How Robust Are ECG Image Digitizers to Image Damage? A Per-Type, Per-Method Study

Abstract

ECG image digitization, the task of recovering a digital signal from a photo or scan of a paper electrocardiogram, is the entry point for using the vast archive of paper ECGs held in hospitals, especially in low-resource settings. Recent digitizers report strong accuracy, but almost all of them summarize performance with a single overall signal-to-noise ratio (SNR), a blind spot that lets a healthy-looking average hide total failure on one kind of image damage. We address this by benchmarking a ladder of methods per damage type and per method on one fixed evaluation split of the PhysioNet 2024 dataset, scoring each of the nine damage types on its own with SNR in decibels (dB). A classical top-down sweep baseline scores -4.83 dB overall and is negative on every one of the nine damage types, meaning its output carries more error than signal everywhere. Adding neural grid rectification lifts each type but stays negative overall at -3.37 dB. A full neural pipeline that rectifies the page and then reads the signal reaches about 25.6 dB overall and, crucially, holds between roughly 25.2 and 25.7 dB across all types, a spread of only about 0.46 dB. The per-type view, not the single average, is what reveals that the classical baseline fails uniformly while the neural pipeline is uniformly robust, turning an aggregate number into practical guidance for clinics.