Skip to content
Conference

SICL-ECGDig: a style-invariant contrastive learning framework for robust ECG image digitization

Nov 2026 · Third International Conference on Image Processing, Intelligent Control, and Computer Engineering (IPICE 2026) · 0 citations

Abstract

Electrocardiogram (ECG) signals are widely used in clinical practice for cardiovascular conditions. However, a large number of ECG records are still stored as paper printouts or scanned images, which limits their reuse in longitudinal analysis, clinical decision support, and machine learning. ECG image digitization addresses this need by recovering digital signals from ECG images. However, existing ECG image digitization methods remain sensitive to real-world style variations, such as wrinkles, artifacts, print degradation, and handwritten marks. To address this problem, we propose SICL-ECGDig, a Style-Invariant Contrastive Learning framework for robust ECG image digitization. Our method is motivated by a medical prior: changes in image style should not alter the underlying semantic content of the ECG signal. Specifically, we design a train-only contrastive learning module named SICL (Style-Invariant Contrastive Learning). During training, images generated from the same ECG signal but with different styles are treated as positive pairs, while images generated from different ECG signals are treated as negative pairs. This design encourages the model to learn style-invariant ECG representations while preserving signal-specific semantics. Since the SICL module is only used during training, it does not increase the inference cost of the baseline model. Extensive experimental results on the PTB-XL dataset and clinical data demonstrate the effectiveness of the SICL-ECGDig framework. The SICL-ECGDig framework improves reconstruction accuracy and cross-style consistency over the baseline. It reduces RMSE from 0.065 to 0.059, increases PCC from 0.918 to 0.933. These results show that style-invariant contrastive learning improves the robustness of ECG image digitization under realistic visual degradation.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.