Clinical electrocardiogram digitization based on U 2 Net

Abstract

Aim: Deep learning has advanced medical artificial intelligence, but many electrocardiogram (ECG) records remain available only as paper-based documents. Scanned paper ECG images often contain grid lines, background artifacts, and unstructured waveform information, which limits their direct use for quantitative ECG analysis. This study aimed to develop an improved ECG image segmentation method for extracting clean ECG traces from scanned paper-based ECG records and supporting subsequent ECG signal digitization. Methods: We proposed CU2-NeXt, an ECG image segmentation framework that extends U2-Net with ConvNeXt modules and incorporates convolutional block attention module (CBAM) to emphasize waveform-related features while suppressing background artifacts. Viterbi dynamic programming was subsequently applied to reconstruct continuous signal trajectories from the segmented waveform pixels. Results: On the evaluated dataset, CU2-NeXt achieved a macro-F1 score of 88.9% and a Dice coefficient of 78.7%, with higher overlap-based performance than the evaluated baseline segmentation models. Conclusion: These findings demonstrate the feasibility of using CU2-NeXt for ECG trace extraction and digitization from scanned paper ECG images. The proposed framework may provide a useful preprocessing approach for converting paper ECG records into analyzable digital signals and supporting downstream quantitative ECG analysis.