Cformer: A Hybrid CNN-Transformer Framework for 12-Lead ECG Denoising

The electrocardiogram (ECG) signal serves as a crucial medical indicator, is prone to interference from various noise sources, leading to misdiagnosis of heart disease. However, previous denoising methods focus more on single-lead ECG, but ignore the correlation of each lead, which is not suitable for 12-lead ECG denoising. In this paper, we propose a novel hybrid CNN-Transformer framework named Cformer to explore the 12-lead ECG denoising method, which mainly includes a structured self-attention (SSA) module and a multi-scale feature aggregation (MSFA) module. For the SSA module, we use a two-stage attention strategy to capture the long-range dependencies and global features between different ECG leads according to their categories. For the MSFA module, we fuse local and global features through convolutions of different scales, which gives our model a wider field of view and produces better performance. Besides, we employ channel and spatial attention to make our model adaptively pay more attention to the critical ECG leads and time series. Experimental results show that our method achieves state-of-the-art performance compared with others on the 12-lead ECG benchmark PTB-XL dataset for four different types of noise.