A lightweight U-net for ECG denoising using knowledge distillation

Objective. Electrocardiogram (ECG) signals are easily polluted by various noises which are likely to have adverse effects on subsequent interpretations. Research on model lightweighting can promote the practical application of deep learning-based ECG denoising methods in real-time processing. Approach. Firstly, grouped convolution and conventional convolution are combined to replace the continuous conventional convolution in the model, and the depthwise convolution with stride is used to compress the feature map in the encoder modules. Secondly, additional identity connections and a local maximum and minimum enhancement module are designed, which can retain the detailed information and characteristic waveform in the ECG waveform while effectively denoising. Finally, we develop knowledge distillation in the experiments, which further improves the ECG denoising performance without increasing the model complexity. The ground-truth ECG is from The China Physiological Signal Challenge (CPSC) 2018, and the noise signal is from the MIT-BIH Noise Stress Test Database (NSTDB). We evaluate denoising performance using the signal-to-noise ratio (SNR), the root mean square error (RMSE) and the Pearson correlation coefficient ( P ). We use the floating point of operations (FLOPs) and parameters to calculate computational complexity. Main Results. Different data generation processes are used to conduct experiments: group 1, group 2 and group 3. The results show that the proposed model (U Lde -net) can improve SNRs by 10.30 dB, 12.16 dB and 12.61 dB; reduce RMSEs by 9.88 × 10 −2 , 20.63 × 10 −2 and 15.25 × 10 −2 ; and increase P s by 14.77 × 10 −2 , 27.74 × 10 −2 and 21.32 × 10 −2 . Moreover, the denoising performance after knowledge distillation is further improved. The U Lde -net has parameters of 6.9 K and FLOPs of 6.6 M, which are much smaller than the compared models. Significance. We designed a lightweight model, but also retain adequate ECG denoising performance. We believe that this method can be successfully applied to practical applications under time or memory limits.