ECGFusion: Adaptive Multi-Model Framework for Noise-Robust Digitization and Fine-Grained Classification of Paper ECGs

This study proposes a multi-model fusion approach for digitization and classification of paper electrocardiograms (ECGs).We employ You Only Look Once version 10 (YOLOv10) for ECG signal area detection, DPLinkNet34 for image binarization, and GoogLeNet for multi-label classification.Our method excels in processing ECG images of varying quality, particularly those with noise and distortions.In the official phase of the TimeBeater team, on the hidden test data, we obtained a signal-to-noise ratio (SNR) of -0.056 for the digitization task, ranking 8th out of 16 teams, and a macro F-measure of 0.183 for the classification task, ranking 13th out of 16 teams.These scores are consistent with the published scores on the Challenge website.Additionally, on 5-fold cross-validation of the training set, we achieved an SNR of 7.273 and a macro Fmeasure of 0.833, though these cross-validation results are not comparable to the official challenge scores.This study is part of the "Digitization and Classification of ECG Images: The George B. Moody PhysioNet Challenge 2024".