Multi-Channel Fetal ECG Denoising With Deep Convolutional Neural Networks

Figures

• 

  Figure 1. The architecture of the proposed multi-channel fetal ECG denoising network. The network consists of an encoder of eight convolutional layers and a decoder of eight transposed convolutional layers, which are linked symmetrically by skip connections.

• 

  Figure 2. Detailed illustration of the way that the skip connections (represented by the arrows) are applied in the network. Only two skip connections are shown for simplicity. Conv stands for convolution and ConvT for transposed convolution.

• 

  Figure 3. Performance of the proposed multi-channel convolutional network in comparison with other denoising methods in terms of improvement in SNR of the denoised fetal ECG signals when compared with the noisy ones.

• 

  Figure 4. Denoising results by the proposed method for two simulated signals (SA and SB) of the test dataset. For both signals: each panel in the left presents one channel of the noisy four-channel fetal ECG signal (red), in the middle the corresponding channels of the clean signal (blue) are shown and in the right the denoised fetal ECG signal by our network (green). The horizontal axis depicts the samples at 500 Hz, while the vertical the amplitude of the signals. The SNR values of the noisy and the denoised fetal ECG for both signals are given in Table 2 (SNRin and SNRout, respectively).

• 

  Figure 5. Performance of the proposed fetal ECG denoising method in a large real dataset (28) in terms of improvement in SNR of the denoised signals.

• 

  Figure 6. The results of denoising two signals (RA and RB) from our private real fetal ECG dataset (28). For each signal: the noisy four-channel fetal ECG signal extracted from the abdominal measurements is presented in the left (red), the running median of 100 heartbeats for each channel in the middle (blue) and the denoised fetal ECG signal by our network in the right (green). The horizontal axis depicts the samples at 500 Hz, while the vertical the amplitude of the signals in μV. The SNR values of the noisy fetal ECG signals (SNRin) together with the values for the denoised ones (SNRout) are given in Table 3.

• 

  Figure 7. The result of denoising a real fetal ECG signal from our private dataset (28) with different denoising algorithms. For simplicity, only one channel is displayed. The panels show: (A) the noisy extracted fetal ECG, (B) the denoised signal by the proposed method, (C) the denoised signal by the single-channel denoising network, (D) the result of 30-complex averaging and (E) the result after wavelet denoising. The horizontal axis depicts the samples at 500 Hz, while the vertical the amplitude of the signals in μV.

• 

  Figure 8. Comparison of the scalp ECG with an estimation of it as a linear combination of the abdominal fetal ECG for 2 records (r10 and r08) of the Abdominal and Direct Fetal ECG Database. In the first row (red) the scalp ECG estimated from the noisy fetal ECG is presented, in the second (blue) the ECG as measured by the scalp electrode (running average of 30 complexes) and in the last one (green) the scalp ECG estimated from the denoised fetal ECG by the proposed network. The horizontal axis depicts the samples at 500 Hz, while the vertical the amplitude of the signals in μV.