Related Experiment Video
Updated: Jan 3, 2026
08:51
Magnetic Resonance Imaging of Multiple Sclerosis at 7.0 Tesla
Published on: February 19, 2021
9.6K
QSMGAN: Improved Quantitative Susceptibility Mapping using 3D Generative Adversarial Networks with increased
Yicheng Chen1, Angela Jakary2, Sivakami Avadiappan2
1From the UCSF/UC Berkeley Graduate Program in Bioengineering, University of California, San Francisco and Berkeley, CA, USA; From the Department of Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, USA.
Neuroimage
|November 25, 2019
Summary
Quantitative susceptibility mapping (QSM) uses MRI to measure tissue susceptibility. QSMGAN, a novel deep learning method, significantly improves QSM accuracy, outperforming traditional algorithms for neurological disorders.
Area of Science:
- Medical Imaging
- Neuroscience
- Artificial Intelligence
Background:
- Quantitative susceptibility mapping (QSM) is a key MRI technique for assessing neurological disorders.
- The accuracy of QSM is limited by the ill-posed dipole inversion problem.
Purpose of the Study:
- To develop an accurate and efficient QSM method using deep learning.
- To address the limitations of traditional dipole inversion algorithms.
Main Methods:
- A 3D deep convolutional neural network (QSMGAN) based on a 3D U-Net architecture was developed.
- The network utilized an increased receptive field and a WGAN with gradient penalty for refinement.
- The method was trained to generate QSM maps from single orientation phase maps.
Main Results:
- QSMGAN efficiently generates accurate QSM maps.
- The proposed method significantly outperforms traditional non-learning-based dipole inversion algorithms.
- The algorithm demonstrated generalization capabilities on unseen pathologies, including radiation-induced microbleeds in brain tumor patients.
Conclusions:
- QSMGAN offers a robust and accurate deep learning solution for QSM.
- This approach enhances the reliability of susceptibility mapping in clinical neuroscience.
- The method shows promise for broader applications in neuroimaging and disease quantification.
