Signal extraction of paper-based ECG using real-world and augmented ECGs

Electrocardiogram (ECG) today largely remains paper-based making them inaccessible for research and innovation. To address this unmet need, we describe an end-to-end framework for ECG signal extraction from images using a four-stage pipeline and lay down the associated data generation requirements. The signal extraction stages include perspective correction, reorientation of paper, lead isolation and signal segmentation that converts signals to voltage units enabling precise digital reconstruction of the original waveforms. Our data generation methodology generates 72,080 simulated ECG images from open source digital signal datasets. We apply novel, realistic augmentations such as coffee stains, imaging shadows and introduce 3D deformation using Blender3D. Furthermore, we employ text-to-image and image-to-image workflows by synthesizing backgrounds around ECG images using the SDXL model with IP Adapter. Our pipeline achieves acceptable signal extraction with a Pearson correlation coefficient of 0.78 and processing speeds of ≈ 1.03 seconds, which is four times faster than existing methods on a related benchmark dataset. This advancement enables practical use of ECG digitization in fast-paced clinical environments, reducing cardiology consultation turnaround times and facilitating digital storage for population health analysis.