ULF-Synth: Physics-Guided Ultra-Low-Field MRI Enhancement for Pediatric Neuroimaging

Abstract

Ultra-low-field (ULF) MRI enables portable and accessible neuroimaging, but suffers from low signal-to-noise ratio and limited spatial resolution relative to high-field (HF) systems. Acquiring paired ULF–HF data for supervised enhancement is often infeasible, particularly in resource-limited settings. We introduce ULF-Synth, a framework combining: (i) acquisition-based synthesis of realistic ULF images from HF volumes for large-scale paired training, and (ii) a spatial-frequency domain objective that prioritizes recovery of high-frequency anatomical detail. The formulation is architecture-agnostic, consistently improving structural similarity and perceptual fidelity across encoder-decoder, adversarial, and diffusion-based translation models. Trained exclusively on synthetic data, our models generalize to real 64 mT ULF acquisitions, improving multiclass brain segmentation and achieving higher radiologist preference and diagnostic acceptability in a blinded reader study.

Simulation Pipeline

The ULF synthesis module models the key physical phenomena distinguishing ULF from HF acquisitions:

Qualitative Results

Installation

git clone https://github.com/toufiqmusah/ULF-Synth.git
cd ULF-Synth
pip install -e .

This installs ulfsynth and all core dependencies (including PyTorch). The bundled nnUNet translation fork is automatically discovered and installed during setup — no extra steps needed.

Install from PyPI (future)

pip install ulfsynth          # simulation only
pip install ulfsynth[full]    # with enhancement support

CLI

The ulfsynth package provides three commands:

ulfsynth simulate — ULF synthesis from HF volumes

# Single volume
ulfsynth simulate input.nii.gz output.nii.gz

# Folder of NIfTI files
ulfsynth simulate /path/to/hf/scans/ /path/to/ulf/scans/

# Reproducible seed
ulfsynth simulate input.nii.gz output.nii.gz --seed 42

ulfsynth enhance — ULF→HF restoration (requires nnUNet)

# Single volume (CPU)
ulfsynth enhance --device cpu input.nii.gz output.nii.gz

# Folder of NIfTI files (GPU)
ulfsynth enhance /path/to/ulf/scans/ /path/to/enhanced/scans/

# Weights are downloaded from HuggingFace on first use.
# Pre-download: ulfsynth download-weights

ulfsynth download-weights — cache pretrained weights

ulfsynth download-weights

Caches model weights from HuggingFace to ~/.cache/ulfsynth/.

Python API

Simulation

from ulfsynth.simulate import simulate_ulf, simulate_file, simulate_folder, sample_params

# Generate one ULF volume with random parameters
ulf_volume, affine, header, params = simulate_ulf("hf_input.nii.gz")

# With a fixed seed
ulf_volume, affine, header, params = simulate_ulf("hf_input.nii.gz", seed=42)

# Custom parameters
params = sample_params()
params["signal_target"] = 30
ulf_volume, affine, header, params = simulate_ulf("hf_input.nii.gz", params=params)

# Single-file convenience (returns params dict)
params = simulate_file("hf_input.nii.gz", "ulf_output.nii.gz", seed=42)

# Batch folder processing (returns list of params)
results = simulate_folder("hf_scans/", "ulf_scans/", seed=42)

Output preserves the input affine and header metadata.

Enhancement

from ulfsynth.enhance import enhance_file, enhance_folder

# Single file
enhance_file("ulf_input.nii.gz", "enhanced_output.nii.gz", device="cpu")

# Batch folder processing
enhance_folder("ulf_scans/", "enhanced_scans/", device="cuda")

Requires nnUNet (pip install -e src/nn-translation/). Weights are auto-downloaded on first call.

Roadmap

• Physics-guided ULF synthesis pipeline
• Pre-trained enhancement weights — ULF→HF restoration models
• Python package — pip install ulfsynth
• Docker image — zero-config containerized pipeline

Citation

@misc{musah2026ulfsynth,
  title        = {ULF-Synth: Physics-Guided Ultra-Low-Field MRI Enhancement for Pediatric Neuroimaging},
  author       = {Toufiq Musah and Salvatore Calcagno and Federica Proietto Salanitri and Xiaomeng Li and Maruf Adewole and Marawan Elbatel},
  year         = {2026},
  eprint       = {2605.24625},
  archivePrefix = {arXiv},
  url          = {https://arxiv.org/abs/2605.24625}
}

License

MIT