Detecting COVID-19 from digitized ECG printouts using 1D convolutional neural networks

Figures

• 

  Fig 1. Schematic diagram of COVID-19 identification system. The process consists of 2 steps: converting ECG images to signals and classifying the resulting ECG signals.

• 

  Fig 2. Procedure for converting paper-based ECG scans to 1D signals. The process of converting image to signal consists of 5 steps: image binarization, isoelectric line detection, lead separation, noise removal and signal extraction.

• 

  Fig 3. Vertical and horizontal histograms of the ECG image. (b) and (c) are the vertical and horizontal histograms of image (a), respectively. (b) has peaks corresponding to positions of the horizontal lines at the top of the paper and the isoelectric lines of the ECG signal. (c) has three peaks corresponding to lines separating the leads. In (d), the black line depicts the positions of all isoelectric lines.

• 

  Fig 4. Image processing for ECG signal extraction. (a) is a binary image of the ECG signal, (b) is an image of the lead V2 signal cropped from (a), (c) is the image of signal after removing noise and artifacts. In (c), there are small noise spots around the signal. These spots were filtered out in (d), leaving only the ECG signal image.

• 

  Fig 5. ECG pixel tracking algorithm. (a) Signal image with noise and artifacts; (b) ECG pixel tracking algorithm with N = 6; (c) Image of signal without noise and artifacts.

• 

  Fig 6. Four types of 12-lead ECG reports in the COVID-19 ECG images dataset.

• 

  Fig 7. (a) Image of an ECG signal in lead II after preprocessing and (b) the resulting signal after image-signal conversion.

• 

  Fig 8. (a, b) Grad-CAM visualization of ECG image with COVID-19. (c, d) Grad-CAM visualization of ECG image with MI. Signal segments significant to the classifier are depicted in darker color.

• 

  Fig 9. t-SNE plots of ECG data of different classes separated at the models’ last convolutional layer.