Fusion of Deep Learning and Rule-Based Techniques for Enhanced Paper-Based ECG Digitization

The electrocardiogram (ECG) is crucial for diagnosing cardiovascular diseases.As a part of the Digitization and Classification of ECG Images: The George B. Moody Phy-sioNet Challenge 2024, our team, PapyrusECG, developed a pipeline combining deep learning and rule-based techniques to digitize paper-based ECGs.The pipeline includes image synthesis, skew correction, lead detection, lead binarization, and signal extraction.We generated images with augmentations using the ECG Image Kit, reserving 6000 for training and 1500 to serve as a held-out subset of the training set.Skew correction was implemented using multiple rule-based computer vision methods.Lead detection employed a fine-tuned YOLOv7 model while Lead binarization used a U-Net model.Signal extraction involved a rule-based algorithm that extracted pixel locations and converted them to millivolts, with challenge signal-to-noise ratio (SNR) results of 2.9797.212and 3.0997.107for our training and held-out subset datasets, respectively.Our team achieved a score of -5.192 on the hidden test set (ranked 15th out of 16 teams).Digitizing paper ECGs can enhance AI-based diagnostic systems and advance cardiac disease diagnosis.Future work will optimize the pipeline, address limitations, and expand its application to diverse ECG datasets to improve cardiovascular healthcare.