💻 computer science

Digitizing Paper ECGs at Scale: An Open-Source Algorithm for Clinical Research

This paper introduces and validates a fully automated, open-source framework that converts millions of existing paper ECG scans into digital signals, significantly outperforming state-of-the-art methods on diverse datasets to enable large-scale retrospective clinical research and AI-driven diagnostics.

Original authors: Elias Stenhede, Agnar Martin Bjørnstad, Arian Ranjbar

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Elias Stenhede, Agnar Martin Bjørnstad, Arian Ranjbar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive library of old medical records, but instead of being typed into a computer, they are all printed on paper strips and tucked away in filing cabinets. These strips contain ECGs (electrocardiograms), which are like the "heartbeat fingerprints" of patients. Today, super-smart computer programs (AI) can read these fingerprints instantly to spot heart problems, but they can only do so if the data is in a digital format.

Right now, millions of these paper records are stuck in the past, invisible to modern AI. This paper introduces a new, free tool that acts like a high-tech translator, turning those old paper photos into digital signals that computers can understand.

Here is how the paper explains their solution, broken down into simple concepts:

The Problem: The "Lost in Translation" Gap

Think of the old paper ECGs as a handwritten letter. A computer can't read handwriting directly; it needs the letter typed out first.

The Issue: Many hospitals still print ECGs on paper. Even if they scan the paper into a photo, the computer sees a picture, not the actual heartbeat data.
The Consequence: AI tools that could save lives by analyzing these heartbeats are useless because they can't "read" the paper photos.

The Solution: A Five-Step Digital Assembly Line

The authors built a fully automated software pipeline (a step-by-step assembly line) that takes a photo of a paper ECG and turns it into a clean digital signal. They compared this process to fixing a crooked, wrinkled photo and then reading the text inside it.

Here are the five steps their "assembly line" performs:

The "Magic Eye" (Segmentation):
Imagine looking at a messy desk covered in papers, coffee stains, and scribbles. The software first puts on "magic glasses" to separate the important stuff (the heartbeat lines) from the junk (the grid lines, the text, and the background). It isolates just the heartbeat drawing.
Straightening the Photo (Perspective Correction):
If you take a photo of a paper ECG with your phone, it might look tilted, stretched, or warped, like a funhouse mirror reflection. This step uses math to "flatten" the image, making the paper look like it was scanned perfectly flat, even if the photo was taken at a weird angle.
Finding the Ruler (Grid Size Extraction):
Paper ECGs have a grid of tiny squares. The software needs to know exactly how big those squares are to measure the heartbeat correctly. It acts like a detective, counting the pixels between the grid lines to figure out the scale (e.g., "One square equals one millisecond").
Reading the Labels (Layout Identification):
ECGs have different sections for different parts of the heart (like Lead I, Lead II, etc.). The software scans the image to find the little text labels (like "V1" or "aVR") and figures out the map of where each heartbeat line belongs.
The Final Translation (2D to 1D Conversion):
Finally, the software traces the squiggly lines on the paper and converts them into a digital wave (a list of numbers) that computers can analyze. It's like tracing a drawing and turning it into a musical note.

How Well Does It Work? (The Test Drive)

The team tested their tool on two huge collections of data:

The "Real World" Test: They used 1,596 photos taken at a hospital in Norway. Some were taken with a flatbed scanner (perfectly flat), and others were taken with iPhone and OnePlus phones (which can be crooked or have bad lighting).
- Result: The tool worked best on scanner photos (very clear), but it was still very good at fixing phone photos, even if they were slightly blurry or tilted.
The "Hard Mode" Test: They also tested it on a massive, hidden dataset from the Emory University challenge, which included photos of papers that were stained, wrinkled, or even moldy.
- Result: Their tool beat every other method in the competition, even on the messy, damaged papers. It was the first to successfully turn all types of photos into usable data.

Why This Matters (According to the Paper)

The authors state two main reasons why this tool is a big deal:

It Opens the Doors: It allows hospitals that don't have fancy digital equipment to still use modern AI tools. You just take a photo of the paper, run it through this software, and suddenly the AI can read it.
It Unlocks History: It lets researchers dig into decades of old paper records. This creates a huge new library of data to train AI, which is especially helpful for studying rare heart conditions that don't have enough modern data to study.

The "Open Source" Promise

Unlike many medical tools that are locked behind paywalls or secret code, the authors released the entire software for free. They also shared the "training data" (the practice problems the AI learned on) and the real hospital data they used for testing. This means other scientists can look at the code, check the work, and improve it, ensuring the tool keeps getting better.

In short: This paper presents a free, automated "translator" that rescues millions of old, paper-based heart records, turning them into digital data so modern AI can finally read them and help diagnose heart disease.

Technical Summary: Digitizing Paper ECGs at Scale

Problem Statement
Despite the widespread adoption of electronic health records, millions of clinical Electrocardiograms (ECGs) remain archived only as paper scans or photographs. These analog records are inaccessible to modern automated diagnostic tools and deep learning algorithms, which require standardized digital time-series data. While AI has surpassed human performance in ECG analysis, its efficacy is significantly diminished when applied directly to scanned images due to artifacts such as perspective distortion, wrinkles, stains, and variable grid configurations. Existing solutions for digitization have historically been constrained by a lack of full automation, insufficient validation on realistic datasets, or the absence of open-source code.

Methodology
The authors propose a fully automated, modular framework designed to convert scanned or photographed paper ECGs into standardized digital signals. The pipeline consists of five distinct modules:

Semantic Segmentation: A residual U-Net architecture (8 encoder-decoder blocks) is employed to segment the input image into four classes: ECG waveforms, grid lines, text annotations, and background. This step removes lighting variations and background noise.
Perspective Correction and Cropping: To address perspective distortion common in mobile phone photos, the framework utilizes a two-step Hough transform method. It first separates grid lines and then maps them into an angle-radius domain, followed by an angle-angle transformation. This allows the algorithm to identify the orientation of parallel grid lines even under significant distortion, enabling robust dewarping and cropping.
Layout Identification: A lightweight U-Net segments lead text annotations (e.g., "II", "aVR"). The system calculates weighted centroids of these text regions to infer the ECG layout (e.g., 3×4, 6×2, 12×1). It supports arbitrary layouts via a configuration file and uses a minimum-weight matching algorithm (modified Jonker-Volgenant) to align detected leads with standard configurations, handling missing markers by assigning fixed distances.
Grid Size Extraction: An iterative grid search algorithm determines the pixel-to-physical-unit conversion factors (mm/s and mm/mV). It collapses the dewarped feature map to 1D, computes autocorrelation, and matches the result against a template of major and minor grid lines to estimate the grid spacing parameter $d$ .
2D to 1D Conversion: The segmented waveforms are converted into distinct lead traces. Connected components are identified, and a "snipping" algorithm subdivides overlapping signals. These components are then re-merged using a linear sum assignment problem solver to reconstruct continuous leads. The system handles cases where signals are partially missing or noisy by assigning NaN values to specific segments rather than rejecting the entire image.

Key Contributions

Open-Source Framework: The release of a complete, modular, open-source software suite for general-purpose ECG digitization.
Synthetic Training Dataset: A dataset of 8,658 simulated high-resolution ECG images with ground-truth labels for segmentation and signal extraction, generated using real ECGs from the CODE15% dataset and augmented with randomized parameters to ensure robustness against domain shifts.
Real-World Clinical Dataset: A novel dataset of 1,596 paper ECG images (scanned and photographed) collected from Akershus University Hospital, paired with raw time-series data for validation.
Algorithmic Advances: Implementation of robust dewarping, alignment, and signal extraction methods that handle common artifacts like stains, wrinkles, and perspective distortion without requiring user intervention.

Results
The framework was validated on two primary datasets:

Ahus Paper Digitization ECG Database (1,596 images):
- Performance: The algorithm achieved a mean Signal-to-Noise Ratio (SNR) of 19.65 dB on flatbed scanner images. Performance was lower but still significant for mobile photos: 12.19 dB (OnePlus) and 10.47 dB (iPhone).
- Efficiency: Average processing time was 5.88 seconds per image on a GPU (NVIDIA RTX 5090).
- Robustness: The system successfully handled various capture conditions, with only 0.21% of samples requiring NaN assignment due to unrecoverable artifacts.
Emory Paper Digitization ECG Database (35,595 images):
- Used as a hidden test set (zero-shot evaluation) from the 2024 George B. Moody PhysioNet Challenge.
- The proposed algorithm outperformed all previously published methods across all subcategories, including clean scans, black-and-white scans, mobile photos, and images of deteriorated or stained papers.
- Notably, it was the first method to achieve positive SNR scores across all categories, including mobile photos of degraded papers (ranging from 1.05 to 2.03 dB).

Significance and Claims
The authors claim that this work addresses the critical bottleneck preventing the utilization of historical paper ECG archives in modern AI-driven diagnostics. By providing a fully automated, open-source solution, the framework aims to:

Democratize Access: Enable advanced diagnostic capabilities for institutions relying on paper records, regardless of their digital infrastructure.
Expand Data Availability: Substantially increase the volume of available training data for AI models, particularly for studying rare cardiovascular conditions where historical data is abundant but currently unusable.
Ensure Reproducibility: The release of code, synthetic training data, and a real-world clinical dataset allows for independent verification and further development.

The paper maintains a modest tone regarding clinical utility, acknowledging that while quantitative metrics (SNR, correlation, RMSE) are strong, further validation is required to confirm clinical efficacy. The authors also note that domain shifts and specific artifact types (e.g., severe curvature in deteriorated papers) remain areas for potential improvement, suggesting future work could explore end-to-end approaches or the utilization of linear dependencies between limb leads.