Projects/05
YOUR-Lead
Recovering digital 12-lead ECG signals from paper images
We developed YOUR-Lead (YOLO and U-Net for Reconstruction of ECG Lead Signals) for the PhysioNet Challenge 2024. The pipeline converts paper ECG images into digital waveforms, combining synthetic training data with rotation correction, trace segmentation, lead detection, and signal reconstruction in physical units.
Why digitize paper ECGs
Digitizing paper ECGs makes archived recordings available for computational analysis. YOUR-Lead reconstructs digital 12-lead waveforms from paper images using image segmentation, lead detection, and signal processing.
From paper image to signal
- Generate training images. We used ECG-Image-Kit to create synthetic paper ECGs from 21,799 PTB-XL recordings. These included variations in layout, grid patterns, and simulated scanning artefacts such as wrinkles, blur, and noise. Clean trace images and lead bounding boxes provided the training targets.
- Isolate and locate the leads. Fourier analysis corrected image rotation. A U-Net isolated the ECG traces from the background, while a fine-tuned YOLOv8n detected each lead region. Geometric post-processing organized the lead regions into common 12-lead layouts and refined their positions; k-means clustering assigned lead identities.
- Reconstruct the waveforms. OpenCV connected-component analysis extracted the trace pixels within each lead box. Grid spacing was estimated through autocorrelation, then used to convert pixel coordinates into amplitude and time using the paper calibration of 10 mm/mV and 25 mm/s.
Challenge evaluation
The system was evaluated in the George B. Moody PhysioNet Challenge 2024 digitization task. It achieved a signal-to-noise ratio (SNR) of 0.447 dB on the hidden validation set, measuring reconstruction quality against reference signals.