A Multi-Stage Framework for Simultaneous Digitization and Classification of Electrocardiogram Images

Aim: This paper addresses the challenge of simultaneous digitization and arrhythmia classification of electrocardiograms (ECGs) captured from images, as presented by the George B. Moody PhysioNet Challenge 2024.Methods: We propose a multi-stage framework to digitize and classify ECG images concurrently.The framework consists of 3 modules: (1) Region of Interest (ROI) extraction, (2) Arrhythmia classification, and (3) ECG waveform digitization.The ROI extraction module uses a DETR model with a ResNet-50 backbone to locate the ECG region.The classification module uses a lightweight Con-vNeXt model to classify the ECG signal.The digitization module uses a U-Net model to extract the ECG waveform from the ROI.The digitization task involves a highly unbalanced segmentation problem, where the ECG waveforms occupy a small portion of the image.We propose a novel loss function to address this issue.Results: Our team ("Revenger") received a score of 0.33 for the classification problem on the hidden data, ranked 9/16; and a signal-to-noise ratio (SNR) of -0.733 for the digitization problem, ranked 13/16.Conclusion: The proposed framework demonstrates the feasibility of digitizing and classifying ECG images concurrently, but needs further improvement to achieve better performance.