Related Experiment Video
Updated: Sep 7, 2025
09:30
Quantitative Magnetic Resonance Imaging of Skeletal Muscle Disease
Published on: December 18, 2016
19.7K
Deep learning-based quantitative susceptibility mapping (QSM) in the presence of fat using synthetically generated
Jannis Hanspach1, Steffen Bollmann2, Johanna Grigo3,4
1Institute of Radiology, University Hospital Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany.
Magnetic Resonance in Medicine
|June 17, 2022
Summary
A novel deep learning method rapidly reconstructs high-quality quantitative susceptibility maps (QSM) from MRI data. This automated approach, trained on synthetic data, effectively handles diverse chemical shifts without background masking.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Biophysics
Background:
- Quantitative Susceptibility Mapping (QSM) is crucial for analyzing tissue composition.
- Existing QSM methods struggle with diverse chemical shifts and require manual background field removal.
- Deep learning offers potential for faster and more automated QSM reconstruction.
Purpose of the Study:
- To develop a fast, automatic deep learning-based QSM reconstruction method.
- To enable QSM in tissues with diverse chemical shifts, particularly outside the brain.
- To improve QSM accuracy and reduce artifacts compared to conventional methods.
Main Methods:
- A UNET deep learning model was trained using synthetically generated phase data.
- The model reconstructed susceptibility maps from unwrapped, multi-echo phase data.
- Performance was evaluated using RMS error on synthetic data and tested on in vivo knee, pelvis, and prostate datasets.
Main Results:
- The UNET approach achieved a lower RMS error (0.123 ppm) compared to conventional methods (0.292 ppm).
- Reconstruction time was significantly reduced to under 10 seconds for UNET, versus minutes for conventional methods.
- UNET largely reduced background artifacts, fat-water swaps, and seed-related artifacts in clinical data.
Conclusions:
- Deep learning-based QSM, trained with synthetic data, enables rapid, high-quality reconstruction.
- The method is effective for tissues with fat and eliminates the need for background field removal masking.
- This automated approach shows promise for routine clinical applications.
