Implementation method of multi-branch enhanced diagnostic network for coronary artery occlusion localization

Abstract

The present invention discloses a method for implementing a multi-branch enhanced diagnostic network for locating coronary artery occlusion. The present invention adopts a multi-branch enhanced network, including a local feature extraction module and a global feature extraction module. The local feature extraction module uses a wavelet attention mechanism to enhance the local feature representation, and the global feature extraction module uses a real-virtual enhanced Fourier neural operator to extract global features in the frequency domain; the multi-branch enhanced network is divided into five branches, four of which are used for lesion detection in different coronary artery regions, and the fifth branch fuses 12-lead signals to extract global frequency domain features. The defined adaptive misclassification penalty loss function effectively addresses the problem of category imbalance by dynamically adjusting the penalty coefficients of false negative and false positive samples. The present invention significantly improves the accuracy and robustness of electrocardiogram signal analysis, and provides an efficient and non-invasive detection tool for the early diagnosis of coronary heart disease.

Description

Technical Field

[0001] The present invention relates to the field of cardiovascular disease diagnosis, and specifically to a method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion localization. The method adopts a multi-branch enhanced coronary artery occlusion localization network (MCao-Net), combined with frequency domain feature processing and time-frequency local feature extraction technology, to significantly improve the accuracy and robustness of electrocardiogram signal analysis. MCao-Net is suitable for spatial localization of coronary artery lesions, and provides an efficient non-invasive detection tool for early diagnosis and treatment of clinical coronary heart disease. Background Art

[0002] Coronary artery disease is one of the leading causes of death and morbidity worldwide. It is usually caused by coronary artery stenosis or occlusion caused by atherosclerosis, leading to myocardial ischemia or infarction. In clinical practice, electrocardiogram (ECG) is often used to diagnose CAD, but due to individual physiological differences, signal complexity, and data imbalance, ECG still faces many challenges in accurately locating the lesion site.

[0003] Existing methods for localizing coronary artery occlusions mainly rely on the manual extraction of single features, such as ST segment elevation or T wave inversion. Although this method performs well in some simple cases, it still lacks sufficient resolution when faced with complex ECG signals. In addition, traditional models based on time domain analysis are difficult to effectively capture the multi-frequency components and local time-frequency features in ECG signals, thus limiting the precise positioning of the lesion area.

[0004] Accurately locating the location of coronary artery occlusion is crucial for early screening and timely intervention to reduce patient risks. However, existing methods of locating coronary artery occlusion are insufficient when faced with complex electrocardiogram signals. They rely on single feature extraction and have difficulty dealing with multi-frequency components in complex signals, resulting in limited positioning accuracy.

[0005] In order to solve the above problems, the present invention proposes a multi-branch enhancement network based on 12-lead electrocardiogram signals, which combines frequency domain and time-frequency feature analysis technology, and significantly improves the positioning accuracy and robustness of coronary artery occlusion through a multi-branch structure and an adaptive classification loss function. Summary of the invention

[0006] The purpose of the present invention is to provide a method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion positioning, and a method for spatial positioning of coronary artery lesions based on 12-lead electrocardiogram signals implemented by a multi-branch enhanced network MCao-Net. The network adopts a multi-branch enhanced network structure and integrates two advanced modules and an advanced loss function. The multi-branch enhanced network contains four detection branches for different coronary artery regions (left main trunk, left anterior descending branch, left circumflex branch and right coronary artery), and significantly improves the positioning accuracy through multi-channel feature extraction. The real-imaginary enhanced Fourier neural operator module converts the electrocardiogram signal from the time domain to the frequency domain, and independently optimizes the real part (amplitude) and imaginary part (phase) of the signal. The wavelet attention mechanism uses wavelet transform to decompose the electrocardiogram signal at multiple scales, and captures the local lesion signal by extracting time-frequency features. The adaptive misclassification penalty loss function (AMPLoss) is based on binary cross entropy, and effectively addresses the problem of class imbalance by dynamically adjusting the penalty coefficients of false negative and false positive samples. The present invention significantly improves the accuracy and robustness of electrocardiogram signal analysis, and provides an efficient and non-invasive detection tool for early diagnosis of coronary heart disease.

[0007] The technical solution included in the present invention to solve the technical problem includes the following steps:

[0008] S1: A coronary artery disease dataset containing 12-lead electrocardiogram signals collected from multiple datasets;

[0009] S2: preprocessing the collected ECG data, mainly including denoising, filtering, signal normalization, etc.;

[0010] S3: sorting and dividing the preprocessed ECG data into a training set and a test set;

[0011] S4: applying data enhancement technology to the ECG data of step S3;

[0012] S5: Building a multi-branch enhancement network for localization of coronary artery occlusion;

[0013] S6: Use the preprocessed ECG dataset to train the multi-branch enhanced network and obtain weights;

[0014] S7: Evaluate the model performance by using the trained weights on the test set, and adjust the model parameters and architecture based on indicators including but not limited to accuracy (ACC), F1 score (F1), area under the curve (AUC), geometric mean (G-mean), sensitivity (Se), specificity (Sp), and precision (PPV) to optimize the model performance.

[0015] Beneficial effects of the present invention

[0016] The coronary artery occlusion localization method based on MCao-Net proposed in the present invention effectively combines the global and local feature extraction of the electrocardiogram signal through a multi-branch architecture, significantly improving the accuracy of locating the lesion area. In addition, the RieFNO module enhances the expression ability of frequency domain features, and the wavelet attention mechanism further improves the effect of time-frequency feature extraction, thereby effectively improving the processing ability of complex electrocardiogram signals. The introduction of the adaptive classification penalty loss function makes this method more adaptable when dealing with data imbalance problems, and improves the detection accuracy of rare lesions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is an overall flow chart of using a multi-branch enhanced network for coronary artery disease localization in an embodiment of the present invention.

[0018] Figure 2 It is a detailed structural diagram of the multi-branch enhanced network architecture in an embodiment of the present invention.

[0019] Figure 3 It is a detailed structural diagram of the RieFNO module in an embodiment of the present invention.

[0020] Figure 4 Detailed structural diagram of the wKAN module in an embodiment of the present invention.

[0021] Figure 5 It is a heat map of the left trunk branch feature extraction in the embodiment of the present invention.

[0022] Figure 6 It is a heat map of the left anterior descending branch feature extraction in an embodiment of the present invention.

[0023] Figure 7 It is a heat map of the left circumflex branch feature extraction in an embodiment of the present invention.

[0024] Figure 8 is a heat map of the feature extraction of the right coronary artery branches in an embodiment of the present invention.

[0025] Fig. 9 It is a heat map of the feature extraction of the wKAN fusion branch in the embodiment of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further explained below with reference to the embodiments.

[0027] like Figure 1 As shown, the present invention proposes a method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion positioning. The method aims to realize coronary artery occlusion positioning based on electrocardiogram signals, which specifically includes the following steps:

[0028] Step 1: Collect 12-lead ECG signal data from multiple medical image datasets PTB and CardioLead-CAD. These data cover various types of coronary artery occlusive lesions, such as lesions of the left main trunk, left anterior descending artery, left circumflex artery and right coronary artery.

[0029] Step 2: The collected ECG data is subjected to denoising and signal standardization in the preprocessing stage to ensure signal quality. A bandpass filter is used to remove high-frequency noise and low-frequency drift in the ECG signal to ensure the stability of the input signal.

[0030] Step 3: Organize and divide the preprocessed ECG data into training set and test set.

[0031] Step 4: Use enhancement technology on the preprocessed ECG data: Specific enhancement methods include random time offset, amplitude scaling, etc., to ensure that the model can maintain a high generalization ability in different scenarios.

[0032] Step 5. In order to accurately locate the lesion area of ​​coronary heart disease, a multi-branch enhanced network MCao-Net based on 12-lead ECG signals is constructed. This network uses a multi-branch architecture to combine local feature extraction modules with global feature extraction modules to specifically identify lesion features in the ECG; the local feature extraction module uses the wavelet attention mechanism (wKAN) to enhance local feature representation, and the global feature extraction module uses the real virtual enhanced Fourier neural operator (RieFNO) to extract global features in the frequency domain, thereby improving positioning accuracy. The overall architecture of the model is divided into five branches, four of which are used for lesion detection in different coronary artery regions (left main coronary artery, left anterior descending artery, left circumflex artery, right coronary artery), and the fifth branch fuses 12-lead signals to extract global frequency domain features.

[0033] like Figure 2 As shown in the figure, each branch network focuses on a specific local area of ​​the coronary artery and uses the relevant lead signals to extract local features. Among them, the left anterior descending branch (LAD) is specifically used to detect abnormalities in the anterior wall and ventricular septum of the myocardium through the V1 to V4 lead signals. The left circumflex branch (LCX) uses leads I, aVL, V5 and V6 to capture changes in electrical activity in the lateral wall area. The left main branch (LM) monitors a wider area of ​​the myocardium, especially the anterior wall, lateral wall and overall activity of the ventricular septum through leads V1 to V6, I and aVL. The right coronary artery branch (RCA) uses leads II, III, aVF, V1 and V2 to capture the electrical activity of the inferior wall and right ventricle, which can effectively detect abnormalities in the right ventricle and inferior wall. Therefore, the design of each branch is based on the correspondence between the ECG data and the anatomical position of the coronary artery to ensure that the features of a specific lesion area can be accurately extracted.

[0034] like Figure 3 As shown in the figure, the real-imaginary enhanced Fourier neural operator (RieFNO) is mainly used to enhance the feature extraction capability of the signal in the frequency domain. By introducing an independent learnable weight matrix, the real part (amplitude) and the imaginary part (phase) can be optimized at the same time, so that the model can better reflect the complex spectrum characteristics. First, the signal of the input ECG data Perform Fourier transform to convert the time domain signal to the frequency domain to obtain its corresponding real and imaginary parts. The Fourier transform formula is as follows:

[0035] (1)

[0036] in, represents the Fourier transform function, is the complex frequency domain representation, Represent the real and imaginary parts of the Fourier transform result respectively.

[0037] Unlike traditional methods, RieFNO does not process these two parts separately, but introduces multiple sets of learnable weight matrices to perform independent linear transformations on the real and imaginary parts of the frequency domain signal, thereby enhancing the frequency domain features. In the feature enhancement process, a combination strategy of addition and subtraction is adopted to capture the phase and amplitude information of the signal at the same time. The real and imaginary parts of the frequency domain signal reflect the phase and amplitude characteristics respectively. The addition operation captures the amplitude information, and the subtraction operation extracts the phase difference, thereby ensuring that the model fully acquires key frequency domain information. At the same time, this strategy enhances the diversity and distinguishing ability of features. The addition operation highlights the commonality of the signal, and the subtraction operation highlights the difference, which helps to improve the performance of the model in distinguishing normal and abnormal signals. Moreover, this combination meets the analytical requirements of complex operations, maintains the symmetry and consistency of mathematical operations, and avoids the loss of important phase information during feature conversion. These enhanced features are expressed as:

[0038] (2)

[0039] (3)

[0040] in, and denote the enhanced features of the real and imaginary parts, respectively, is the RELU activation function, is the learnable weight matrix, is the bias term, represents the number of input channels. This step ensures independent enhancement of real and imaginary information, enabling the model to capture the complex frequency domain changes caused by lesions.

[0041] Furthermore, the real-imaginary enhanced Fourier neural operator (RieFNO) adds and fuses the enhanced features of the real and imaginary parts to form a new composite feature , which effectively combines the amplitude and phase information. The calculation formula for this composite feature is as follows:

[0042] (4)

[0043] in, is the learnable weight matrix used for fusion processing, This fusion process helps capture the abnormal electrical activity caused by myocardial lesions and better reflect the changes in frequency domain features.

[0044] Then the fused composite features Compared with previous enhancement features and Combine again to further enhance the characteristic expression of the signal. The combined frequency domain features are expressed as follows:

[0045] (5)

[0046] (6)

[0047] in, is the new learnable weight matrix, is a new bias term. This secondary enhancement strategy further improves the representation ability of frequency domain features.

[0048] Next, in order to reduce noise and further optimize the frequency domain features, RieFNO applies a soft threshold function for sparse processing:

[0049] (7)

[0050] in, is the sparseness threshold parameter, Respectively represent the optimized frequency domain features, represents the soft threshold function. This step helps to eliminate unimportant frequency components and make the model focus more on key pathological features.

[0051] Finally, the enhanced refined frequency domain features are remapped back to the time domain to preserve and exploit the diagnostic information enhanced in the frequency domain:

[0052] (8)

[0053] in, Represents the inverse Fourier transform, which is used to obtain the enhanced ECG signal in the time domain .

[0054] like Figure 4 As shown in the figure, the wavelet attention mechanism (wKAN) aims to combine time-frequency feature extraction with the attention mechanism to capture local lesion features in ECG data. The wavelet attention mechanism decomposes ECG data into time-frequency features of different frequency scales through wavelet transform, which can effectively reflect the local lesions caused by coronary heart disease. Compared with traditional time domain or frequency domain analysis methods, the wavelet attention mechanism can enhance the detection effect of lesion areas at multiple scales.

[0055] The channel attention mechanism unit first compresses the spatial dimension of the feature map, extracts different feature representations respectively, and then generates the channel attention weight through the fully connected layer. The calculation formula is as follows:

[0056] (9)

[0057] in, is the sigmoid activation function, and Respectively represent the average pooling and maximum pooling operations in the time dimension, and represents the attention weight, represents the RELU activation function, Represents the input data of the channel attention mechanism unit.

[0058] The spatial attention mechanism unit compresses the channel dimension. First, it performs average pooling and maximum pooling operations on the input data to obtain two pooling feature maps. Then, the two pooling feature maps are concatenated in the channel dimension to form a new input tensor. Finally, a wavelet convolution-based operation is used to capture the important spatial information in the feature map. The whole process is expressed by the following formula:

[0059] (10)

[0060] in, is the sigmoid activation function, represents the wavelet KAN convolution operation, Represents the splicing operation on the channel axis, and the features are obtained after splicing .

[0061] Wavelet KAN convolution extracts richer feature information by combining wavelet transform and standard convolution. The specific process is as follows:

[0062] First, the input features Perform scale transformation and translation operations. This process makes the ECG data consistent at different scales by performing scale transformation and translation operations on the ECG data, so as to better perform wavelet transformation. The standardization formula is as follows:

[0063] (11)

[0064] in, and are learnable scale and translation parameters, They represent the number of channels of wavelet convolution input and output respectively.

[0065] Next, the input signal is transformed using the biorthogonal wavelet as the mother wavelet. Specifically, the low-pass filter coefficients of the biorthogonal wavelet are used here. Perform a one-dimensional convolution operation on the signal to extract the characteristics of the signal:

[0066] (12)

[0067] in, are the low-pass filter coefficients of the biorthogonal wavelet.

[0068] On this basis, the signal The generalized wavelet transform of , which can decompose the signal into multi-scale time-frequency features:

[0069] (13)

[0070] Finally, the wavelet KAN convolution combines generalized wavelet transform and standard convolution operation to extract features. First, generalized wavelet transform is performed on the input signal to obtain the time-frequency features after wavelet transform. One-dimensional convolution operation is performed on the time-frequency features to extract local features I. Sigmoid weighted linear unit (SiLU) activation function is applied to local features I to obtain local features II after nonlinear activation. At the same time, the features of the input Then perform convolution operation to extract nonlinear features. Then add the two results of local feature II and nonlinear feature, fuse the feature information from wavelet transform and nonlinear activation, and finally standardize the addition result through batch normalization operation to improve the training stability and generalization ability of the model. The formula is as follows:

[0071] (14)

[0072] in, represents the convolution operation, is the Sigmoid linear unit activation function, Represents a normalization operation.

[0073] Step 6. Use the preprocessed ECG dataset to train the MCao-Net network. During the training process, a new loss function, the adaptive misclassification penalty loss function (AMPLoss), is proposed to enhance the robustness of the model to class imbalanced data by adaptively adjusting the penalty weights of false negative and false positive samples. The definition of this loss function is based on the binary cross entropy loss (BCE) and the focus adjustment mechanism. First, define the BCE loss as follows:

[0074] (15)

[0075] in, is the true label, is the Sigmoid activation function.

[0076] On this basis, the adaptive misclassification penalty loss (AMPLoss) is defined as:

[0077] (16)

[0078] in, For the positioning index, there are a total of four positioning predictions, represents the predicted probability of positioning, Position the sample balance factor of c to alleviate the sample imbalance problem, The positioning difficult and easy sample balance factor is used to adjust the weight reduction rate of easy-to-classify samples. and represents the penalty coefficients for false negatives and false positives of positioning c, is the binary cross entropy loss for locating c, represents the true label of location c.

[0079] Step 7: Apply the trained weights to the test set to evaluate the segmentation effect. We take the main and key evaluation indicators to evaluate the model, including accuracy (ACC), F1 score (F-measure, F1), area under the curve (AUC), geometric mean (G-mean), sensitivity (Se), specificity (Sp), and precision (PPV). They are defined as follows:

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] Among them, FP is false positive, which represents the number of locations that are judged to have lesions but actually do not have lesions; FN is false negative, which represents the number of locations that are judged to have no lesions but actually have lesions; TN and FP are true negative and true positive, respectively.

[0088] According to the index results, it can be seen that the multi-branch enhanced network framework of the present invention has reached the highest level in all indicators of almost all categories. It can be concluded that the multi-branch enhanced network is extremely competitive in the localization of coronary heart disease lesions based on ECG signals.

[0089] Table 1 Comparison of detection results of CardioLead-CAD dataset

[0090] Network Acc F1 PPV Se AUC G_mean Sp CBAM_Encoder 62.70 50.18 38.82 71.64 61.23 42.03 40.68 CNN_A 63.32 49.39 39.35 67.36 59.97 42.62 43.00 DSCSSA 56.26 49.17 36.01 79.50 54.51 18.32 24.24 ecgTransform 72.20 53.28 54.72 54.97 67.36 57.31 68.77 MLResNet 66.03 51.38 43.87 63.89 64.60 51.05 51.94 ResNet_BiLSTM 56.85 48.97 36.10 78.08 55.13 19.88 25.79 ResNet_CBAMN 72.50 51.80 54.58 51.44 68.78 57.33 72.15 ShuffleGhostNet 69.86 54.69 50.72 62.79 67.83 53.72 58.17 ShuffleNet 66.79 52.96 44.14 67.08 63.53 48.97 50.07 MCADNN 65.43 51.28 42.41 66.62 63.43 46.81 47.38 ECGNet 55.83 47.81 35.50 76.76 53.27 29.30 27.93 MCA_Net 58.71 48.17 36.14 74.60 57.07 23.05 29.14 Our Model 74.67 55.65 62.1 52.7 71.31 61.33 76.75

[0091] Table 1 shows the quantitative results of the CardioLead-CAD dataset. MCao-Net performs best in all key performance indicators, including accuracy, F1 score, precision, area under the curve, geometric mean, and specificity. MCao-Net improves accuracy to 74.67% and F1 score to 55.65%, significantly outperforming other models, especially in the case of multi-indicator balance, which shows that the model has significant advantages in extracting complex ECG signal features.

[0092] The AUC of the trained MCao-Net on the CardioLead-CAD dataset reaches 0.71, indicating that it has high sensitivity and specificity in the low false positive rate area, especially in distinguishing pathological and non-pathological samples.

[0093] Table 2 Comparison of detection results on the PTB dataset

[0094] Network Acc F1 PPV Se AUC G_mean Sp MFB-CBRNN 63.79 - - 54.41 - - 63.04 ML-ResNet 55.74 47.94 56.06 47.58 - - 55.37 MLA-CNN-BiGRU 62.94 - - 63.97 - - 63.00 ML-Net 66.85 - - 62.64 - - 68.70 MCA_Net 68.90 - - 55.35 - - 67.96 SRTNet 68.45 60.50 - 67.95 - - 61.82 Our 85.25 60.53 60.66 70.85 85.25 77.04 87.10

[0095] Table 2 compares the performance of MCao-Net and other models in localizing myocardial infarction on the PTB dataset. MCao-Net significantly outperforms other models in terms of accuracy, F1 score, precision, sensitivity, etc. The accuracy of MCao-Net increased from 68.90% to 85.25%, and the F1 score increased from 47.94% to 60.53%, indicating that the model not only performs well in CAD localization, but also has strong generalization ability in the detection and localization of myocardial infarction.

[0096] In order to verify the effectiveness of each branch, module and loss function in our proposed multi-branch enhancement network framework, we conducted extensive ablation experiments.

[0097] Table 3 Ablation experiment of branch enhancement network performance

[0098] LM LAD LCX RCA Fusion Acc F1 PPV Se AUC G_mean Sp ✓ ✓ ✓ ✓ 73.97 55.24 55.56 55.48 71.60 61.60 73.78 ✓ ✓ ✓ ✓ 73.94 54.87 56.92 55.03 71.02 60.24 72.93 ✓ ✓ ✓ ✓ 73.76 54.59 55.29 54.07 70.72 61.89 75.22 ✓ ✓ ✓ ✓ 73.83 55.03 55.36 55.79 71.25 60.87 72.72 ✓ ✓ ✓ ✓ 72.46 53.74 54.52 54.69 69.90 58.98 70.62 ✓ 73.19 53.66 53.24 54.23 69.09 61.13 73.81 ✓ ✓ ✓ ✓ ✓ 74.67 55.65 62.1 52.7 71.31 61.33 76.75

[0099] The ablation results of the branch-enhanced network performance are shown in Table 3, which shows the impact of different feature extraction branches on network performance. Removing any branch will lead to performance degradation, especially when the wKAN fusion branch is removed, the performance degradation is most obvious, indicating that it plays an important role in integrating global features. Other branches such as the left anterior descending branch and the left circumflex branch are also crucial for feature extraction in specific regions.

[0100] Table 4 Formulas and parameters of different mother wavelets in wKAN module

[0101] Wavelet Type Mother wavelet formula parameter db4 #timg# #timg# sym4 #timg# #timg# bior13 #timg# #timg# gaussian #timg# #timg# haar #timg# #timg# shannon #timg# #timg# meyer #timg# #timg# dog #timg# #timg# morlet #timg# #timg# mexican_hat #timg# #timg#

[0102] Table 5 Ablation analysis of different mother wavelets in wKAN module

[0103] Network Acc F1 PPV Se AUC G_mean Sp db4 74.49 55.02 59.04 54.07 71.71 60.18 74.22 sym4 73.80 53.54 55.05 53.29 71.65 59.78 74.02 bior13 74.67 55.65 62.1 52.7 71.31 61.33 76.75 gaussian 73.31 54.34 54.84 54.73 70.26 60.43 72.74 haar 74.34 53.31 56.49 51.74 71.35 59.68 75.86 shannon 72.42 54.04 53.77 54.95 69.01 60.14 71.43 meyer 74.62 53.71 58.03 51.92 71.28 59.75 76.05 dog 73.45 54.73 54.45 55.48 69.46 61.04 72.76 morlet 72.71 53.34 53.30 54.16 69.85 59.60 71.92 mexican_hat 74.48 53.98 55.66 52.88 71.13 61.06 76.22 Conv_KAN 73.44 53.38 54.51 53.02 71.44 59.88 73.93 Conv 72.95 54.01 52.12 56.35 69.26 60.17 70.75 whole 74.67 55.65 62.1 52.7 71.31 61.33 76.75

[0104] Table 5 shows the ablation analysis of different mother wavelets in the wKAN module. The different mother wavelet formulas and parameters used in the ablation experiment can be obtained in Table 4. The bior1.3 wavelet performs best in terms of accuracy, F1 score, G-mean and other indicators, indicating its effectiveness in multi-scale signal decomposition. Other wavelets, such as db4 and sym4, perform well but cannot handle the complex dynamics of non-stationary signals.

[0105] Table 6 Ablation study of removing RieFNO module in different feature extraction branches

[0106] LM LAD LCX RCA Fusion Acc F1 PPV Se AUC G_mean Sp ✓ ✓ ✓ ✓ 73.32 54.80 54.78 56.46 70.51 59.90 70.81 ✓ ✓ ✓ ✓ 73.95 55.42 55.80 55.56 70.48 61.92 74.01 ✓ ✓ ✓ ✓ 73.20 54.21 57.63 53.55 70.55 59.55 72.88 ✓ ✓ ✓ ✓ 73.32 54.25 54.37 55.15 70.55 60.41 72.48 ✓ ✓ ✓ ✓ 73.82 54.05 54.75 54.44 69.29 59.85 72.98 ✓ 72.59 53.27 53.40 54.36 68.54 58.58 70.73 ✓ ✓ ✓ ✓ ✓ 74.67 55.65 62.10 52.70 71.31 61.33 76.75

[0107] Table 6 shows the impact of removing the RieFNO module on the model performance in different feature extraction branches. The results show that the RieFNO module has a significant contribution in each branch. Especially in the left trunk branch, when the RieFNO module is removed, the F1 score drops to 54.80%, which shows the key role of the RieFNO module in enhancing the frequency domain feature extraction. In addition, removing the RieFNO module in the LAD branch causes the F1 score to drop to 55.42%, which, although the impact is slightly smaller, still shows its importance in processing complex frequency domain information.

[0108] Table 7 Ablation study of removing wKAN module in different branches of MCao-Net

[0109] LM LAD LCX RCA Fusion Acc F1 PPV Se AUC G_mean Sp ✓ ✓ ✓ ✓ 74.44 55.18 57.92 53.91 70.32 61.45 75.52 ✓ ✓ ✓ ✓ 73.42 54.76 55.78 54.27 71.75 61.33 74.13 ✓ ✓ ✓ ✓ 73.60 54.34 55.29 54.32 71.58 60.82 73.84 ✓ ✓ ✓ ✓ 73.75 53.68 55.30 53.20 71.71 60.14 74.37 ✓ ✓ ✓ ✓ 73.98 54.27 55.33 53.93 72.61 60.76 74.43 ✓ 72.45 53.04 50.43 56.63 69.00 58.87 69.14 ✓ ✓ ✓ ✓ ✓ 74.67 55.65 62.10 52.70 71.31 61.33 76.75

[0110] Table 7 shows the impact of removing the wKAN module on the model performance, especially the performance after removal in different feature extraction branches. After removing the wKAN module, the accuracy of the left trunk branch dropped to 74.44%, and other branches also showed significant performance degradation, especially the decline in F1 score. The removal of the wKAN module weakened the ability of specific branches to capture local time-frequency features, which shows the importance of the wKAN module for multi-scale feature extraction, especially in complex ECG signal scenarios. The wKAN module improves the overall performance of each branch by enhancing the extraction of time and frequency domain features.

[0111] Table 8 Analysis of the impact of different loss functions on MCao-Net performance

[0112] Network Acc F1 PPV Se AUC G_mean Sp BCELoss 73.06 52.72 51.86 54.57 69.44 58.64 71.25 FocalLoss 73.01 54.52 57.37 54.64 70.22 59.40 71.62 CustomLoss 73.61 54.21 54.03 55.10 70.51 60.49 72.80 AMPLoss 74.67 55.65 62.1 52.7 71.31 61.33 76.75

[0113] Table 8 compares several different loss functions, including BCELoss, FocalLoss, CustomLoss, and AMPLoss, and shows their impact on the performance of the MCao-Net model. The results show that AMPLoss performs best in all evaluation indicators, including accuracy (74.67%), F1 score (55.65%), and area under the curve (71.31%). AMPLoss effectively solves the problem of class imbalance by adaptively reweighting the penalty of misclassified samples, especially showing higher robustness in detecting rare and complex coronary heart disease lesions.

[0114] In order to reveal the working mechanism of the MCao-Net model in the localization of coronary artery lesions, a visualization technique based on hook functions was used to generate a heat map by capturing the output of the neural layer in the forward propagation, showing how the model extracts pathological features from ECG signals, especially the characteristic localization of coronary artery branches such as the left main trunk, left anterior descending artery, left circumflex artery and right coronary artery. The specific analysis is as follows;

[0115] Figure 5 Shows the regions activated by the model when detecting the left main coronary artery (LMCA) signal. The heat map highlights the regions that the model focuses on when analyzing the LMCA, which are associated with extensive myocardial ischemia, typical of LMCA occlusion.

[0116] Figure 6 The model's feature activation when detecting left anterior descending (LAD) lesions is shown. The LAD branch is mainly responsible for extracting the electrical activity characteristics of the anterior and septal myocardium. The area focused by the model in the heat map is consistent with the LAD-related lesion characteristics, showing high sensitivity to the anterior wall myocardial electrical signals.

[0117] Figure 7 Demonstrated feature extraction of the model when dealing with a left circumflex artery (LCX) lesion. The model showed significant activation in the high lateral and lateral wall regions, reflecting the effect of left circumflex artery lesions on the ECG signal, which are areas typically supplied by the LCX or the distal left anterior descending artery.

[0118] Figure 8 The feature activations when the right coronary artery (RCA) is lesioned are depicted, especially focusing on the right ventricle and inferior wall regions, which are important myocardial regions supplied by the left circumflex artery. The heat map shows that the model can effectively capture the lesion characteristics in these regions.

[0119] Fig. 9 The activation of the model after fusing the global features from all branches is shown. The wKAN fusion branch generates the final feature representation by combining the global frequency domain information with the local features of each branch. The heat map shows the comprehensive evaluation of the model on the lesion at the global average pooling layer, which enhances the diagnostic accuracy of the overall coronary artery lesions.

[0120] The above contents are only preferred implementation modes of the present invention and do not limit the protection scope of the present invention. It should be emphasized that any minor adjustments and improvements made by technicians in this field without departing from the basic principles of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A method for implementing a multi-branch enhanced diagnostic network for localizing coronary artery occlusion, characterized in that: The steps include: Step 1: Collect a coronary artery disease dataset containing 12-lead electrocardiogram signals from multiple datasets; Step 2: Preprocess the collected ECG data, mainly including denoising, filtering, and signal normalization; Step 3: sort and divide the preprocessed ECG data into a training set and a test set; Step 4: Apply data enhancement technology to the ECG data in step 3; Step 5: Building a multi-branch enhanced network for coronary artery occlusion localization; Step 6: Use the preprocessed ECG dataset to train the multi-branch enhanced network and obtain weights; Step 7: Use the trained weights to evaluate the model performance on the test set; The multi-branch enhancement network includes a local feature extraction module and a global feature extraction module. The local feature extraction module uses the wavelet attention mechanism to enhance the local feature representation, and the global feature extraction module uses the real-virtual enhancement Fourier neural operator to extract the global features in the frequency domain. The multi-branch enhancement network is divided into five branches, four of which are used for lesion detection in different coronary artery regions, and the fifth branch fuses 12-lead signals to extract global frequency domain features. The real-imaginary enhanced Fourier neural operator RieFNO is used to enhance the feature extraction capability of the signal in the frequency domain. By introducing an independent learnable weight matrix, it can optimize the real and imaginary parts at the same time. (1) The input ECG data signal x(t) is subjected to Fourier transform to convert the time domain signal into the frequency domain to obtain its corresponding real part and imaginary part. The Fourier transform formula is as follows: in, represents the Fourier transform function, X(f)=X r (f)+iX i (f) is the complex frequency domain representation, X r (f) and X i (f) represent the real and imaginary parts of the Fourier transform results respectively; (2) The enhanced features after RieFNO are expressed as: O r =φ(W r X r -W i H i +b r ) (2) O i =φ(W i X r +W r X i +b i ) (3) Among them, O r and O i Represent the enhanced features of the real part and the imaginary part respectively, φ is the RELU activation function, is the learnable weight matrix, is the bias term, C in Indicates the number of input channels; (3) The enhanced features of the real part and the imaginary part are added and fused through RieFNO to form a new composite feature H. The calculation formula of the composite feature is as follows: H=φ(W h (X r +X i )+b h ) (4) in, is the learnable weight matrix used for fusion processing, is the bias term; (4) Combine the fused composite feature H with the previous enhanced feature O r and O i Combining again, the combined frequency domain features are expressed as follows: O' r =W' r O r -W' i O i +W' h H+b' r (5) O' i =W' i O r +W' r O i +W' h H+b' i (6) in, is the new learnable weight matrix, is the new bias term (5) RieFNO applies a soft threshold function to perform sparse processing on frequency domain features: O″ r ,O″ i =SoftShrink(O' r ,O' i ;λ) (7) Among them, λ is the sparse threshold parameter, O″ r ,O″ i They represent the optimized frequency domain features respectively, and SoftShrink represents the soft threshold function; (6) Re-map the optimized frequency domain features back to the time domain to retain and utilize the diagnostic information enhanced in the frequency domain: in, represents inverse Fourier transform, through which the enhanced ECG signal x(t) in the time domain is obtained; The implementation of the wavelet attention mechanism wKAN needs to include channel attention mechanism units and spatial attention mechanism units; The channel attention mechanism unit first compresses the spatial dimension of the feature map, extracts different feature representations respectively, and then generates the channel attention weight through the fully connected layer. The calculation formula is as follows: Where σ is the sigmoid activation function, AvgPool(F) and MaxPool(F) represent the average pooling and maximum pooling operations in the time dimension, W1 and W2 represent the attention weights, φ represents the RELU activation function, and F represents the input data of the channel attention mechanism unit; The spatial attention mechanism unit compresses the channel dimension. First, it performs average pooling and maximum pooling operations on the input data to obtain two pooling feature maps respectively. Then, the two pooling feature maps are concatenated in the channel dimension to form a new input tensor. Finally, a wavelet convolution-based operation is used to capture the important spatial information in the feature map. The whole process is expressed by the following formula: Where σ is the sigmoid activation function, f wvA (·) represents the wavelet KAN convolution operation, [;] represents the concatenation operation on the channel axis, and the feature X(τ) is obtained after concatenation.

2. The method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion location according to claim 1, characterized in that: The coronary artery area includes the left main coronary artery, left anterior descending artery, left circumflex artery, and right coronary artery. Each branch network focuses on a specific coronary artery area and uses relevant lead signals to extract local features. The left anterior descending branch uses V1 to V4 lead signals to specifically detect abnormalities in the anterior myocardial wall and ventricular septum. The left circumflex branch uses I, aVL, V5, and V6 leads to capture changes in electrical activity in the lateral wall area. The left trunk branch monitors a wider area of ​​myocardium through V1 to V6, I, and aVL leads, focusing on the overall activity of the anterior wall, lateral wall, and ventricular septum. The right coronary artery branch uses leads II, III, aVF, V1, and V2 to capture the electrical activity of the inferior wall and right ventricle and detect abnormalities of the right ventricle and inferior wall.

3. The method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion location according to claim 1, characterized in that: Wavelet KAN convolution extracts richer feature information by combining wavelet transform and standard convolution. The specific process is as follows: First, the input feature X(τ) is scaled and translated, and the standardization formula is as follows: in, and are learnable scale and translation parameters, C out and C in Respectively represent the number of channels of wavelet convolution input and output; Next, a one-dimensional convolution operation is performed on the signal using the low-pass filter coefficient h0[k] of the biorthogonal wavelet to extract the signal features: The generalized wavelet transform of feature X(τ) is obtained, which can decompose the feature into multi-scale time-frequency features: Finally, the wavelet KAN convolution combines generalized wavelet transform and standard convolution operation to extract features. First, the input signal is generalized wavelet transform to obtain the time-frequency features after wavelet transform; the time-frequency features are subjected to one-dimensional convolution operation to extract local feature I; the sigmoid weighted linear unit activation function is applied to the local feature I to obtain the local feature II after nonlinear activation, and the input feature X(τ) is convolved again to extract nonlinear features; then the two results of local feature II and nonlinear feature are added to fuse the feature information from wavelet transform and nonlinear activation, and finally the addition result is standardized by batch normalization operation, and the formula is expressed as follows: Among them, Conv() represents the convolution operation, SiLU() is the Sigmoid linear unit activation function, and BN() represents the normalization operation.

4. The method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion location according to claim 3, characterized in that: During the training process, an adaptive misclassification penalty loss function is used to adaptively adjust the penalty weights of false negative and false positive samples; the definition of this loss function is based on binary cross entropy loss and focus adjustment mechanism.

5. The method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion location according to claim 4, characterized in that: The implementation of the adaptive misclassification penalty loss function is as follows: First, define the BCE loss as follows: Among them, y∈{0,1} is the true label, is the Sigmoid activation function; Secondly, define the adaptive misclassification penalty loss as: Among them, c is the positioning index, and there are a total of four positioning predictions. represents the predicted probability of positioning, Position the sample balance factor of c to alleviate the sample imbalance problem, γ (c) The positioning difficult and easy sample balance factor is used to adjust the weight reduction rate of easy-to-classify samples. and represents the penalty coefficient for false negative and false positive of positioning c, is the binary cross entropy loss for locating c, y (c) represents the true label of location c.

6. The method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion location according to claim 1, characterized in that: Step 2 uses a bandpass filter to remove high-frequency noise and low-frequency drift in the ECG signal.

7. The method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion location according to claim 1, characterized in that: The data enhancement means described in step 4 include random time offset signal and random scaling signal increase.

8. The method for implementing a multi-branch enhanced diagnostic network for coronary artery occlusion location according to claim 1, characterized in that: Evaluation indicators include accuracy, F1 score, area under the curve, geometric mean, sensitivity, specificity, and precision.

Images