MCao: Multi-Branch Coronary Artery Occlusion Localization Using Real-Imaginary Enhancement Fourier Wavelet-KAN
Figure 1: Detailed network structure of our proposed MCao.
Figure 2: The system structure of our proposed CAD localization architecture.
The proposed MCao-Net is an advanced ECG-based diagnostic model tailored for localizing coronary artery lesions. This network combines multi-branch feature extraction with the Real-Imaginary Enhancement Fourier Neural Operator (RieFNO) and the wavelet-KAN attention module (wKAN) to address the complexities of ECG signals. It outperforms 12 other recent methods on the CardioLead-CAD and PTB datasets, establishing state-of-the-art accuracy in identifying key coronary artery occlusions.
We will first introduce our method and underlying principles, explaining how MCao-Net uses specialized branches and attention mechanisms to improve feature extraction from ECG signals. Next, we provide details on the experimental setup, performance metrics, and GitHub links to previous methods used for comparison. Finally, we present the experimental results, showing how MCao-Net achieves high performance across multiple datasets.
We run MCao and previous methods on a system running Ubuntu 22.04, with Python 3.9, PyTorch 2.0.0, and CUDA 11.8.
Figure 3: Comparison of ECG detection performance between MCao-Net and other methods on the CardioLead-CAD dataset.
Figure 4: Comparison of ECG detection performance between MCao-Net and other methods on the PTB dataset.
Our method demonstrates the best performance in accuracy and sensitivity for coronary artery lesion localization, surpassing previous models on both the CardioLead-CAD and PTB datasets. The integration of the Real-Imaginary Enhancement Fourier Neural Operator (RieFNO) and wavelet-KAN attention module (wKAN) significantly enhances MCao-Net’s capability to detect rare and complex lesion patterns, thereby improving its precision in ECG-based coronary artery disease diagnosis.
Figure 5: Ablation study assessing the impact of individual branch-specific enhancements on network performance in the CardioLead-CAD dataset.
Figure 6: Ablation analysis of the effectiveness of different mother wavelets in the wKAN module using the CardioLead-CAD dataset.
Figure 7: Ablation study of the RieFNO module across different feature extraction branches in the MCao-Net architecture.
Figure 8: Ablation study on the performance impact of removing the wKAN module from specific feature extraction branches in MCao-Net.
Figure 9: Effect of different loss function variants on the performance of the MCao-Net model in ECG signal classification.
Figure 9: Heatmaps generated by MCao-Net showing feature extraction intensities across coronary artery branches.
Each image corresponds to a specific input category: Absence of CAD, LMCA, LAD, LCX, and RCA. The heatmaps capture outputs from targeted neural layers during forward propagation, illustrating how the model identifies pathological features in ECG signals. Specifically: (a) LMCA branch focuses on left main coronary artery signals, (b) LAD branch detects features related to the left anterior descending artery, (c) LCX branch highlights left circumflex artery features, (d) RCA branch focuses on right coronary artery signals, and (e) wKAN fusion branch integrates global features from all coronary territories at the global average pooling layer. Each heatmap demonstrates how the model distinguishes between normal and pathological cases for specific coronary branches.
if you have any questions, please contact '[email protected]'