Automated ECG Image Digitization for Multi-Lead Time Series Reconstruction Using Deep Learning

The electrocardiogram (ECG) is essential for diagnosing cardiovascular disorders. Despite growing digital medical records, large volumes of past ECG data remain as printed, scanned, or photographed images, limiting their usability in automated medical systems. This research presents an automated approach for ECG image digitization.

The model was developed using the PhysioNet ECG Image Digitization challenge dataset. The methodology involved image processing, artifact handling, waveform extraction, and reconstruction of multi-channel ECG signals from images. Image enhancement techniques addressed distortions caused by scanning, rotation, contrast variation, and noise, and reconstructed signals were converted into time-series representations.

Model performance was evaluated using a modified Signal-to-Noise Ratio (SNR) metric designed for ECG reconstruction assessment, accounting for time alignment and amplitude normalization between reconstructed and reference signals. The proposed method achieved a private leaderboard score of 0.10574 in the PhysioNet ECG Image Digitization Kaggle competition, which had 1,424 participating teams. While top-ranked solutions typically used hybrid pipelines combining deep learning-based segmentation with signal-processing refinement to achieve higher scores, our results show that a lighter, image-processing-driven pipeline can still recover clinically relevant waveforms under moderate noise and distortion. Performance declined for heavily overlapping leads, severe skew, and low-contrast scans, indicating key areas for improvement.

These findings demonstrate the feasibility of automated ECG digitization for converting archival ECG records into digitally readable signals, supporting the development of AI-based clinical decision tools and cardiovascular research. Future work will explore deep learning-based waveform segmentation to narrow the performance gap with top leaderboard solutions and improve robustness against severe scan artifacts.