wherein the image cropping processing specifically crops the original rectangular field of view into a circular area with the center of the rectangle as the circle center and the length of the short side of the rectangle as the diameter;

[0042] Step S15: performing periodic cyclic shifting and superposition on the circular consistent image data to obtain periodic enhanced data;

[0043] Step S16: performing signal enhancement and optimization processing based on time difference enhancement, spatial filtering, intensity normalization and data smoothing on the periodic enhancement data, thereby obtaining optical-physiological signal mapping data.

[0044] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S1 in the embodiment of the present invention, step S1 includes the following steps:

[0045] Step S11: collecting optical detection image data of myocardial tissue under experimental conditions through an optical imaging system, wherein the optical imaging system is specifically a fluorescent calcium imaging combined with a mechanical wave imaging system;

[0046] Under experimental conditions, the present invention embodiment uses a fluorescent calcium imaging system combined with a mechanical wave imaging system to perform optical detection of myocardial tissue. The experiment was set up in a constant temperature environment of 37°C to ensure the stability of the physiological state. The fluorescent calcium imaging system illuminates the myocardial tissue with excitation light of a specific wavelength, records the flow of calcium ions within the cells, and forms a fluorescent image. The mechanical wave imaging system captures the mechanical movement of the myocardium during contraction and relaxation. The imaging frequency is set to 100 frames per second, ensuring accurate capture of the rapid contraction and relaxation of the heart, ultimately obtaining optical detection image data with high temporal resolution.

[0047] Step S12: synchronously collecting electrophysiological signal data of myocardial tissue using a calcium imaging system, wherein the electrophysiological signal data includes a calcium ion concentration change signal;

[0048] While acquiring optical image data, the calcium imaging system in this embodiment of the present invention simultaneously collects electrophysiological signal data from myocardial tissue. Calcium ion concentration change signals are reflected by changes in fluorescence intensity after labeling with a calcium-sensitive dye. These signals represent changes in the action potential of myocardial cells. During the experiment, the acquisition frequency of calcium ion concentration changes was consistent with that of the optical imaging system, also at 100 frames per second, to ensure signal synchronization. These optical signals were converted into digital signals via a photoelectric converter to form electrophysiological signal data. This data includes the rising and falling phases of calcium ions, reflecting the contraction and relaxation processes of the myocardium.

[0049] Step S13: performing image registration on the optical detection image data and the electrophysiological signal data based on the field of view size, position and angle, thereby obtaining image consistency data;

[0050] The embodiment of the present invention performs image registration on the collected optical detection image data and electrophysiological signal data, with the aim of eliminating errors at different time points and spatial positions. The image registration process is adjusted based on the size, position and angle of the field of view. First, clear and obvious structural feature points in the optical image and electrophysiological signal data are selected as reference points, and then the images are adjusted using a registration algorithm based on nearest neighbor interpolation to ensure the consistency of each frame image in the time series. The field of view size is set to 10x10 mm by the initial imaging system, and the error after registration is less than 1 pixel, resulting in precisely aligned image consistency data.

[0051] Step S14: performing image cropping processing on the image consistency data to obtain circular consistent image data, wherein the image cropping processing specifically crops the original rectangular field of view into a circular area with the center of the rectangle as the circle center and the length of the short side of the rectangle as the diameter;

[0052] After image registration is completed, the embodiment of the present invention performs image cropping on the consistent image data. First, the center point of the rectangular field of view is determined and set as the center of the cropped circular image. Next, the length of the short side of the rectangular field of view is used as the diameter of the cropping circle, and the pixels on both sides of the short side are used as boundaries to crop a circular area concentric with it. This circular area retains the core part of the myocardial tissue while avoiding the interference of edge blur or distortion. Ultimately, the consistent circular image data is obtained, with a circular diameter of approximately 5 mm, ensuring that the data is concentrated in the important myocardial area.

[0053] Step S15: performing periodic cyclic shifting and superposition on the circular consistent image data to obtain periodic enhanced data;

[0054] In an embodiment of the present invention, the cropped circularly consistent image data is subjected to periodic cyclic shifting and superposition processing. First, the periodic characteristics of myocardial contraction are extracted by calculating a Fourier transform, resulting in the main frequency components of the cardiac cycle. Then, based on the period length of the periodic signal, the image data is cyclically shifted, with each shift taking the period length as the step size, generating multiple sets of shifted data. These shifted data are then superimposed to form a periodically enhanced data set. This process can enhance the periodic variation characteristics in the signal, making the rhythm of myocardial contraction and relaxation more distinct.

[0055] Step S16: performing signal enhancement and optimization processing based on time difference enhancement, spatial filtering, intensity normalization and data smoothing on the periodic enhancement data, thereby obtaining optical-physiological signal mapping data.

[0056] The embodiment of the present invention performs a series of signal enhancement and optimization processes on the obtained periodic enhancement data. First, time difference enhancement is performed to adjust the time delay of each pixel point and eliminate the timing deviation caused by the difference in signal propagation speed between each frame image. Then, a spatial filter based on the Bartlett window is used to spatially filter the image data to filter out the interference of high-frequency noise. Next, the pixel intensity of the image is normalized to unify all signal values ​​into the range of [0,1]. Finally, a Savitzky-Golay smoothing filter is applied to smooth the data to reduce the impact of noise on the signal, obtain optical-physiological signal mapping data, and ensure the continuity and consistency of the signal.

[0057] The fluorescent calcium imaging system of the present invention can efficiently capture changes in calcium ion concentration within myocardial cells, reflecting cardiac electrophysiological activity. Combined with mechanical wave imaging, it can simultaneously acquire mechanical motion information, providing more physical parameters for subsequent modeling. This combined imaging approach provides dual data support, enabling more accurate capture of the electrophysiological properties of myocardial tissue and its relationship to mechanical motion during model training. The synchronously acquired calcium signals can directly reflect the dynamic changes in calcium ions during cardiac excitation-contraction coupling. Since calcium ions are a key regulator of myocardial cell contraction, the synchronous acquisition of electrophysiological and calcium signals ensures accurate temporal matching between the electrical signals and the optical images during subsequent analysis, reducing timing errors in data processing. Image registration ensures precise alignment of images at different time points and angles. This eliminates displacement or angular deviations caused by experimental setup or tissue motion, ensuring image and signal consistency during subsequent analysis, which is particularly important for constructing accurate optical-physiological signal mapping. Cropping, which converts rectangular regions into circular regions, eliminates distortion or invalid information that may exist at the edges of the field of view, improving the overall quality of the image data and the accuracy of the analysis. This process simplifies the data structure, facilitating subsequent periodic analysis and signal enhancement. Through periodic enhancement, the periodic features in the image can be amplified, making repetitive movements such as cardiac contraction and relaxation in the signal more prominent. This enhancement process improves the model's sensitivity to periodic phenomena and helps to better extract the key contraction and electrical activity characteristics of the myocardium. By enhancing dynamic changes through time differences, the subtle timing features in the signal are amplified, making it easier to capture instantaneous changes in myocardial activity, remove image noise, retain important spatial features, and ensure a clearer signal; standardization processing makes the intensity values ​​of different pixels fall within a uniform scale range, which helps to eliminate intensity differences in the data and improve the consistency of the model; it reduces random fluctuations in the signal, improves the stability and reliability of the signal, and provides cleaner input data for subsequent model training.

[0058] Preferably, step S15 includes the following steps:

[0059] Step S151: performing Fourier transform on the circular consistent image data and determining the main cycle of myocardial tissue contraction, thereby obtaining periodic characteristic data;

[0060] The embodiment of the present invention performs Fourier transform processing on circular consistent image data, with the goal of extracting the main periodic features of myocardial tissue contraction. First, each frame of the image is converted into a frequency domain signal, and the amplitude spectrum of each frequency component in the image is calculated by fast Fourier transform (FFT). Then, based on the amplitude spectrum, the most significant frequency peak is determined, which represents the main period of myocardial tissue contraction. In order to enhance accuracy, image data of multiple time periods can be selected for Fourier transform, and the average value of the periodic features can be obtained to obtain stable periodic feature data. The main period at this time is usually between 1Hz and 2Hz, corresponding to a heart rate of 60-120 beats per minute.

[0061] Step S152: determining the cyclic shift step length according to the periodic characteristic data, thereby obtaining shift step length data;

[0062] The embodiment of the present invention determines the step size of the cyclic shift based on the periodic characteristic data. The cyclic shift step size is the frame number interval calculated based on the myocardial contraction cycle. In a specific implementation, the main cycle of the myocardium (in seconds) is combined with the imaging frame rate (for example, 100 frames / second) to calculate the corresponding number of frames in one cycle. For example, if the periodic characteristic data is 1.2 seconds, then a complete cycle contains 120 frames of images. These frame numbers are evenly divided into several steps, for example, one step size for every 20 frames, thereby obtaining shift step size data, which will be used for subsequent cyclic shift processing.

[0063] Step S153: cyclically shifting the circular consistent image data input signal according to the shift step data to generate multiple groups of shifted image data, thereby obtaining a cyclically shifted data set;

[0064] In an embodiment of the present invention, circularly consistent image data is cyclically shifted based on shift step data. Each frame of the image is shifted backward by the number of frames corresponding to the step, and this operation is repeated until the entire myocardial contraction cycle is covered. For example, assuming a cycle of 120 frames and a shift step of 20 frames, image data with different shift amounts, such as 20 frames, 40 frames, and 60 frames, are generated in sequence to form multiple shifted image sets. This process ensures that dynamic changes during myocardial contraction are captured, and the multiple sets of shifted data generated represent the myocardial state at different times, ultimately resulting in a cyclically shifted data set containing multiple time points.

[0065] Step S154: performing pixel-level weighted superposition on the images in the cyclically shifted data set to obtain periodically enhanced data, wherein the weights are distributed in inverse proportion to the shift amount.

[0066] The embodiment of the present invention performs pixel-level weighted superposition on the cyclically shifted data set to enhance the periodic signal of myocardial contraction. First, different weights are assigned to each group of shifted images, and the weight value is inversely proportional to the shift amount, that is, the smaller the shift step, the greater the image weight, and the larger the shift amount, the smaller the image weight. For example, the image weight with a step size of 20 frames can be set to 0.9, the image weight with a step size of 40 frames is set to 0.8, and so on. Then, the weighted average of each frame image at the same pixel position is performed, and the final value of each pixel point is calculated to obtain periodic enhanced data. Through this method, the noise and irregular changes in the original data are weakened, the periodic signal is enhanced, and finally a clear myocardial contraction rhythm image is obtained.

[0067] The Fourier transform of the present invention is a powerful tool for frequency domain analysis. It can decompose signals in time (or space) into distinct frequency components, highlighting periodic variations. Through the Fourier transform, the primary cycle of myocardial contraction can be accurately identified, thereby capturing the dominant characteristics of cardiac motion. This process simplifies complex signals into periodic components, facilitating subsequent shift processing and improving sensitivity to myocardial contraction characteristics. Based on the extracted periodic characteristics, the cyclic shift step size is set to ensure that each shift corresponds to a specific time point in the myocardial contraction cycle, thereby ensuring that the shifted image remains consistent with the periodic characteristics. This step size selection is crucial, ensuring that the subsequently generated shifted dataset contains sufficient information variation without disrupting the original periodic structure. This ensures that all shift operations are coordinated with the myocardial contraction process and reduces noise interference. By performing multiple cyclic shifts, the myocardial contraction process can be observed from different angles, forming a dataset containing multiple sets of shifted images. This data augmentation method not only enriches the training dataset but also better captures the complete dynamic characteristics of myocardial tissue throughout its entire cycle. Circular shifting is equivalent to multiple sampling of myocardial contractions in time, increasing robustness to temporal variations and enhancing the ability to capture periodic phenomena. Through pixel-level weighted superposition, key information from data with different shifts can be effectively integrated, further amplifying periodic features. The weight distribution in weighted superposition is based on the inverse of the shift amount, meaning that images with larger shifts have less impact on the final result, while images with smaller shifts have larger weights, thereby preserving the integrity of the core periodic features. This strategy helps eliminate unnecessary noise introduced by shifting, enhances periodic information, improves image signal quality, and provides clearer input data for subsequent signal extraction and modeling.

[0068] Preferably, step S16 includes the following steps:

[0069] Step S161: setting time difference parameters for the periodic enhancement data according to the actual cell waveform conduction velocity, thereby obtaining time difference parameter data;

[0070] The embodiment of the present invention sets the time difference parameter for the periodic enhancement data based on the actual cell waveform conduction velocity in the myocardial tissue. First, the conduction velocity of myocardial cells is generally between 0.3 and 1.0 m / s, depending on the specific experimental conditions. Based on the known conduction velocity and the physical size of the myocardial tissue, the time difference of the conduction signal between different pixels is calculated. For example, if the conduction velocity of the myocardium is 0.5 m / s and the spatial resolution of the imaging system is 100 microns / pixel, the signal conduction time difference between adjacent pixels is approximately 0.2 milliseconds. Based on these calculations, the time difference between different pixel points is set to obtain the time difference parameter data for subsequent processing.

[0071] Step S162: Calculating the absolute value of the difference between the current value of each pixel position and the value at a past time point determined according to the time difference parameter data based on the periodic enhancement data, thereby obtaining time difference enhancement data;

[0072] The embodiment of the present invention calculates the absolute value of the difference between the value of each pixel at the current time point and the value at the past time point based on the time difference parameter data. Specifically, for each pixel, the intensity value at the current moment is obtained, and the intensity value at the corresponding past moment is determined based on the time difference parameter. Then, the intensity difference between the two moments is calculated, and its absolute value is taken to obtain enhanced data reflecting the time difference. For example, if the intensity of a pixel at the current moment is 0.8, and the intensity at the past moment is 0.6, then its time difference enhanced data is |0.8-0.6|=0.2. The goal of this processing step is to highlight the changing trend of the signal and reduce the influence of static signals.

[0073] Step S163: performing spatial filtering processing based on the Bartlett filter on the time difference enhancement data to obtain spatial filtered data;

[0074] An embodiment of the present invention performs spatial filtering processing based on the Bartlett filter on the time difference enhancement data. The Bartlett filter is a weighted average filter with a linearly decreasing weight, gradually decreasing from the center to the outside. The specific operation is to apply the Bartlett filter to each pixel point and its adjacent area, and calculate the weighted average of each pixel in the area, with the center pixel having the largest weight and the weights of the surrounding pixels gradually decreasing. For example, the filter window size is 3x3 pixels, the weight of the center pixel is set to 1, and the weights of the surrounding 8 pixels are set to 0.5 or lower. After processing with the Bartlett filter, the noise in the time difference enhancement data will be effectively smoothed, and the local continuity of the signal will be enhanced, thereby obtaining spatially filtered data.

[0075] Step S164: performing the maximum and minimum values ​​of the time series of each pixel on the spatial filtering data to obtain pixel intensity range data;

[0076] An embodiment of the present invention performs time series analysis on spatial filtering data, extracts the maximum and minimum values ​​in the time series of each pixel point, and then obtains pixel intensity range data. The specific method is to traverse the intensity values ​​of each pixel point in all time frames, and record the maximum and minimum values ​​of the pixel in the entire cycle. For example, the intensity of a certain pixel point in the 5th frame is 0.9, which is the maximum value; the intensity in the 20th frame is 0.1, which is the minimum value. In this way, the fluctuation range of each pixel in the time series can be determined, thereby obtaining the intensity range data of each pixel as the basis for subsequent normalization processing.

[0077] Step S165: performing linear normalization on each pixel according to the pixel intensity range data, and mapping the value to the interval [0, 1], thereby obtaining intensity normalized data;

[0078] The embodiment of the present invention linearly normalizes the intensity value of each pixel based on the pixel intensity range data and maps the value to the interval [0, 1]. The formula for linear normalization is: (current value - minimum value) / (maximum value - minimum value). Through this formula, the intensity value of each pixel can be standardized to a uniform range. For example, if the minimum value of a pixel is 0.1 and the maximum value is 0.9, then its current value of 0.5 is normalized to (0.5-0.1) / (0.9-0.1) = 0.5. The normalized data will ensure that the intensity values ​​of all pixels are within a uniform range, which is convenient for subsequent analysis and processing, thereby obtaining intensity normalized data.

[0079] Step S166: performing data smoothing processing on the intensity normalized data to obtain optical-physiological signal mapping data.

[0080] The embodiment of the present invention performs data smoothing on the intensity normalized data to reduce intensity changes caused by noise or random fluctuations. The specific smoothing method used can be a moving average method or a Gaussian smoothing method. The moving average method performs an average calculation on each pixel point and several adjacent time points in its time series. For example, the intensity average of the three previous and next time points is taken to replace the current value. The Gaussian smoothing rule performs a weighted average on the time series according to the weight of the Gaussian function. Through smoothing, the noise can be effectively reduced, and the stability and continuity of the signal can be improved, thereby obtaining the final optical-physiological signal mapping data and ensuring the accuracy of subsequent data analysis.

[0081] The electrical signal conduction velocity of myocardial cells in the present invention directly affects the timing of signal propagation. Precisely setting the time difference parameter ensures that differences at different time points in the signal processing process are properly accounted for. By adjusting the time difference parameter based on the actual conduction velocity, the time difference enhanced data can reflect the actual changes in the electrophysiological activity of myocardial cells, improving the ability to capture dynamic signals, and thus facilitating more accurate analysis of the electrophysiological characteristics of myocardial tissue. By calculating the time difference between pixels, signal changes at different time points are enhanced. Time difference enhancement can amplify rapidly changing signal features, such as electrophysiological activity in myocardial tissue, thereby increasing sensitivity to short-term, drastic changes. This enhancement method not only highlights periodic fluctuations but also enhances the synergistic effect of spatiotemporal information, providing clearer time series data for subsequent filtering and processing. The Bartlett filter is a simple and efficient low-pass filter that can smooth noise and irregular signal fluctuations while preserving key signal features. Spatial filtering can reduce image noise generated during data acquisition, such as random fluctuations or high-frequency interference, thereby extracting smoother and more coherent physiological signal data. This process enhances the spatial consistency of the signal and is particularly suitable for removing high-frequency noise from surface signals of muscle tissue. By calculating the intensity range of each pixel, the amplitude of signal fluctuations throughout the entire time series can be effectively identified, capturing the signal variation characteristics of different regions within the image. This is crucial for further normalization and signal amplification, helping to identify pixels with large electrophysiological activity fluctuations and providing a standardized range for subsequent normalization. The goal of linear normalization is to standardize the intensity values ​​of each pixel, eliminating distortion caused by inconsistent signal intensities between pixels. Within the same image, different parts of myocardial tissue may have different signal intensities. Normalization can unify them to a comparable scale, enhancing the consistency of feature extraction. This process improves the accuracy of model training and prediction, ensuring that signals from different regions are analyzed at the same amplitude. Data smoothing can further eliminate irregular variations in the signal, ensuring signal continuity and stability. This is particularly important when processing biological signals, where excessively sharp fluctuations may be caused by noise. Smoothing can highlight the main trends of the electrophysiological signal without being affected by occasional noise or minor errors, thereby providing high-quality input data for the model and improving the robustness of the prediction model.

[0082] Preferably, step S166 includes the following steps:

[0083] Step S1661: Calculate the power spectrum density of the intensity normalized data and determine the main frequency components of the signal to obtain frequency characteristic analysis data;

[0084] The embodiment of the present invention performs power spectral density (PSD) calculation on the intensity normalized data to determine the main frequency components of the signal. The specific operation is to perform Fourier transform on the time series of each pixel point to obtain an energy distribution diagram in the frequency domain, and determine the main frequency components of the signal by analyzing the power spectral density curve. For example, if the acquisition frequency of the myocardial signal is 1000Hz, the PSD analysis finds that the main frequency distribution is between 0.5Hz and 5Hz, which represents the main frequency characteristics of myocardial contraction and electrophysiological signals. By analyzing the data with this frequency characteristic, the parameters of the smoothing process can be further optimized.

[0085] Step S1662: determining smoothing method parameters based on Rloess according to the frequency feature analysis data, thereby obtaining smoothing method parameter data, wherein the window size parameter in the smoothing method parameter data is set to 20% of the acquisition frame rate;

[0086] In an embodiment of the present invention, the parameters of the Rloess smoothing method are determined based on the frequency characteristic analysis data. Rloess is a local regression smoothing method, and the parameter settings include the window size and the polynomial order. In combination with the frequency characteristic analysis, the smoothing window size is usually set to 20% of the acquisition frame rate. For example, if the acquisition frame rate of the myocardial data is 500 frames / second, the window size is set to 100 frames to ensure that high-frequency noise can be effectively removed during the smoothing process while retaining the main signal characteristics. The polynomial order is selected according to the complexity of the signal and is usually set to 2 or 3.

[0087] Step S1663: smoothing each pixel time series of the intensity normalized data according to the smoothing method parameter data, and calculating the mean square error before and after smoothing, thereby obtaining smoothing effect evaluation data;

[0088] The embodiment of the present invention performs smoothing processing on each pixel time series of intensity normalized data based on smoothing method parameter data. The specific operation of the smoothing process is to apply the Rloess method to perform local regression fitting on the time series of each pixel to reduce the influence of noise. After processing, the mean square error (MSE) before and after smoothing is calculated to evaluate the smoothing effect. For example, if the MSE of a pixel point before smoothing is 0.05 and the MSE after smoothing is 0.01, it indicates that the smoothing effect is good. By calculating the mean square error, the effectiveness of the smoothing process is further judged, and smoothing effect evaluation data is obtained.

[0089] Step S1664: performing edge preservation processing on pixels with poor smoothing effect based on the smoothing effect evaluation data, thereby obtaining edge preservation data, wherein the edge preservation processing specifically involves detecting sudden changes and edges in the signal and performing local adjustments;

[0090] The embodiment of the present invention performs edge preservation processing on pixels with poor smoothing effects based on smoothing effect evaluation data. The specific method is to detect mutation points and edge locations in the signal and retain these signal features through local adjustment. First, edge points are identified by detecting mutations in the first-order derivative of the signal, and then the smoothing parameters are locally adjusted in these areas to reduce the weakening of the edge signal caused by the smoothing process. For example, at certain key electrophysiological signal mutation locations, the Rloess window size can be appropriately reduced to ensure that the edge features are retained, thereby obtaining edge-preserving data.

[0091] Step S1665: analyzing the signal changes between adjacent frames of the edge preservation data and performing time consistency correction to obtain time-corrected data;

[0092] The embodiment of the present invention performs a signal change analysis between adjacent frames on the edge retention data and performs time consistency correction. The specific operation is to calculate the signal change rate of each pixel between adjacent frames and determine whether there is an inconsistent signal jump. For pixels with poor time consistency, interpolation correction or local smoothing is performed to make the signal more continuous in the time dimension. For example, if the signal jump of a pixel between adjacent frames exceeds a set threshold (such as 20%), the frame signal is time-corrected to obtain time-corrected data.

[0093] Step S1666: Calculating the signal-to-noise ratio change before and after smoothing the time-corrected data to obtain signal-to-noise ratio improvement data;

[0094] The present invention uses time-corrected data to calculate the change in the signal-to-noise ratio (SNR) before and after smoothing. Specifically, the average signal value and standard deviation of the noise in each pixel time series are calculated, and the SNR before and after smoothing is calculated using the SNR formula (SNR = average signal value / standard deviation of noise). For example, if the SNR of a pixel is 10 before smoothing and 20 after smoothing, the SNR has doubled. By comparing the changes in the SNR, the effect of the smoothing process is evaluated, and SNR improvement data is obtained.

[0095] Step S1667: performing weighted fusion on the time-corrected data and the intensity-normalized data according to the signal-to-noise ratio improved data, thereby obtaining optical-physiological signal mapping data.

[0096] This embodiment of the present invention performs a weighted fusion of time-corrected data and intensity-normalized data based on the signal-to-noise ratio (SNR) improvement data. Specifically, weights are assigned to the time-corrected and intensity-normalized data based on changes in the SNR. For example, if the SNR improvement for a particular pixel is significant, a higher weight is assigned to the time-corrected data; conversely, a higher weight is assigned to the intensity-normalized data. Ultimately, the weighted fusion preserves both intensity accuracy and temporal consistency, resulting in the final optical-physiological signal mapping data, ensuring the integrity and reliability of the signal data.

[0097] The power spectral density analysis of the present invention helps extract the primary frequency components in a signal and evaluate the signal's primary fluctuation characteristics from a frequency domain perspective. This is particularly true in myocardial electrophysiological signals, where different frequencies correspond to distinct physiological processes. By analyzing frequency characteristics, key frequencies in the signal can be identified, providing a basis for subsequent smoothing parameter settings and ensuring that these key physiological features are not diminished during smoothing. Rloess is a smoothing method based on local regression that eliminates noise by fitting local data while preserving key changes in the signal. Setting the window size to 20% of the acquisition frame rate can significantly smooth short-term fluctuations without affecting the signal's primary trends or edge information. This method makes the smoothing process more adaptive and is particularly suitable for processing myocardial electrophysiological data containing random noise. The effectiveness of the smoothing process can be quantitatively evaluated by calculating the mean squared error (MSE). A lower MSE indicates that the smoothed signal differs less from the original signal, preserving more signal features. Furthermore, this evaluation data provides a basis for subsequent detection of pixels with poor smoothing performance, facilitating further adjustments to signals that have not been adequately optimized. In some pixels, smoothing may result in over-smoothing the signal, weakening breakpoints and edge information. Therefore, edge-preserving processing aims to preserve signal transition points and key edge features, preventing the smearing of important physiological signal features. This processing not only smooths noise but also preserves transitions and boundary signals, ensuring signal integrity and accuracy. Temporal consistency correction ensures temporal continuity between frames, reducing errors caused by temporal drift during data acquisition or processing. By analyzing signal variations between frames and correcting for potential inconsistencies, the temporal stability of myocardial electrophysiological signals is ensured. This processing is crucial for subsequent improvement in signal-to-noise ratio (SNR) and model prediction accuracy. Improvement in SNR is a key metric for measuring the effectiveness of signal processing. By calculating this improvement in SNR, the effectiveness of smoothing on denoising can be quantified. A significant improvement in SNR indicates that the smoothing and correction steps have successfully reduced noise and improved signal quality, providing better input data for subsequent analysis and model training. Weighted fusion effectively combines the time-corrected and intensity-normalized signals to produce higher-quality optical-to-physiological signal mapping data. Weighted fusion not only preserves signal strength information but also further enhances temporal consistency and optimizes the signal-to-noise ratio, generating more accurate and biologically meaningful mapping data. This step ensures that the final data accurately reflects the electrophysiological activity of myocardial tissue, providing strong support for subsequent model training and prediction.

[0098] Preferably, step S2 includes the following steps:

[0099] Step S21: extracting local displacement information and spatial variation trend of myocardial contraction deformation in the image from the optical-physiological signal mapping data, thereby constructing a mechanical contraction curve;

[0100] Step S22: mapping the electrical signal intensity fluctuations at each pixel in the optical-physiological signal data, and using time as the horizontal coordinate and the electrical signal intensity as the vertical coordinate to obtain an electrophysiological curve reflecting the electrical activity characteristics of the myocardial cells;

[0101] Step S23: dividing the mechanical contraction curve and the electrophysiological curve into cardiac contraction cycles, determining the start and end time points of each cardiac cycle, and calibrating the peak and valley values, thereby obtaining cardiac cycle segmentation data;

[0102] Step S24: performing feature extraction based on contraction amplitude, maximum contraction rate, peak and valley values ​​of the electrical signal, and rise time on the mechanical contraction curve and the electrophysiological curve in the cardiac cycle according to the cardiac cycle segmentation data, thereby obtaining feature vector data;

[0103] Step S25: performing scale difference elimination processing on the feature vector data between different samples, and mapping the data values ​​to a unified numerical interval, thereby obtaining mechanical contraction curve data and electrophysiological curve data.

[0104] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 1 Detailed step flow diagram of step S2 in the embodiment of the present invention, step S2 includes the following steps:

[0105] Step S21: extracting local displacement information and spatial variation trend of myocardial contraction deformation in the image from the optical-physiological signal mapping data, thereby constructing a mechanical contraction curve;

[0106] The embodiment of the present invention extracts the local displacement information from the optical-physiological signal mapping data. The specific operation is to detect the displacement of the myocardial tissue in each frame of the image through the optical flow algorithm to capture the motion trajectory of the myocardial tissue. Then, the displacement information is spatially interpolated to construct the spatial variation trend of the myocardial contraction deformation. By analyzing these trend changes, the displacement curves of the myocardium in different time periods can be drawn, and finally a complete mechanical contraction curve can be constructed. For example, in an experimental scenario, by detecting the local displacement of the myocardium, it was found that the maximum displacement of the myocardium in each cardiac cycle was about 1.5 mm, and the displacement curve reflects the dynamic contraction deformation process of the myocardium.

[0107] Step S22: mapping the electrical signal intensity fluctuations at each pixel in the optical-physiological signal data, and using time as the horizontal coordinate and the electrical signal intensity as the vertical coordinate to obtain an electrophysiological curve reflecting the electrical activity characteristics of the myocardial cells;

[0108] The embodiment of the present invention analyzes the fluctuation of the electrical signal intensity of each pixel in the optical-physiological signal mapping data. Specifically, the time series data of each pixel is selected, and the electrophysiological curve of each pixel is drawn with time as the horizontal coordinate and the electrical signal intensity as the vertical coordinate. These curves reflect the electrical activity characteristics of myocardial cells, such as the process of depolarization and repolarization. For example, in actual operation, the electrophysiological curve of a certain myocardial area shows that the electrical signal intensity rises rapidly to a peak value of 0.8V during depolarization, and slowly decreases to a valley value of 0.2V during repolarization, which can clearly reflect the changes in the electrical activity of myocardial cells.

[0109] Step S23: dividing the mechanical contraction curve and the electrophysiological curve into cardiac contraction cycles, determining the start and end time points of each cardiac cycle, and calibrating the peak and valley values, thereby obtaining cardiac cycle segmentation data;

[0110] The embodiment of the present invention divides the cardiac contraction cycle by the mechanical contraction curve and the electrophysiological curve. The specific method is to determine the start and end time points of each cardiac cycle by calculating the local extreme points in the mechanical contraction curve and the electrophysiological curve. First, the main peaks and valleys reflecting the cardiac cycle in the electrophysiological curve are determined, these extreme points are calibrated, and then the peak and valley positions of the mechanical contraction curve are synchronized. For example, in a certain experimental data, the peak voltage of the cardiac cycle is 1.0V, the valley value is 0.1V, and the cycle length is about 800 milliseconds. Through this extreme value calibration, the cardiac cycle can be clearly divided.

[0111] Step S24: performing feature extraction based on contraction amplitude, maximum contraction rate, peak and valley values ​​of the electrical signal, and rise time on the mechanical contraction curve and the electrophysiological curve in the cardiac cycle according to the cardiac cycle segmentation data, thereby obtaining feature vector data;

[0112] The embodiment of the present invention performs feature extraction on the mechanical contraction curve and electrophysiological curve in each cycle based on cardiac cycle segmentation data. The specific steps of feature extraction include: first, calculating the mechanical contraction amplitude (the difference between the peak and valley values) and the maximum contraction rate (the maximum slope of the contraction curve) of each cycle; then, calculating the peak, valley, and rise time (the time difference from valley to peak) of the electrophysiological curve. For example, the mechanical contraction amplitude in a certain cardiac cycle is 1.2 mm, the maximum contraction rate is 2.5 mm / s, the electrophysiological peak is 0.9 V, the valley is 0.2 V, and the rise time is 150 milliseconds. These extracted feature values ​​constitute the characteristic vector data of the cycle.

[0113] Step S25: performing scale difference elimination processing on the feature vector data between different samples, and mapping the data values ​​to a unified numerical interval, thereby obtaining mechanical contraction curve data and electrophysiological curve data.

[0114] The embodiment of the present invention performs a process to eliminate scale differences between different samples of the feature vector data. Specifically, the data is mapped to a uniform numerical range (such as [0, 1]) through a normalization method. For example, the numerical range of the mechanical contraction amplitude may vary greatly among different samples. Through normalization, the contraction amplitude of all samples is mapped to the [0, 1] interval. At the same time, the electrophysiological curve data is subjected to the same normalization process, and the peak and valley values ​​of the electrical signal are respectively mapped to the interval [0, 1]. After the processing is completed, the mechanical contraction curve data and the electrophysiological curve data of different samples are compared and analyzed under a unified scale to ensure the comparability between the data.

[0115] The local displacement information and spatial variation of myocardial contractile deformation in this invention are core parameters reflecting myocardial mechanical activity. Extracting these features can intuitively demonstrate the degree of deformation at different locations during cardiac contraction, helping to understand myocardial contractile properties. Constructing a mechanical contraction curve can quantitatively describe the amplitude and rate of myocardial motion, providing essential basic data for subsequent electrophysiological signal analysis and biomechanical modeling. The electrophysiological curve reflects the electrical activity characteristics of myocardial cells and can demonstrate how electrical signals change at different time points. This curve accurately reflects the excitation and recovery processes of myocardial cells during electrophysiological activity, making it particularly crucial for studying the spatiotemporal distribution of cardiac electrical signals. This curve provides direct input data for training ECG signal prediction models. Cardiac cycle segmentation helps isolate specific contraction and electrical activity events within each cardiac cycle, enabling precise analysis of the characteristics within each cycle. By calibrating peak and valley values, the moments of maximum contraction and minimum relaxation of the heart during each cycle can be captured, providing a periodic structure for subsequent feature extraction and model training. Cycle segmentation also improves data processing accuracy and helps identify abnormal periodic behavior. Feature vector extraction can extract core information from complex time series data, helping to identify and distinguish important physiological features in different cardiac cycles. By extracting features from mechanical contraction curves and electrophysiological curves, the mechanical and electrical activity states of the heart can be quantitatively described, which helps further data analysis, modeling, and prediction; this process reduces the complexity of the data while retaining key information. Due to the different scales and contraction characteristics of different heart samples, direct comparison of raw data will lead to errors. By eliminating scale differences and normalizing data, systematic differences between different samples can be eliminated, making data from different samples comparable. This step enables the model to process data of various sizes during training, while enhancing the consistency and interpretability of feature data, facilitating subsequent model training and analysis.

[0116] Preferably, step S3 includes the following steps:

[0117] Step S31: using the mechanical contraction curve data as an independent variable and the electrophysiological curve data as a dependent variable, and matching them based on the time series to form a training data set;

[0118] In this embodiment of the present invention, mechanical contraction curve data is used as the independent variable and electrophysiological curve data as the dependent variable. These two are matched point by point using timestamps to form a complete time series dataset. Specifically, the matching process involves finding the corresponding time point in the electrophysiological curve for each time point in the mechanical contraction curve, ensuring that the two are aligned in the temporal dimension. For example, if the mechanical contraction curve value at a certain time point is 0.75 mm, and the electrophysiological curve voltage value is 0.65 V, then this matching point will be used as a sample in the training data. The resulting training dataset contains thousands of similar samples, forming the time series data.

[0119] Step S32: Divide the training data set into a training set, a validation set, and a test set based on a ratio of 80:10:10;

[0120] This embodiment of the present invention divides the training dataset into training, validation, and test sets in an 80:10:10 ratio. Specifically, the entire dataset is first randomly shuffled to avoid model bias caused by serial correlation. Then, 80% of the data is used to train the model, 10% is used to verify the model's performance, and the remaining 10% is used for final testing. For example, if the dataset contains 10,000 samples, 8,000 samples will be used for model training, 1,000 for validation, and 1,000 for testing, ensuring the independence of the different datasets and the consistency of the data distribution.

[0121] Step S33: Initializing a gated recurrent unit network model based on the mechanical contraction curve data and the electrophysiological curve data, and setting hyperparameters of the gated recurrent unit network model based on the learning rate, optimizer type, batch size, and number of training rounds, thereby obtaining an initial GRU network model;

[0122] In an embodiment of the present invention, a gated recurrent unit (GRU) network model is initialized based on mechanical contraction curve and electrophysiological curve data. Model initialization includes determining the network structure and setting hyperparameters. Specific hyperparameters include: a learning rate of 0.001, an Adam optimizer, a batch size of 32, and 100 training rounds. The GRU network contains two hidden layers, each with a width of 64 neurons, to capture dynamic features in time series. This network structure can effectively process time-dependent long- and short-term information, regulating the flow of information through a gating mechanism. The initialized GRU network model is ready for training.

[0123] Step S34: input the training set into the initial GRU network model for training, calculate the predicted value through forward propagation, and then adjust the weight through backpropagation to reduce the prediction error, thereby obtaining a trained dynamic time series prediction model;

[0124] In this embodiment of the present invention, the training set is input into the initial GRU network model for training. The training process includes: first, forward propagation is used to calculate the predicted value of each sample, that is, the model predicts the electrophysiological curve value based on the input mechanical contraction curve data; then, the error between the predicted value and the actual electrophysiological curve value is calculated, using the mean square error (MSE) as the loss function; finally, the model adjusts the weights through backpropagation and optimizes the model parameters to reduce the error. With each round of training, the model's prediction accuracy gradually improves. After 100 rounds of training, a trained dynamic time series prediction model is obtained.

[0125] Step S35: Validate the dynamic time series prediction model using the validation set, and adjust the training strategy based on the error of the validation result, thereby obtaining an adaptive spatiotemporal attention GRU model;

[0126] The embodiment of the present invention uses a validation set to verify the trained dynamic time series prediction model. The specific operation is: input the validation set into the model and calculate the error between the predicted value and the true value. According to the performance of the validation error, the parameters such as the learning rate and the number of training rounds of the model are adjusted in a timely manner. For example, when it is found that the error on the validation set no longer decreases significantly, an early stopping strategy is adopted to prevent overfitting. At the same time, by introducing an adaptive spatiotemporal attention mechanism, the model can dynamically adjust the weights according to the different importance of time and space, thereby further improving the prediction accuracy of the model and forming an adaptive spatiotemporal attention GRU model.

[0127] Step S36: Performing a performance test based on mean square error and root mean square error on the adaptive spatiotemporal attention GRU model using the test set to obtain model performance evaluation data;

[0128] The embodiment of the present invention uses a test set to perform a performance test on the adaptive spatiotemporal attention GRU model. The specific operation is as follows: the test set is input into the model, and the model is used to predict the electrophysiological curves in the test set. Then, the mean square error (MSE) and root mean square error (RMSE) between the predicted value and the true value are calculated. For example, in one test scenario, the model's prediction results for electrophysiological activity during the cardiac cycle showed an MSE of 0.004 and an RMSE of 0.063, which well reflects the model's generalization ability on unknown data, and ultimately obtains the model's performance evaluation data.

[0129] Step S37: Perform integrated learning optimization processing on the adaptive spatiotemporal attention GRU model based on the model performance evaluation data to obtain an electrophysiological signal prediction model.

[0130] In this embodiment of the present invention, an adaptive spatiotemporal attention GRU model is optimized through ensemble learning based on model performance evaluation data. Specifically, multiple GRU models are introduced and optimized using ensemble learning. For example, algorithms such as random forests or gradient boosted decision trees are used to weightedly fuse the prediction results of multiple models, thereby improving the overall model's predictive performance. After ensemble learning, the resulting electrophysiological signal prediction model is able to maintain high prediction accuracy and robustness even with more complex myocardial electrophysiological data.

[0131] The time series matching of the mechanical contraction curve and the electrophysiological curve in the present invention ensures consistency in input and output data, helping to capture the temporal relationship between cardiac contraction and electrophysiological activity. This provides the model with a complete input-output pair, enabling the network to learn the mapping relationship between the two and improving the accuracy of model predictions. Reasonable data set division can effectively prevent model overfitting: the training set is used for initial model training, the validation set is used to adjust the model's hyperparameters during training, and the test set is used to ultimately evaluate model performance. This three-part data set division allows for comprehensive testing of the model's generalization capabilities and ensures its predictive effectiveness on new data. The GRU model is suitable for processing time series data and can effectively capture the long- and short-term dependencies between myocardial contraction and electrophysiological activity. Hyperparameter settings can optimize the model training process. An appropriate learning rate and optimizer can accelerate model convergence, while the batch size and number of training rounds ensure that the model fully learns from the data, improving training effectiveness. Forward propagation allows the model to learn to predict electrophysiological signals from the input data, while backpropagation adjusts the model's weights by minimizing the error. As training progresses, the model gradually converges to a low-error state, resulting in continuously improved prediction accuracy and gradually optimizing the model's predictive capabilities. Performance testing on the validation set allows the model to adjust its training strategy promptly to avoid overfitting or underfitting. The introduction of an adaptive spatiotemporal attention mechanism allows the model to more effectively focus on features at key time points and spatial regions, thereby more accurately capturing the dynamic relationship between mechanical contraction and electrophysiological activity and enhancing the model's predictive capabilities. The test set, independent of the training process, provides a true assessment of the model's generalization performance. Mean squared error and root mean squared error are important metrics for measuring model prediction accuracy. These metrics can quantitatively evaluate the model's performance on real-world data, providing a basis for further optimization and application. Ensemble learning, by combining the strengths of multiple models, can further enhance prediction accuracy and stability. Fusion of the results from multiple models with different weights helps eliminate the shortcomings of individual models and improve overall prediction performance. The resulting electrophysiological signal prediction model exhibits enhanced generalization and robustness, enabling accurate prediction of dynamic changes in myocardial electrophysiological activity.

[0132] Preferably, step S33 includes the following steps:

[0133] The gated recurrent unit network model is initialized according to the mechanical contraction curve data and the electrophysiological curve data, wherein the gated recurrent unit network model includes an input layer, two hidden layers and an output layer. The input dimension of the input layer is the time step of the mechanical contraction curve; the number of GRU units in the first hidden layer is 64, and its task is to extract primary temporal features from the input mechanical contraction curve data; the number of GRU units in the second hidden layer is 64, which receives the output from the first hidden layer and performs high-level feature integration and complex pattern recognition; the output dimension of the output layer is the time step of the electrophysiological curve; the hyperparameters of the gated recurrent unit network model are set based on the learning rate, optimizer type, batch size and number of training rounds to obtain the initial GRU network model.

[0134] This embodiment of the present invention initializes a gated recurrent unit (GRU) network model based on mechanical contraction curve data and electrophysiological curve data. First, the input dimension of the input layer is equal to the time step of the mechanical contraction curve. Specifically, each time step corresponds to a feature vector, meaning each time step represents a mechanical contraction state. For this embodiment, assuming the time step of the mechanical contraction curve data is 100 points, the input dimension of the input layer is set to 100. Next, the first hidden layer contains 64 GRU units, whose task is to extract primary temporal features from the input mechanical contraction curve data. GRU units can capture long-term dependencies in time series information and, through their gating mechanism, effectively filter out unimportant information. The second hidden layer, also containing 64 GRU units, is responsible for receiving the output from the first hidden layer and integrating high-level features and recognizing complex patterns, helping the network establish a deep relationship between mechanical contraction and electrophysiological signals. The output dimension of the output layer is equal to the time step of the electrophysiological curve. That is, each mechanical contraction time step corresponds to an electrophysiological signal output. Assuming the time step of the electrophysiological curve is also 100, the dimension of the output layer is 100. To ensure effective model learning, key hyperparameters are set: the learning rate is set to 0.001, the Adam optimizer is selected to accelerate model convergence, the batch size is set to 32, meaning 32 sets of data are input for training at each iteration, and the number of training rounds is set to 100. In practice, the model initialization process also includes random initialization of the weights. The weights are optimized through each iterative round to ultimately obtain the initial GRU network model. In practical applications, such as sequence prediction of myocardial data, this architecture can capture the complex relationship between mechanical contraction and electrophysiological activity.

[0135] The present invention uses the time step of the mechanical contraction curve as an input dimension, ensuring that the model can capture the timing information of myocardial contraction and reflect the dynamic changes of the mechanical contraction process. Each contraction and relaxation process of the myocardium is closely related to electrophysiological activity. Using a complete time series as input can provide the model with sufficient information foundation to ensure that subsequent layers can accurately extract timing features. The main function of the first hidden layer is to extract primary features from the timing data, such as the basic patterns in the cardiac contraction and relaxation cycles. The 64 GRU units have sufficient expressive power to perform preliminary analysis of the mechanical contraction signals in different time periods and extract the main characteristic patterns in the myocardial contraction process, such as changes in contraction amplitude and frequency information. This layer helps to capture the basic relationship between mechanical movement and electrophysiological activity, laying the foundation for subsequent deep feature integration. By adding a second hidden layer, the model has higher expressive power and the ability to learn complex patterns. Although the primary features extracted by the first hidden layer can characterize the basic changes in mechanical contraction, the dynamic behavior of myocardial activity is very complex, requiring advanced feature integration through deeper network layers. The second hidden layer can identify higher-order correlations between myocardial contraction and electrophysiological signals, such as subtler waveform changes and their interactions with electrical activity, improving the model's generalization and prediction accuracy. Setting the output dimension to the time step of the electrophysiological curve ensures that the model generates output signals that correspond to actual myocardial electrical activity. By analyzing the mechanical contraction signal, the model can predict the electrophysiological response of the heart at each time step, thereby providing dynamic predictions of electrical signal changes. This design ensures that the model's output is consistent with the heart's electrical activity and can reflect complex physiological phenomena. A learning rate that is too large can cause the model to skip the optimal solution during training, while a too small learning rate can result in excessive training time. With an appropriate learning rate setting, the model can stably converge to the global optimal solution. The choice of optimizer type determines how the model weights are updated. Optimizers such as Adam, which adaptively adjust the learning rate, offer faster convergence and better performance. The batch size affects the model's generalization and training efficiency. A moderate batch size (such as 32 or 64) ensures model stability during training without incurring excessive computational overhead. The number of training epochs determines how well the model is trained on the dataset. A higher number of epochs ensures that the model fully learns the data features, but also avoids overfitting.

[0136] Preferably, step S4 includes the following steps:

[0137] Step S41: inputting the new mechanical contraction curve data into the electrophysiological signal prediction model, and predicting changes in cardiac electrophysiological activity through forward propagation calculation, thereby obtaining electrophysiological prediction data;

[0138] In this embodiment of the present invention, new mechanical contraction curve data is input into a pre-trained electrophysiological signal prediction model. Using a forward propagation method, the model calculates predicted changes in cardiac electrophysiological activity time-step by time-step based on the input mechanical contraction data, generating predicted electrophysiological data. During this process, the GRU units within the model infer the corresponding electrophysiological activity signals based on previously learned patterns. For example, in a practical application scenario, if the time-step length of the new mechanical contraction data is 100, the model will predict the electrical signal for each time-step and output the corresponding electrophysiological signal.

[0139] Step S42: comparing the electrophysiological prediction data with the actually measured electrophysiological data, and calculating the relative error, mean square error, and root mean square error to obtain prediction error data;

[0140] The embodiment of the present invention compares the predicted electrophysiological signal with the electrophysiological data actually measured in the experiment, focusing on the difference between the two. Commonly used error indicators are calculated, including relative error, mean square error (MSE), and root mean square error (RMSE). These error indicators can accurately reflect the difference between the model prediction value and the actual value. For example, if the actual value of a time step is 0.8 and the model prediction value is 0.75, the relative error and MSE / RMSE can be calculated according to the formula.

[0141] Step S43: performing prediction performance evaluation on the electrophysiological signal prediction model at different time scales and signal amplitude ranges based on the prediction error data, thereby obtaining multi-scale performance evaluation data;

[0142] The present invention uses prediction error data to evaluate the prediction performance of electrophysiological signal prediction models at different time scales (e.g., short-term versus long-term trends) and signal amplitude ranges (e.g., small versus large signal fluctuations). This multi-scale analysis allows us to understand the model's prediction accuracy at different levels of the signal. For example, whether the model performs well in the presence of rapidly changing ECG signals. The evaluation results form multi-scale performance evaluation data.

[0143] Step S44: Visualizing the prediction results based on the multi-scale performance evaluation data to obtain visualized performance evaluation data, wherein the visualization includes generating a comparison chart of the predicted value and the actual value and an error distribution chart;

[0144] The embodiments of the present invention visualize multi-scale performance evaluation data, generating a variety of charts to help researchers more intuitively understand the model's performance. Common visualization methods include plotting a comparison curve between predicted and actual values, and error distribution diagrams that display the magnitude of the error at different time steps. For example, in an application scenario, a line chart can be used to visually see whether the model's predicted trend is consistent with the actual electrophysiological signal, as well as the error fluctuations at different time points.

[0145] Step S45: Generate a detailed model performance report for the multi-scale performance evaluation data and the visualization performance evaluation data, including the overall prediction accuracy, statistical analysis of various error indicators, and the performance of the model under different conditions, thereby obtaining model performance evaluation data.

[0146] The embodiment of the present invention generates a detailed model performance report based on the aforementioned multi-scale performance evaluation data and visualization performance evaluation data. The report covers the overall prediction accuracy, statistical analysis of various error indicators (such as average error, maximum error, etc.), and evaluates the performance of the model under different experimental conditions. For example, the report will show whether the model performs better with high-frequency electrophysiological fluctuations than with low-frequency fluctuations, whether it is more accurate in predicting small-amplitude signals, etc. This report will be used to evaluate the pros and cons of the model and help further optimize and improve the model.

[0147] The present invention utilizes a previously trained model for forward propagation to quickly and accurately predict cardiac electrophysiological activity. This prediction is based on the dynamic relationship between mechanical contraction and electrophysiological activity. The model leverages temporal characteristics to efficiently simulate the complex changes in cardiac activity. The advantage of this step is that new electrophysiological data can be predicted without actual measurements, significantly improving the efficiency of diagnosis and assessment. Comparing predicted data with actual data allows for intuitive assessment of model accuracy. Relative error reflects the difference between the predicted and actual values, while mean square error (MSE) and root mean square error (RMSE) measure the degree of deviation from the predicted data. This precise error calculation can identify model deficiencies and provide a basis for subsequent improvements. Root mean square error (RMSE) can further amplify the impact of larger errors, helping to identify time periods or scenarios where errors are concentrated. Multi-scale performance evaluation provides a more detailed assessment of model performance. Because the temporal characteristics of cardiac electrophysiological activity can exhibit different patterns across different timeframes, evaluating the model's prediction performance over short and long timeframes can help determine whether the model is adaptable to varying signal variations. Furthermore, analyzing the predictive capabilities for different signal amplitudes (such as large abnormal cardiac fluctuations and small normal activity) can improve the model's universality across different cardiac states and enhance its clinical application value. Visualizing the predicted results against actual data provides a more intuitive demonstration of the model's predictive performance. Comparison charts provide a visual indication of the degree of agreement between predicted and actual values, while error distribution plots clearly reveal the specific distribution of errors. This visualization tool not only facilitates rapid identification of model strengths and weaknesses but also provides a reference for model improvement, playing a particularly important role in analyzing model performance at different time scales. Generating a detailed model performance report systematically summarizes the model's performance in predicting cardiac electrophysiological signals. The report includes not only an overall accuracy assessment but also detailed performance across different time scales and signal amplitude ranges. Statistical analysis of multiple error metrics (such as relative error, mean squared error, and root mean square error) provides a comprehensive understanding of the differences in model performance under different experimental conditions. This report provides a strong basis for subsequent model optimization and improvement and provides detailed data support for validating the model's feasibility in clinical applications.

[0148] Preferably, step S5 includes the following steps:

[0149] Step S51: performing DFD80 extraction on the mechanical contraction curve data to obtain mechanical contraction characteristic data; performing APD80 extraction on the electrophysiological curve data to obtain electrophysiological characteristic data;

[0150] The embodiment of the present invention extracts DFD80 (delay to peak 80%) from the mechanical contraction curve data. Specifically, DFD80 refers to the time required for the myocardial contraction curve to reach 80% of the maximum contraction amplitude from the starting position. By detecting the time value at that moment in each mechanical contraction curve, the characteristic data of the mechanical contraction is extracted. Similarly, APD80 (action potential duration 80%) is extracted from the electrophysiological curve data, which represents the time it takes for the electrophysiological signal to drop from the peak to 80% of the amplitude. In this way, the dynamic characteristics between myocardial electrical activity and mechanical contraction can be reflected. The process of extracting these features depends on the curve shape within the cardiac cycle, and the specific time points are usually defined in the electrocardiogram or optical detection data.

[0151] Step S52: Calculating the Pearson correlation coefficient of the mechanical contraction characteristic data and the electrophysiological characteristic data to obtain correlation data;

[0152] The present invention calculates the Pearson correlation coefficient between mechanical contraction feature data and electrophysiological feature data to quantitatively assess their correlation. The Pearson correlation coefficient is calculated by dividing the covariance of two variables by the product of their standard deviations, yielding a value between -1 and 1, where 1 indicates a perfect positive correlation and -1 indicates a perfect negative correlation. In practice, each pair of data points (e.g., DFD80 and APD80) is first matched. Then, the correlation calculation formula is used to analyze the mechanical contraction and electrophysiological features at all time points to generate correlation data.

[0153] Step S53: Calculate the average correlation coefficient, standard deviation, and confidence interval of the correlation data to obtain correlation statistics;

[0154] Based on the correlation data, the present invention further calculates the average correlation coefficient, standard deviation, and confidence interval. The average correlation coefficient reflects the general level of correlation across the sample, the standard deviation indicates the degree of dispersion in the correlation, and the confidence interval can help assess the reliability of the correlation. For example, in medical applications, a high average correlation coefficient and a low standard deviation indicate a high correlation between mechanical contraction and electrophysiological signals, while a narrow confidence interval further confirms the reliability of this result.

[0155] Step S54: quantitatively evaluating the degree of correlation between the label-free optical detection and the electrophysiological signal based on the correlation statistical data, thereby obtaining correlation data;

[0156] The present invention uses correlation statistics to quantitatively assess the degree of correlation between label-free optical detection (such as optical cardiac imaging) and electrophysiological signals. The assessment is based on the mean and confidence interval of the correlation statistics. If the correlation is strong, it can be considered that the label-free optical signal has the potential to predict electrophysiological activity. This step is commonly used in cardiac research to evaluate the correlation between different detection methods and help determine the effectiveness of noninvasive detection methods.

[0157] Step S55: Optimizing the parameters of the electrophysiological signal prediction model according to the model performance evaluation data and the correlation data to achieve the model's prediction accuracy for the electrophysiological signal of the myocardial tissue.

[0158] The present invention optimizes the parameters of the electrophysiological signal prediction model based on model performance evaluation data (such as error data and signal-to-noise ratio) and correlation data. Model optimization involves adjusting hyperparameters such as the model's learning rate, batch size, and number of training rounds to improve the model's accuracy in predicting myocardial tissue electrophysiological signals. Specifically, the optimization goal is to reduce error while maintaining high prediction accuracy. The resulting electrophysiological signal prediction model is able to more accurately capture the correlation between mechanical contraction and electrical activity.

[0159] By extracting DFD80 and APD80, the present invention captures key features of cardiac contraction and electrophysiological activity, which are of great physiological significance in assessing myocardial tissue function. This extraction provides basic data for subsequent correlation analysis, ensuring the accuracy and scientificity of the analysis. The Pearson correlation coefficient is a commonly used statistical method for quantifying the linear relationship between two variables. By calculating the correlation between mechanical contraction and electrophysiological signals, it can reveal their mutual influence on cardiac function. Understanding this relationship helps identify underlying physiological mechanisms and provides data support for clinical decision-making. Statistical analysis provides a more comprehensive understanding of the relationship between mechanical contraction and electrophysiological signals. The average correlation coefficient provides an overview of the overall degree of association, while the standard deviation and confidence interval reveal the consistency and reliability of the correlation. This analysis provides important information for evaluating the reliability and applicability of the model. Through quantitative evaluation, a clear understanding of the relationship between label-free optical detection and electrophysiological signals is achieved. This evaluation not only improves data interpretation but also helps to demonstrate the effectiveness and reliability of label-free optical detection in predicting electrophysiological signals, providing important evidence for subsequent research and application. By combining model performance evaluation with feature correlation for optimization, model parameters can be adjusted in a targeted manner to improve the model's predictive accuracy. This process not only enhances model performance but also increases its applicability under different experimental conditions. The optimized model is more adaptable to the characteristics of real cardiac electrophysiological signals, thereby improving the ability to monitor and analyze the electrophysiological state of myocardial tissue.

[0160] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0161] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A label-free optical detection method for myocardial tissue electrophysiology based on deep learning, characterized in that: The following steps are involved: Step S1: Acquire optical detection image data and electrophysiological signal data of myocardial tissue; Performing image preprocessing on the optical detection image data and the electrophysiological signal data to obtain optical-physiological signal mapping data, wherein the image preprocessing includes image registration, image cropping, periodic enhancement, time difference enhancement, spatial filtering, intensity normalization, and data smoothing; Step S2: extracting characteristic signals based on the mechanical contraction curve and the electrophysiological curve according to the optical-physiological signal mapping data, thereby obtaining mechanical contraction curve data and electrophysiological curve data; Step S3: Using the mechanical contraction curve data as an independent variable and the electrophysiological curve data as a dependent variable, a gated recurrent unit network model is trained to construct and train the model, thereby obtaining an electrophysiological signal prediction model, wherein the electrophysiological signal prediction model includes two hidden layers, and the width of each hidden layer is 64; Step S3 includes: Step S31: using the mechanical contraction curve data as an independent variable and the electrophysiological curve data as a dependent variable, and matching them based on the time series to form a training data set; Step S32: Divide the training data set into a training set, a validation set, and a test set based on a ratio of 80:10:10; Step S33: Initializing a gated recurrent unit network model based on the mechanical contraction curve data and the electrophysiological curve data, and setting hyperparameters of the gated recurrent unit network model based on the learning rate, optimizer type, batch size, and number of training rounds, thereby obtaining an initial GRU network model; Step S34: input the training set into the initial GRU network model for training, calculate the predicted value through forward propagation, and then adjust the weight through backpropagation to reduce the prediction error, thereby obtaining a trained dynamic time series prediction model; Step S35: Validate the dynamic time series prediction model using the validation set, and adjust the training strategy based on the error of the validation result, thereby obtaining an adaptive spatiotemporal attention GRU model; Step S36: Performing a performance test based on mean square error and root mean square error on the adaptive spatiotemporal attention GRU model using the test set to obtain model performance evaluation data; Step S37: performing ensemble learning optimization processing on the adaptive spatiotemporal attention GRU model according to the model performance evaluation data, thereby obtaining an electrophysiological signal prediction model; Step S4: using the electrophysiological signal prediction model to predict changes in cardiac electrophysiological activity based on the mechanical contraction curve data, thereby obtaining electrophysiological prediction data; performing a prediction performance evaluation on the electrophysiological prediction data based on relative error, mean square error, and root mean square error, thereby obtaining model performance evaluation data; Step S5: Extract characteristic indicators based on DFD80 and APD80 for the mechanical contraction curve data and the electrophysiological curve data, and calculate the Pearson correlation coefficient of the characteristic indicators to obtain correlation data between label-free optical detection and electrophysiological signals; optimize the parameters of the electrophysiological signal prediction model based on the model performance evaluation data and the correlation data to achieve the model's prediction accuracy for the electrophysiological signals of myocardial tissue.

2. The method for label-free optical detection of myocardial tissue electrophysiology based on deep learning according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: collecting optical detection image data of myocardial tissue under experimental conditions through an optical imaging system, wherein the optical imaging system is specifically a fluorescent calcium imaging combined with a mechanical wave imaging system; Step S12: synchronously collecting electrophysiological signal data of myocardial tissue using a calcium imaging system, wherein the electrophysiological signal data includes a calcium ion concentration change signal; Step S13: performing image registration on the optical detection image data and the electrophysiological signal data based on the field of view size, position and angle, thereby obtaining image consistency data; Step S14: performing image cropping processing on the image consistency data to obtain circular consistent image data, wherein the image cropping processing specifically crops the original rectangular field of view into a circular area with the center of the rectangle as the circle center and the length of the short side of the rectangle as the diameter; Step S15: performing periodic cyclic shifting and superposition on the circular consistent image data to obtain periodic enhanced data; Step S16: performing signal enhancement and optimization processing based on time difference enhancement, spatial filtering, intensity normalization and data smoothing on the periodic enhancement data, thereby obtaining optical-physiological signal mapping data.

3. The method for label-free optical detection of myocardial tissue electrophysiology based on deep learning according to claim 2, characterized in that: Step S15 includes the following steps: Step S151: performing Fourier transform on the circular consistent image data and determining the main cycle of myocardial tissue contraction, thereby obtaining periodic characteristic data; Step S152: determining the cyclic shift step length according to the periodic characteristic data, thereby obtaining shift step length data; Step S153: cyclically shifting the circular consistent image data input signal according to the shift step data to generate multiple groups of shifted image data, thereby obtaining a cyclically shifted data set; Step S154: performing pixel-level weighted superposition on the images in the cyclically shifted data set to obtain periodically enhanced data, wherein the weights are distributed in inverse proportion to the shift amount.

4. The method for label-free optical detection of myocardial tissue electrophysiology based on deep learning according to claim 3, characterized in that: Step S16 includes the following steps: Step S161: setting time difference parameters for the periodic enhancement data according to the actual cell waveform conduction velocity, thereby obtaining time difference parameter data; Step S162: Calculating the absolute value of the difference between the current value of each pixel position and the value at a past time point determined according to the time difference parameter data based on the periodic enhancement data, thereby obtaining time difference enhancement data; Step S163: performing spatial filtering processing based on the Bartlett filter on the time difference enhancement data to obtain spatial filtered data; Step S164: performing the maximum and minimum values ​​of the time series of each pixel on the spatial filtering data to obtain pixel intensity range data; Step S165: performing linear normalization on each pixel according to the pixel intensity range data, and mapping the value to the interval [0, 1], thereby obtaining intensity normalized data; Step S166: performing data smoothing processing on the intensity normalized data to obtain optical-physiological signal mapping data.

5. The method for label-free optical detection of myocardial tissue electrophysiology based on deep learning according to claim 4, characterized in that: Step S166 includes the following steps: Step S1661: Calculate the power spectrum density of the intensity normalized data and determine the main frequency components of the signal to obtain frequency characteristic analysis data; Step S1662: determining smoothing method parameters based on Rloess according to the frequency feature analysis data, thereby obtaining smoothing method parameter data, wherein the window size parameter in the smoothing method parameter data is set to 20% of the acquisition frame rate; Step S1663: smoothing each pixel time series of the intensity normalized data according to the smoothing method parameter data, and calculating the mean square error before and after smoothing, thereby obtaining smoothing effect evaluation data; Step S1664: performing edge preservation processing on pixels with poor smoothing effect based on the smoothing effect evaluation data, thereby obtaining edge preservation data, wherein the edge preservation processing specifically involves detecting sudden changes and edges in the signal and performing local adjustments; Step S1665: analyzing the signal changes between adjacent frames of the edge preservation data and performing time consistency correction to obtain time-corrected data; Step S1666: Calculating the signal-to-noise ratio change before and after smoothing the time-corrected data to obtain signal-to-noise ratio improvement data; Step S1667: performing weighted fusion on the time-corrected data and the intensity-normalized data according to the signal-to-noise ratio improved data, thereby obtaining optical-physiological signal mapping data.

6. The method for label-free optical detection of myocardial tissue electrophysiology based on deep learning according to claim 5, characterized in that: Step S2 includes the following steps: Step S21: extracting local displacement information and spatial variation trend of myocardial contraction deformation in the image from the optical-physiological signal mapping data, thereby constructing a mechanical contraction curve; Step S22: mapping the electrical signal intensity fluctuations at each pixel in the optical-physiological signal data, and using time as the horizontal coordinate and the electrical signal intensity as the vertical coordinate to obtain an electrophysiological curve reflecting the electrical activity characteristics of the myocardial cells; Step S23: dividing the mechanical contraction curve and the electrophysiological curve into cardiac contraction cycles, determining the start and end time points of each cardiac cycle, and calibrating the peak and valley values, thereby obtaining cardiac cycle segmentation data; Step S24: performing feature extraction based on contraction amplitude, maximum contraction rate, peak and valley values ​​of the electrical signal, and rise time on the mechanical contraction curve and the electrophysiological curve in the cardiac cycle according to the cardiac cycle segmentation data, thereby obtaining feature vector data; Step S25: performing scale difference elimination processing on the feature vector data between different samples, and mapping the data values ​​to a unified numerical interval, thereby obtaining mechanical contraction curve data and electrophysiological curve data.

7. The method for label-free optical detection of myocardial tissue electrophysiology based on deep learning according to claim 6, characterized in that: Step S33 includes the following steps: The gated recurrent unit network model is initialized according to the mechanical contraction curve data and the electrophysiological curve data, wherein the gated recurrent unit network model includes an input layer, two hidden layers and an output layer. The input dimension of the input layer is the time step of the mechanical contraction curve; the number of GRU units in the first hidden layer is 64, and its task is to extract primary temporal features from the input mechanical contraction curve data; the number of GRU units in the second hidden layer is 64, which receives the output from the first hidden layer and performs high-level feature integration and complex pattern recognition; the output dimension of the output layer is the time step of the electrophysiological curve; the hyperparameters of the gated recurrent unit network model are set based on the learning rate, optimizer type, batch size and number of training rounds to obtain the initial GRU network model.

8. The method for label-free optical detection of myocardial tissue electrophysiology based on deep learning according to claim 7, characterized in that: Step S4 includes the following steps: Step S41: inputting the new mechanical contraction curve data into the electrophysiological signal prediction model, and predicting changes in cardiac electrophysiological activity through forward propagation calculation, thereby obtaining electrophysiological prediction data; Step S42: comparing the electrophysiological prediction data with the actually measured electrophysiological data, and calculating the relative error, mean square error, and root mean square error to obtain prediction error data; Step S43: performing prediction performance evaluation on the electrophysiological signal prediction model at different time scales and signal amplitude ranges based on the prediction error data, thereby obtaining multi-scale performance evaluation data; Step S44: Visualizing the prediction results based on the multi-scale performance evaluation data to obtain visualized performance evaluation data, wherein the visualization includes generating a comparison chart of the predicted value and the actual value and an error distribution chart; Step S45: Generate a detailed model performance report for the multi-scale performance evaluation data and the visualization performance evaluation data, including the overall prediction accuracy, statistical analysis of various error indicators, and the performance of the model under different conditions, thereby obtaining model performance evaluation data.

9. The method for label-free optical detection of myocardial tissue electrophysiology based on deep learning according to claim 8, characterized in that: Step S5 includes the following steps: Step S51: performing DFD80 extraction on the mechanical contraction curve data to obtain mechanical contraction characteristic data; performing APD80 extraction on the electrophysiological curve data to obtain electrophysiological characteristic data; Step S52: Calculating the Pearson correlation coefficient of the mechanical contraction characteristic data and the electrophysiological characteristic data to obtain correlation data; Step S53: Calculate the average correlation coefficient, standard deviation, and confidence interval of the correlation data to obtain correlation statistics; Step S54: quantitatively evaluating the degree of correlation between the label-free optical detection and the electrophysiological signal based on the correlation statistical data, thereby obtaining correlation data; Step S55: Optimizing the parameters of the electrophysiological signal prediction model according to the model performance evaluation data and the correlation data to achieve the model's prediction accuracy for the electrophysiological signal of the myocardial tissue.

Images