Published December 2, 2025 | Version v3.0
Model Open

An nnUNet for fat-water mapping from Dixon-MRI magnitude images

  • 1. ROR icon University of Sheffield

Description

🚀 Summary

This repository contains a trained deep-learning nnUNet model that is used to generate fat and water images from magnitude-only in-phase and opposed-phase Dixon MRI data. 

📚 Background

Computation of fat and water images from a 2-point MRI Dixon acquisition is usually done in-line by the scanner software, and requires access to the phase and magnitude data.

In some cases one may want to compute fat and water images retrospectively - for instance when they were not originally exported, or in order to reconstruct them with different models (e.g. with correction for T2* decay, B0-effects, etc). This causes a practical problem when, as is common, phase images or not stored and only magnitude images of in-phase and opposed-phase scans are available.

The crucial bit of information that is missing with magnitude-only data is the sign of the opposed phase image - does the pixel contain mostly water or mostly fat? This pipeline trains a deep learning model to recover this binary information from magnitude images of in-phase and opposed-phase data.

🛠️ Methods

Training data are taken from 1,143 3D Dixon scans of the abdomen acquired in the iBEAt study on diabetic kidney disease. The dataset includes pre- and post contrast agent images from patients, and precontrast scans in volunteers.

Data were acquired at 3T with a field of view of 400 mm in both the read and phase directions, a slice thickness of 1.5 mm, and 144 slices per slab. The repetition time (TR) was 4.01 ms, with echo times (TE) at 1.34 ms and at 2.57 ms. 

The trained model is a 3D nnUNet which predicts a binary image with value=1 in pixels that contain mostly water, and 0 otherwise. The model has two input channels for in-phase and opposed-phase magnitude images.

The complete training pipeline to derive this version can be found here

📦 Usage

The most convenient way of using the trained model is via the function fatwater in the python package miblab-dl.

📜 Version History

Version Date Description
v3.0 Dec 2, 2025 Uploaded new and improved model weights trained on a larger and cleaner dataset.
v2.0 May 28, 2025 Added an example .gif file: water_dominant_map_nnunet.gif.
v1.0 May 22, 2025 Initial release. Two-channel nnU-Net (.zip file with .pth model and .json file).

Files

FatWaterPredictor.zip

Files (1.1 GB)

Name Size Download all
md5:d7f01060de4e8caab35f02107550e149
1.1 GB Preview Download

Additional details

Funding

European Commission
BEAt-DKD - Biomarker Enterprise to Attack DKD - Sofia ref.: 115974 115974

Dates

Submitted
2025-12-02