LeadSeg-ECG: Web-Based 12-Lead ECG Lead Segmentation Using nnU-Net

Title

LeadSeg-ECG: Automatic Lead-Wise Segmentation of 12-Lead ECG Images Using Synthetic Data and nnU-Net

Description

LeadSeg-ECG is a Flask-based research platform for 12-lead electrocardiogram (ECG) image segmentation. The tool allows users to upload ECG images through a browser interface, run semantic segmentation using an nnU-Net v2 model, and visualize the resulting lead masks through multiple outputs, including the original ECG image, raw segmentation mask, colored lead segmentation, segmentation overlay, Grad-CAM style heatmap, and SHAP style attribution map.

The platform is designed for research workflows involving ECG image analysis, model inspection, and visual quality assessment of lead-level segmentation results. It includes a web interface, model inference pipeline, runtime nnU-Net folder preparation, result export utilities, and institution-branded HTML templates. The tool name displayed in the interface is LeadSeg-ECG.

The implemented inference workflow detects the available nnU-Net checkpoint from the project model/ directory, prepares the required nnU-Net runtime structure, converts uploaded ECG images to grayscale before inference, runs prediction on CPU by default, and stores generated visual outputs in the results/ directory. The segmentation display treats class 0 as background and hides it from colored visualizations so that only foreground ECG lead regions are shown. The Grad-CAM style and SHAP style explanation outputs are displayed on white backgrounds with standard color ranges and color bars for interpretation.

This release contains the source code, web templates, static assets, inference utilities, dependency specification, model directory structure, and project documentation needed to run the LeadSeg-ECG platform locally.

Main Features

• Browser-based upload and segmentation of 12-lead ECG images.
• nnU-Net v2 based semantic segmentation workflow.
• Automatic checkpoint detection from the local model/ directory.
• Grayscale conversion before passing images to nnU-Net.
• Colored segmentation visualization with white background.
• Raw mask and overlay visualization.
• Grad-CAM style heatmap visualization with color bar.
• SHAP style attribution visualization with color bar.
• Class 0 background suppression in foreground visualizations.
• Institution-branded interface with M3B Lab and Politecnico di Torino links.
• Local Flask deployment with a Python virtual environment.

Project Contents

The project includes the following main directories and files:

• app.py: Flask application entry point and web routes.
• inference.py: Image preprocessing, nnU-Net inference, fallback demo logic, and visualization generation.
• requirements.txt: Python dependencies required by the project.
• README.md: Complete installation, usage, and troubleshooting documentation.
• templates/: HTML templates for the upload and results pages.
• static/: CSS files and institutional icon assets.
• model/: nnU-Net dataset metadata, plans, and checkpoint files.
• uploads/: Runtime directory for uploaded ECG images.
• results/: Runtime directory for generated segmentation and explanation outputs.
• nnunet_runtime/: Runtime nnU-Net folder structure generated during inference.

Software Environment

The project is intended to run in a Python virtual environment. On Windows, a short-path virtual environment is recommended to avoid long path issues during installation of large packages such as PyTorch:

The tested environment includes Flask, Pillow, NumPy, PyTorch CPU, SHAP, and nnU-Net v2. Dependencies are listed in requirements.txt.

Running the Platform

After installing dependencies in the virtual environment, the web application can be launched with:

The local interface is then available at:

DOI

DOI: 10.5281/zenodo.20082175

Suggested Citation

If you use this project, please cite:

Creators and Contact

For questions, installation issues, model-related problems, or reuse requests, please contact:

Affiliation and Acknowledgement

This work is developed under M3B Lab, Department of Mechanical and Aerospace Engineering, Politecnico di Torino, Italy.

Keywords

• ECG segmentation
• 12-lead ECG
• Lead segmentation
• nnU-Net
• nnU-Net v2
• Flask
• Biomedical image analysis
• Semantic segmentation
• Explainable AI
• Grad-CAM
• SHAP
• M3B Lab
• Politecnico di Torino

Recommended Zenodo Metadata

Upload type: Software

Publication date: Use the date of the Zenodo release.

Version: Use the project release version assigned by the depositor.

Language: English

Access right: Open Access, if permitted by the associated model and data constraints.

License: Select the intended software license before public redistribution. If no license has been selected yet, the repository should not be assumed to be open-source beyond the permissions explicitly granted by the depositor.

Related identifier:

Relation: Is cited by / is supplement to / is new version of, depending on the final Zenodo deposit configuration.

Notes

This software is intended for research and development use. It is not presented as a certified medical device or clinical decision-support system. Users are responsible for validating the model, data, and deployment environment before applying the tool in any scientific, educational, or operational workflow.