ULF-EnC-Challenge

ULF-EnC: Ultra-Low-Field MRI Image Enhancement Challenge

📌 Project Overview

The ULF-EnC Challenge is designed to advance deep learning techniques for enhancing ultra-low-field (64mT) MRI to match high-field (3T) MRI quality. This challenge provides a paired dataset of 3T and 64mT MRI scans and aims to push the boundaries of medical image enhancement by encouraging participants to develop robust algorithms that improve diagnostic capabilities.

👉 Official Challenge Website on Synapse (https://www.synapse.org/Synapse:syn65485242/wiki/631224)

🏆 Challenge Goals

Develop deep learning models to enhance 64mT MRI images.

Maintain anatomical integrity and clinical relevance.

Submissions are benchmarked using a weighted combination of SSIM, PSNR, MAE, and NMSE, with SSIM contributing the most to the final ranking to emphasize structural fidelity.

Promote accessibility to high-quality MRI in low-resource settings.

📦 ULF-EnC-Challenge

ULF-EnC-Challenge/
├── data/
│   ├── test/
│   │   ├── 3T/         # Ground truth high-field MRI
│   │   ├── 64mT/       # Ultra-low-field MRI
├── submissions/
│   ├── team1/          # Example team submission
│   ├── team2/          # Additional submissions
├── evaluation.py       # Evaluation script for ranking submissions
├── generate_dummy_data.py  # Script for testing with synthetic data
├── leaderboard.csv     # Generated results after evaluation
├── evaluation.log      # Log file with warnings/errors
├── README.md           # Project documentation

🖥️ Installation & Setup

1️⃣ Install Required Dependencies

Ensure you have Python 3.x installed, then run the following commands:

pip install numpy nibabel pandas scikit-image

📥 Clone the Repository

git clone https://github.com/BioMedAnalysis/ULF-EnC-Challenge.git
cd ULF-EnC-Challenge

🧪 Generate Dummy Data

python generate_dummy_data.py

✅ Run Evaluation

python evaluation.py

📊 View Leaderboard

cat leaderboard.csv

📊 Evaluation Process

Submissions are evaluated per subject and per modality.

Two metrics are used:

Structural Similarity Index (SSIM)

Peak Signal-to-Noise Ratio (PSNR)

Mean Absolute Error (MAE)

Normalised Mean Square Error (NMSE)

Leaderboard ranks teams based on a weighted combination of SSIM (70%), PSNR (10%), MAE (10%), and NMSE (10%) is used to evaluate reconstruction quality, emphasizing structural fidelity while incorporating intensity-based and error-based metrics.

Results are stored in leaderboard.csv.

🚀 Submission Guidelines

🔹 File Naming

Each team must submit enhanced MRI images using the following naming format:

✅ Validation Folder Structure

📁 validation/
├── 📁 POCEMR001/
│   └── 📁 Enhanced/
│       ├── POCEMR001_T1.nii.gz
│       ├── POCEMR001_T2.nii.gz
│       └── POCEMR001_FLAIR.nii.gz
├── 📁 POCEMR002/
│   └── 📁 Enhanced/
│       ├── POCEMR002_T1.nii.gz
│       ├── POCEMR002_T2.nii.gz
│       └── POCEMR002_FLAIR.nii.gz
├── 📁 POCEMR003/
│   └── 📁 Enhanced/
│       ├── POCEMR003_T1.nii.gz
│       ├── POCEMR003_T2.nii.gz
│       └── POCEMR003_FLAIR.nii.gz
...

✅ Test Folder Structure

📁 test/
├── 📁 POCEMR001/
│   └── 📁 Enhanced/
│       ├── POCEMR001_T1.nii.gz
│       ├── POCEMR001_T2.nii.gz
│       └── POCEMR001_FLAIR.nii.gz
├── 📁 POCEMR002/
│   └── 📁 Enhanced/
│       ├── POCEMR002_T1.nii.gz
│       ├── POCEMR002_T2.nii.gz
│       └── POCEMR002_FLAIR.nii.gz
├── 📁 POCEMR003/
│   └── 📁 Enhanced/
│       ├── POCEMR003_T1.nii.gz
│       ├── POCEMR003_T2.nii.gz
│       └── POCEMR003_FLAIR.nii.gz
...

🔹 Naming Convention

• Each subject folder is named after the subject ID (e.g., POCEMR001)

• Inside each subject folder, an Enhanced/ subfolder contains multi-contrast 3D NIfTI files

• File naming format: SUBJECTID_CONTRAST.nii.gz Example: POCEMR001_T1.nii.gz, POCEMR001_FLAIR.nii.gz

🔹 Submission Format

• All reconstructed images must be submitted in .nii.gz format.
• Each team must upload a Docker image containing their trained model.
• The submission should produce 3D reconstructions for each test subject.

📜 License

This project is licensed under MIT License.

📬 Contact & Support

For inquiries, please reach out via GitHub Issues or email the organizers.

🎯 Future Work

Extend the dataset with more paired 3T-64mT scans.

Introduce additional deep learning-based evaluation metrics.

Collaborate with medical experts to refine clinical impact.

🚀 Join us in pushing the boundaries of medical imaging!

CMT Acknowledgment

The Microsoft CMT service was used for managing the peer-reviewing process for this conference. This service was provided for free by Microsoft and they bore all expenses, including costs for Azure cloud services as well as for software development and support.