A non-invasive method for myocardial transmembrane potential reconstruction and abnormal point location

Abstract

The present invention discloses a non-invasive myocardial transmembrane potential reconstruction and abnormal point location method. The method uses a denoising structure combined with a ConvLSTM network structure and a spatiotemporal dual attention mechanism, fully leveraging the flexibility and generalization advantages of deep learning. The neural network structure replaces the parameters that require manual and continuous adjustment in traditional methods, and the network automatically learns universal features in electrocardiogram data. It also retains the patient's physiological information and a rigorous physical model. The convergence of the algorithm and the accuracy of reconstruction are theoretically supported. The spatiotemporal attention mechanism in the present invention introduces the idea of self-attention into the spatial attention module and the temporal channel attention module, and encodes the global features of the feature map in the spatial and temporal dimensions respectively. Then, the temporal global features and the spatial global features are fused to achieve global dependence on both space and time.

Description

Technical Field

[0001] The present invention belongs to the technical field of cardiac electrophysiological imaging, and in particular relates to a method for non-invasive myocardial transmembrane potential reconstruction and abnormal point positioning. Background Art

[0002] Cardiovascular disease caused by arrhythmias has become one of the leading causes of death and health worldwide. According to the American Heart Association and the World Health Organization, nearly 10 million people die each year from heart disease due to arrhythmias. Despite the advanced medical system in the United States, nearly 380,000 people die each year from coronary heart disease, accounting for a staggering 17% of all deaths in the United States. The Chinese Cardiovascular Disease Report reveals that the situation in my country is equally dire. The number of cardiovascular disease patients currently stands in the hundreds of millions, and the incidence and mortality rates continue to rise, currently exceeding the mortality rate from all other diseases, including cancer. Therefore, the development of diagnostic technologies for heart disease is of vital importance.

[0003] Because the electrical potential on the heart is difficult to measure directly, simple preliminary diagnoses of cardiac abnormalities are often made clinically through methods such as electrocardiograms. Doctors can roughly infer the type of heart disease and the vague location of the lesion by observing the characteristic information of the electrocardiograms of different leads and based on years of clinical experience. However, this is only a simple qualitative assessment. During clinical intraoperative diagnosis and treatment, certain frequent arrhythmias require very precise lesion location information, and judgments based solely on surface electrocardiograms obviously cannot fully meet clinical needs. Invasive potential mapping systems require invading the patient's heart to measure the surface potential of the heart and identify the points that need to be ablated. This can cause damage to the body and has a certain degree of risk. Moreover, clinical use of such systems often only allows for local measurements and partial heart reconstruction, which is not conducive to the complete presentation of cardiac electrical signals.

[0004] In recent years, myocardial transmembrane potential reconstruction technology has developed rapidly. Myocardial transmembrane potential reconstruction is based on the conversion of surface ECG data into myocardial transmembrane potential. However, the number of surface potential leads is much smaller than the number of cardiac nodes to be determined. In addition, high-frequency noise is introduced during the surface ECG data measurement process and the forward model solution process, making the results obtained by direct solution without any constraints non-unique. To address this problem, classic methods include direct regularization methods such as the Tikhonov method. The literature [Hofmann B, Kaltenbacher B, Poeschl C, et al. A convergence rates result for Tikhonov regularization in Banach spaces with non-smooth operators [J]. Inverseproblems, 2007, 23 (3): 987] adds a norm constraint after the data fidelity term. Such direct regularization methods are relatively simple and low-cost, but they all lead to over-smoothing of the reconstruction results, which is not conducive to the diagnosis of cardiac foci and has poor generalization. Therefore, an iterative regularization method was proposed. The literature [Steidl G, Weickert J, Brox T, et al. On the equivalence of soft wavelet shrinkage, total variation diffusion, total variation regularization, and SIDEs [J]. SIAM journal on numerical analysis, 2004, 42 (2): 686-713] proposed the ISTA iterative soft threshold shrinkage algorithm, which eventually converges to the optimal solution through continuous iterations, gradient descent steps and soft threshold shrinkage steps.

[0005] With the rapid development of hardware computing devices and deep learning algorithms, end-to-end deep learning has begun to be applied to the reconstruction of myocardial transmembrane potentials. This is equivalent to directly replacing the physical model in traditional methods with a neural network. Without data fidelity terms and regularization penalties, it directly converts the surface ECG data input into the myocardial transmembrane potential output regression problem. The literature [Dhamala J, Ghimire S, Sapp JL, et al. Bayesian optimization on large graphs via a graph convolutional generative model: Application in cardiac model personalization [C]. International conference on medical image computing and computer-assisted intervention. Springer, Cham, 2019: 458-467] introduced a deep convolutional neural network approach, eliminating the need for the transformation matrix required in traditional methods. Once the end-to-end neural network structure is successfully trained, forward reasoning is simple and highly real-time. However, the training process requires a large data set and is highly dependent on the data distribution of the training set. Due to the difficulty in collecting cardiac potential data in clinical practice, the training set is easily incomplete, resulting in poor generalization performance of the trained model. Most of them directly process the surface ECG data and then regress and output the myocardial transmembrane potential. This network structure needs to learn the characteristics of two domains, which puts a lot of pressure on training.

[0006] Traditional regularization methods require manual determination of the regularization form and parameters, and the iterations are time-consuming and result in low reconstruction accuracy. End-to-end deep learning methods, on the other hand, require extensive data training, lack convergence theory, and are prone to overfitting or underfitting. To address these issues, deep learning methods based on physical models have entered the field of myocardial transmembrane potential reconstruction. The paper [Wang L, Wu W, Chen Y, et al. An ADMM-net solution to the inverse problem of electrocardiology [C]. 2018 5th International Conference on Systems and Informatics (ICSAI). IEEE, 2018: 565-569] proposed the ISTA-Net network structure. This network, based on the traditional ISTA iterative algorithm, replaces the soft threshold shrinkage step with a neural network. The network is embedded in the iterative process and automatically learns the regularization operator and parameters, significantly improving its robustness and generalization.

[0007] However, these methods do not fully consider the spatial global characteristics of ECG data and the temporal correlations during its propagation. Therefore, these efforts still have certain limitations, and reconstruction accuracy still has significant room for improvement. Therefore, based on the depth information of surface ECG signals, considering the temporal and spatial characteristics of the data, and combining it with the patient's physiological information, it is of great clinical significance to simply and non-invasively obtain cardiac potential information and locate abnormal heart foci. Summary of the Invention

[0008] In view of the above, the present invention provides a non-invasive myocardial transmembrane potential reconstruction and abnormal point positioning method, which can achieve non-invasive reconstruction of myocardial transmembrane potential and accurate positioning of heart disease foci.

[0009] A non-invasive myocardial transmembrane potential reconstruction and abnormal point location method comprises the following steps:

[0010] (1) Collect the patient's surface electrocardiogram data and obtain the patient's myocardial transmembrane potential information and the location of the lesion through clinical measurement;

[0011] (2) performing multiple acquisitions and measurements on different patients according to step (1) to obtain a large number of samples, each set of samples including surface electrocardiogram data and corresponding myocardial transmembrane potential information and lesion location information, and then dividing all samples into a training set and a test set;

[0012] (3) Construct a neural network model (CLDA-Net) based on ConvLSTM and spatiotemporal attention mechanism, which includes:

[0013] Initialization module, used to convert surface ECG data into initial myocardial transmembrane potential;

[0014] A denoising module is used to denoise the initial myocardial transmembrane potential;

[0015] The ConvLSTM module is used to encode the denoised myocardial transmembrane potential in the spatial local and temporal dimensions to obtain the corresponding feature encoding data;

[0016] A spatiotemporal attention module is used to perform information processing using an attention mechanism on the feature-encoded data in both the temporal and spatial dimensions, and then fuse them to obtain a spatiotemporal global dependency feature map. The spatiotemporal global dependency feature map can be used to reconstruct the output myocardial transmembrane potential through a convolution block, or to predict the output location information of the lesion point through multi-layer perceptron regression.

[0017] (4) Using the surface ECG data in the training set samples as model input and the myocardial transmembrane potential information or lesion location information as labels, the above network model is trained;

[0018] (5) By inputting the surface ECG data in the test set samples into the trained network model, the corresponding myocardial transmembrane potential information or lesion location information can be directly predicted and output.

[0019] Furthermore, the denoising module is composed of a cascade of four residual learning blocks, each residual learning block is composed of three convolutional layers D1 to D3 connected in sequence, and the output of each convolutional layer is activated by the ReLU function; the convolution kernel size of D1 is 7×7, and the number of output channels is 64; the convolution kernel size of D2 is 1×1, and the number of output channels is 32; the convolution kernel size of D3 is 5×5, and the number of output channels is consistent with the number of input channels of the residual learning block; the output of D3 is added to the input of D1 as the output of the residual learning block.

[0020] Furthermore, the specific operation process of the ConvLSTM module is as follows:

[0021]

[0022]

[0023] g t =tanh(W xc *x t +W hc *h t-1 +b c )

[0024]

[0025]

[0026]

[0027] Where: x t is the input data of the ConvLSTM module for the tth iteration, h t and h t-1 are the output results of the ConvLSTM module at the t-th iteration and the t-1-th iteration, respectively. * indicates the convolution operation. represents the dot product operator, σ() represents the sigmoid activation function, f t 、i t 、g t 、o t are the intermediate results obtained by the ConvLSTM module operation at the tth iteration, W xf 、W hf 、W cf 、W xi 、W hi 、W ci 、W xc 、W hc、W xo 、W ho 、W co are weight matrices to be learned, b f 、b i 、b c 、b o are bias vectors to be learned, and t is a natural number greater than 0.

[0028] Furthermore, the spatiotemporal attention module includes a spatial attention module and a temporal channel attention module. The spatial attention module is based on the self-attention idea, and obtains three different feature vectors A1, A2 and A3 by three 1×1 convolutions on the input data respectively. The transpose of A1 is multiplied by A2 and processed by the softmax function to obtain the attention map X, and then A3 is multiplied by X and added to the input data to obtain a feature map E with spatial global dependency. The processing flow of the temporal channel attention module is the same as that of the spatial attention module, except that the temporal channel dimension is retained when calculating the attention map X, and finally a feature map F with temporal global dependency is obtained. Finally, the feature maps E and F are fused to obtain the spatiotemporal global dependency feature map.

[0029] Furthermore, the process of training the network model in step (4) is as follows:

[0030] 4.1 Initialize model parameters, including the bias vector and weight matrix of each layer, learning rate, and optimizer;

[0031] 4.2 Input the surface ECG data in the training set samples into the model, forward propagate the model output to obtain the corresponding prediction result, and calculate the loss function L between the prediction result and the label;

[0032] 4.3 Based on the loss function L, the optimizer is used to iteratively update the model parameters through the gradient descent method until the loss function L converges and the training is completed.

[0033] Furthermore, when the network model is applied to reconstruct the myocardial transmembrane potential function, the loss function L is expressed as follows:

[0034]

[0035] Where: u is the label, i.e. the myocardial transmembrane potential obtained by clinical measurement, u k The model finally outputs the predicted result of myocardial transmembrane potential, u1 is the output result of the denoising module, λ is the weight coefficient, and || ||2 represents the L2 norm.

[0036] Furthermore, when the network model is applied to the lesion location prediction function, the loss function L is expressed as follows:

[0037]

[0038] Among them: S (x, y, z) is the label, that is, the three-dimensional coordinates of the lesion point obtained by clinical measurement, The model finally outputs the predicted result about the location of the lesion, and || ||2 represents the L2 norm.

[0039] The non-invasive myocardial transmembrane potential reconstruction and abnormal point localization method of the present invention simultaneously uses a denoising structure plus a ConvLSTM network structure and a spatiotemporal dual attention mechanism, fully leveraging the flexibility and generalization advantages of deep learning. It uses a neural network structure to replace the parameters that require manual continuous adjustment in traditional methods, and uses the network to automatically learn the universal features in the ECG data. It also retains the patient's physiological information and a rigorous physical model. The convergence of the algorithm and the accuracy of reconstruction are both theoretically supported.

[0040] The initialization module is retained in the denoising module of the model of the present invention, and domain-to-domain conversion is achieved through linear transformation before data is input into the network, thereby obtaining cleaner input and significantly improving the reconstruction effect. In addition, the present invention uses a ConvLSTM network structure to replace the traditional CNN+LSTM structure. The difference is that convolution layers are added in the process from input to state and from state to state, which reduces the structural redundancy while relying on the structural design of the classic LSTM to extract temporal relationships, and additional convolution layers can be added to extract spatial features to obtain better results. The spatiotemporal attention mechanism in the present invention introduces the idea of self-attetion into the spatial attention module and the temporal channel attention module and encodes the global features of the feature maps in the spatial and temporal dimensions respectively, and then fuses the temporal global features and the spatial global features to achieve global dependence on both space and time. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the overall structure of the network model CLDA-Net of the present invention.

[0042] Figure 2 Schematic diagram of the structure of the denoising module in the model of the present invention.

[0043] Figure 3 Schematic diagram of the structure of the spatiotemporal attention mechanism.

[0044] Figure 4 Schematic diagram of the structure of the spatial attention module.

[0045] Figure 5 Schematic diagram of the structure of the temporal channel attention module.

[0046] Figure 6This is a line graph of the relative error between the reconstructed values and the true values of each model in the ablation experiment.

[0047] Figure 7 The line graph shows the structural similarity between the reconstructed values and the true values of each model in the ablation experiment.

[0048] Figure 8 This is a spatial three-dimensional reconstruction of the location of ectopic pacemakers and infarction scars using deep learning methods of the present invention and existing ReconNet, DR2-Net and CLDA-Net.

[0049] Figure 9 This is a diagram of the myocardial transmembrane waveform reconstruction of different cardiac nodes using the present invention and the existing Tikhonov and TV methods. DETAILED DESCRIPTION

[0050] In order to describe the present invention more clearly, the technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific implementation methods.

[0051] The specific implementation steps of the non-invasive myocardial transmembrane point reconstruction and abnormal point positioning method of the present invention are as follows:

[0052] (1) Obtain the patient's surface ECG data, with the number of leads set to 64.

[0053] (2) Based on the surface ECG data obtained above, the initialization module converts it into an initial solution of the myocardial transmembrane potential. The initialization module converts the surface ECG data domain into the myocardial transmembrane potential domain.

[0054] (3) The myocardial transmembrane potential data obtained in step (2) is denoised by a denoising module, and the obtained cleaner data is used as input data.

[0055] The denoising module of the present invention is modified based on ReconNet and DR2-Net in the field of compressed sensing; Figure 2 As shown in the figure, the entire denoising module contains 4 residual learning blocks. In each residual learning block, the convolution kernel size of the first convolution layer is 7*7, and the number of output channels is 64; the convolution kernel size of the second convolution layer is 1*1, and the number of output channels is 32; the convolution kernel size of the third convolution layer is 5*5, and the number of output channels is the same as the initial input of each residual block, and then superimposed with the input of the residual block to obtain the output of the residual block; each layer of convolution in each residual block is activated by the ReLU function after each convolution.

[0056] (4) Construct a neural network based on ConvLSTM and spatiotemporal attention mechanism, which includes a ConvLSTM module and a spatiotemporal attention mechanism module, wherein the conventional CNN+LSTM structure is replaced by the ConvLSTM module, and the spatiotemporal attention mechanism includes a spatial attention mechanism module and a temporal channel attention mechanism module.

[0057] (5) Based on the input data obtained in step (3), feature encoding in the spatial local and temporal dimensions is performed through the ConvLSTM module, where the ConvLSTM operation process is as follows:

[0058]

[0059]

[0060] g t =tanh(W xc *x t +W hc *h t-1 +b c )

[0061]

[0062]

[0063]

[0064] Where: * represents the convolution operation, W represents the weight matrix to be learned, σ() represents the sigmoid activation function, tanh() represents the tanh activation function, represents the dot product of the matrix (element-by-element multiplication), and b represents the bias vector to be learned.

[0065] (6) Based on the feature encoding obtained in step (5), the fused spatiotemporal global dependency feature map is obtained through the spatiotemporal attention module, such as Figure 3 As shown in Figure 2, the spatiotemporal attention mechanism includes a spatial attention module and a temporal channel attention module.

[0066] The spatial attention module is based on the self-attention idea, such as Figure 4 As shown in the figure, the input feature map A (C*H*W) is respectively subjected to three 1*1 convolutions to obtain three different feature vectors A1, A2 and A3, which are then converted into a size of C*HW; the transpose of A1 is multiplied by A2 and processed by the softmax function to obtain a spatial attention map X of size HW*HW; A3 is multiplied by X and converted back to the size of C*H*W, and then A itself is added to obtain a feature map E with spatial global dependency.

[0067] The temporal channel attention module can be obtained based on the same idea as the spatial attention module, such as Figure 5 As shown in the figure, the difference is that the time channel dimension is retained when calculating the attention map, so the size of the attention map Y in the time channel attention module is C*C, and the other operations are shown by the arrows in the figure, and finally the time global dependent feature map F is obtained.

[0068] (7) Based on the feature map obtained in step (6), the final reconstructed myocardial transmembrane potential is output through the convolution block or the lesion point coordinate value (X, Y, Z) is output through the multi-layer perceptron.

[0069] In the function of reconstructing the myocardial transmembrane potential, the MSE function is used to calculate the loss between the reconstructed myocardial transmembrane potential and the true value, and the output loss of the denoising module is added. The total loss function is:

[0070]

[0071] Where: u represents the true value of the myocardial transmembrane potential, u k represents the reconstructed value output by the network, u1 represents the output of the denoising module, and λ represents the weight coefficient, which is set to 0.01 in this example.

[0072] The three-dimensional coordinate value of the output lesion point uses the L2 norm of the output coordinate value and the true coordinate value to describe its loss, which is expressed as:

[0073]

[0074] Among them: S (x, y, z) represents the true coordinate value of the lesion point, Represents the coordinate value of the network output.

[0075] Combining the above-mentioned initialization and denoising modules, ConvLSTM network structure, spatiotemporal attention mechanism, and convolutional blocks and multi-layer perceptrons used for reconstruction, the model structure of the non-invasive myocardial transmembrane potential reconstruction and abnormal point positioning method of the present invention is as follows: Figure 1 As shown in the figure, the initialization module is responsible for converting the input φ into u0 through linear transformation. The denoising module denoises the initial data u0 of the myocardial transmembrane potential. The ConvLSTM module uses two layers of ConvLSTM. The input-to-state and state-to-state convolution kernel sizes in each layer are set to 5*5, and the number of hidden states in the two layers is 128 and 64 respectively. The structure of the spatiotemporal attention mechanism module is shown in Figure 3As shown in the figure, the convolution kernel size of the encoded feature vector is 1*1, and the output feature maps E and F are fused by element-wise addition to obtain the final feature map. The fused feature map is then passed through a 3*3 convolutional layer to output the reconstructed myocardial transmembrane point or the coordinate value (X, Y, Z) is output through multi-layer perceptron regression.

[0076] In the specific implementation process, the entire algorithm model CLDA-Net of the present invention is tested in the Ubuntu system, where the CPU is Intel(R) Core(TM) i9-7900x CPU@3.30GHz 3.31GHz, the host memory is 32GB RAM, the graphics card model is NVIDIA TITAN RTX, and the main programming environment is Tensorflow1.14 and Python 3.7.

[0077] The deep learning training and testing in this paper were conducted in a PyTorch and Python environment. A single NVIDIA TITAN RTX graphics card with 24GB of video memory was used. Basic settings in CLDA-Net included a batch size of 32, a learning rate of 0.001, 500 training epochs, and the Adam optimizer.

[0078] To verify the effectiveness of the present invention in reconstructing myocardial transmembrane potentials and the accuracy of localizing cardiac foci, we applied the above method to the reconstruction of myocardial transmembrane potentials in a simulated data center, localizing the center of myocardial infarction, and ectopic pacemakers. Furthermore, to verify the effectiveness of the initialization module, denoising module, ConvLSTM module, and dual attention module in the proposed model, we conducted relevant ablation studies. The ablation test results are shown in Table 1.

[0079] Table 1

[0080]

[0081]

[0082] In the experiment, we used six different models: (a) using the output of the denoising network directly as the reconstruction value; (b) adding a ConvLSTM module to the denoising module, then inputting the feature map output by the ConvLSTM module into the output layer and outputting the TMP reconstruction value; (c) directly following the denoising module with the spatiotemporal dual attention module, which is equivalent to the complete network omitting the ConvLSTM module; (d) omitting the denoising module and directly inputting the subsequent ConvLSTM module and spatiotemporal dual attention module after initialization; (e) the complete CLDA-Net omitting the initial initialization module, which is equivalent to the denoising module completing both the denoising task and the reconstruction task of the initial solution; (f) the complete CLDA-Net, including the initialization module, denoising module, ConvLSTM module, and dual attention module. We used the structural similarity and relative error between the reconstructed myocardial transmembrane potential value and the true value as the evaluation indicators for this experiment, calculated as follows:

[0083]

[0084] Among them: μ represents the mean, σ 2 represents the variance, σ uku is u k The covariance of y and u, c is a constant to avoid the denominator being zero.

[0085]

[0086] Where: N represents the number of cardiac nodes.

[0087] like Figure 6 and Figure 7 As shown, the optimization modules proposed in this paper have varying degrees of positive effects on both structural similarity and relative error evaluation metrics. In particular, the ConvLSTM module and the dual attention module significantly enhance feature extraction during the reconstruction process. The addition of the DA module reduces relative error by 11.3% and improves structural similarity by 3.0%. This is because the module utilizes both the temporal channel attention module and the spatial attention module to simultaneously extract spatiotemporal non-local features, effectively mining the data's hidden information. The ConvLSTM module also demonstrates significant results, reducing relative error by 7.3% and improving structural similarity by 3.4%. This module replaces the CNN+LSTM architecture while capturing both spatial local features and global dependencies in the temporal dimension. Furthermore, the denoising module plays a crucial role in this network, ensuring high-quality input for subsequent networks. Its addition reduces relative error by -3.5% and structural similarity by +1.7%, respectively. Inspired by DR2-Net, the addition of a linear transformation initialization module also reduces relative error by -1.4% and structural similarity by +1.1%.

[0088] In order to verify the effect of the CLDA-Net model algorithm of the present invention on the reconstruction of myocardial transmembrane potential, we conducted a quantitative comparative experiment between the present invention and the traditional methods Tikhonov and TV, as well as other deep learning methods ReconNet and DR2-Net. The experimental results are shown in Table 2. The evaluation indicators are consistent with the ablation experiment, using structural similarity (SSIM) and relative error (RE). The results in the table are calculated on the test set (data volume of 15,900 groups), including mean and standard deviation. It can be seen that compared with traditional methods, CLDA-Net greatly improves the reconstruction effect, and compared with the results of other deep learning methods, it also has a relatively obvious advantage. The main reason is that compared with DR2-Net and ReconNet, the method of the present invention pays more attention to the global dependence in time and space dimensions when performing feature extraction, which is more in line with the particularity of ECG data.

[0089] Table 2

[0090]

[0091] In order to qualitatively analyze the reconstruction effect of CLDA-Net in the spatial dimension, we performed spatial 3D reconstruction of the ectopic pacemaker location and infarct scar, and compared three different deep learning methods: ReconNet, DR2-Net, and CLDA-Net. The results are shown in Figure 2. Figure 8 It can be seen that the present invention is closest to the true value in terms of both ectopic pacemaker location and infarct scar reconstruction, which demonstrates that the present invention is very powerful in terms of global spatial feature extraction.

[0092] In order to qualitatively analyze the effect of CLDA-Net on the reconstruction of myocardial transmembrane potential waveform in the time dimension, we compared the reconstruction of myocardial transmembrane potential waveforms of different nodes on the heart with the two traditional methods of Tikhonov and TV. The results are as follows: Figure 9 As shown in the figure, compared with the two traditional methods, the results of CLDA-Net suppress a large amount of noise and reconstruct waveform shape, pacing time, action potential amplitude, and excitation duration very close to the true value. This shows that the denoising module, ConvLSTM module, and dual attention module in the present invention play a very significant role in the reconstruction process.

[0093] To verify the effectiveness of CLDA-Net in locating the ectopic pacemaker (EP) and myocardial infarction (MI) scar center in simulated data, comparative experiments were conducted with RandomForest (RF), LightGBM, SVM, Res-Conv1D, and Seq2Seq methods. The experimental results are shown in Table 3 (experimental results based on 64-lead BSP as input). It can be seen that CLDA-Net far outperforms other methods in both ectopic pacemaker localization and infarct center localization, achieving EP (5.62 mm) and MI (5.83 mm), an improvement of 30.4% (EP) and 29.1% (MI) relative to RF. In addition, it was found that the positioning accuracy of ectopic pacemakers was better than that of myocardial infarction to a certain extent. In fact, from the previous reconstruction experiments, it can also be found that the network's reconstruction effect on pacing time points is better than that on action potential amplitudes. Because most nodes in a heart in the simulation data are normal, the number of nodes with ischemic infarction is relatively far less, so the training effect will be slightly worse in the case of small samples.

[0094] Table 3

[0095]

[0096] To verify the redundancy of the high-lead surface ECG data required for reconstruction, we conducted comparative positioning experiments using 12-lead and 64-lead inputs. The results are shown in Table 4. CLDA-Net achieves the highest positioning accuracy compared to other methods for both high-lead and standard 12-lead inputs. Furthermore, when the number of leads is significantly reduced to the standard 12-lead model, the positioning accuracy remains well below 10 mm. Therefore, if only the coordinates of the lesion location are needed, a standard 12-lead ECG with fewer leads can be used. In clinical settings, the acquisition cost and difficulty of a standard 12-lead ECG are significantly lower than those of a high-lead BSP.

[0097] Table 4

[0098]

[0099]

[0100] The above description of the embodiments is intended to facilitate understanding and application of the present invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention should fall within the scope of protection of the present invention.

Claims

1. A non-invasive myocardial transmembrane potential reconstruction and abnormal point location method, comprising the following steps: (1) Collect the patient's surface electrocardiogram data and obtain the patient's myocardial transmembrane potential information and the location of the lesion through clinical measurement; (2) performing multiple acquisitions and measurements on different patients according to step (1) to obtain a large number of samples, each set of samples including surface electrocardiogram data and corresponding myocardial transmembrane potential information and lesion location information, and then dividing all samples into a training set and a test set; (3) Construct a neural network model based on ConvLSTM and spatiotemporal attention mechanism, which includes: Initialization module, used to convert surface ECG data into initial myocardial transmembrane potential; A denoising module is used to denoise the initial myocardial transmembrane potential; The ConvLSTM module is used to encode the denoised myocardial transmembrane potential in the spatial local and temporal dimensions to obtain the corresponding feature encoding data; A spatiotemporal attention module is used to perform information processing using an attention mechanism on the feature-encoded data in both the temporal and spatial dimensions, and then fuse them to obtain a spatiotemporal global dependency feature map. The spatiotemporal global dependency feature map can be used to reconstruct the output myocardial transmembrane potential through a convolution block, or to predict the output location information of the lesion point through multi-layer perceptron regression. (4) Using the surface ECG data in the training set samples as model input and the myocardial transmembrane potential information or lesion location information as labels, the above network model is trained; (5) By inputting the surface ECG data in the test set samples into the trained network model, the corresponding myocardial transmembrane potential information or lesion location information can be directly predicted and output.

2. The non-invasive myocardial transmembrane potential reconstruction and abnormal point location method according to claim 1, characterized in that: The denoising module consists of a cascade of four residual learning blocks, each of which consists of three convolutional layers D1 to D3 connected in sequence. The output of each convolutional layer is activated by the ReLU function; the convolution kernel size of D1 is 7×7, and the number of output channels is 64; the convolution kernel size of D2 is 1×1, and the number of output channels is 32; the convolution kernel size of D3 is 5×5, and the number of output channels is consistent with the number of input channels of the residual learning block; The output of D3 is added to the input of D1 and used as the output of the residual learning block.

3. The non-invasive myocardial transmembrane potential reconstruction and abnormal point location method according to claim 1, characterized in that: The specific operation process of the ConvLSTM module is as follows: g t =tanh(W xc *x t +W hc *h t-1 +b c ) Where: x t is the input data of the ConvLSTM module for the tth iteration, h t and h t-1 are the output results of the ConvLSTM module at the t-th iteration and the t-1-th iteration, respectively. * indicates the convolution operation. represents the dot product operator, σ() represents the sigmoid activation function, f t 、i t 、g t 、o t are the intermediate results obtained by the ConvLSTM module operation at the tth iteration, W xf 、W hf 、W cf 、W xi 、W hi 、W ci 、W xc 、W hc 、W xo 、W ho 、W co are weight matrices to be learned, b f 、b i 、b c 、b o are bias vectors to be learned, and t is a natural number greater than 0.

4. The non-invasive myocardial transmembrane potential reconstruction and abnormal point location method according to claim 1, characterized in that: The spatiotemporal attention module includes a spatial attention module and a temporal channel attention module. The spatial attention module is based on the self-attention idea. It obtains three different feature vectors A1, A2 and A3 by three 1×1 convolutions on the input data respectively. The transpose of A1 is multiplied by A2 and processed by the softmax function to obtain the attention map X. Then, A3 is multiplied by X and added to the input data to obtain a feature map E with spatial global dependency. The temporal channel attention module has the same processing flow as the spatial attention module, except that the temporal channel dimension is retained when calculating the attention map X. Finally, a feature map F with temporal global dependency is obtained. Finally, the feature maps E and F are fused to obtain the spatiotemporal global dependency feature map.

5. The non-invasive myocardial transmembrane potential reconstruction and abnormal point location method according to claim 1, characterized in that: The process of training the network model in step (4) is as follows: 4.1 Initialize model parameters, including the bias vector and weight matrix of each layer, learning rate, and optimizer; 4.2 Input the surface ECG data in the training set samples into the model, forward propagate the model output to obtain the corresponding prediction result, and calculate the loss function L between the prediction result and the label; 4.3 Based on the loss function L, the optimizer is used to iteratively update the model parameters through the gradient descent method until the loss function L converges and the training is completed.

6. The non-invasive myocardial transmembrane potential reconstruction and abnormal point location method according to claim 5, characterized in that: When the network model is applied to reconstruct the myocardial transmembrane potential function, the expression of the loss function L is as follows: Where: u is the label, i.e. the myocardial transmembrane potential obtained by clinical measurement, u k The model finally outputs the predicted result of myocardial transmembrane potential, u1 is the output result of the denoising module, λ is the weight coefficient, and || ||2 represents the L2 norm.

7. The method for non-invasive myocardial transmembrane potential reconstruction and abnormal point location according to claim 5, characterized in that: When the network model is applied to the lesion location prediction function, the expression of the loss function L is as follows: Among them: S (x, y, z) is the label, that is, the three-dimensional coordinates of the lesion point obtained by clinical measurement, The model finally outputs the predicted result about the location of the lesion, and || ||2 represents the L2 norm.

Images