MCao: Multi-branch coronary artery occlusion localization using real-imaginary enhancement Fourier wavelet-KAN

Abstract

Coronary artery disease (CAD) is a highly lethal disease caused primarily by atherosclerosis, which leads to arterial blockage and myocardial ischemia or infarction. Currently, electrocardiography (ECG) is commonly used for CAD diagnosis, but CAD-based diagnosis is challenging due to individual physiological differences, signal complexity, and data imbalance. To address this issue, this study introduces the Multi-Branch Enhanced Coronary Artery Occlusion Localization Network (MCao-Net), which apply a multi-branch neural network to locate coronary artery lesions in specific regions based on 12-lead ECG signals, including the left main coronary artery (LMCA), left anterior descending (LAD), left circumflex artery (LCX), and right coronary artery (RCA). The network incorporates two key modules: Real-Imaginary Enhanced Fourier Neural Operator (RieFNO) for enhancing multi-frequency domain feature extraction, and the wavelet-KAN attention (wKAN) mechanism, which improves the precision of time-frequency localized feature detection. Additionally, the adaptive misclassification penalty loss (AMPLoss) function addresses data imbalance in different arteries, particularly improving the detection of rare lesions. Empirical tests on the CardioLead-CAD dataset demonstrated MCao-Net’s performance, achieving an accuracy of 74.67% and an F1 score of 55.65%. Furthermore, the PTB dataset was employed for a Myocardial Infarction (MI) localization task, functioning as a secondary validation of our model’s core feature extraction components, where an accuracy of 85.25% and an F1 score of 60.53% were achieved. MCao-Net surpassed state-of-the-art methods and has potential for clinical use. The project code is publicly available at https://github.com/IMOP-lab/MCao-Pytorch.git.