Related Experiment Video
Updated: Sep 16, 2025
07:13
Author Spotlight: An Efficient and Robust Software for Automated Fusion of Multiple Preclinical Imaging Modalities
Published on: October 27, 2023
1.4K
A preliminary attempt to harmonize using physics-constrained deep neural networks for multisite and multiscanner MRI
Gawon Lee1, Dong Hye Ye2, Se-Hong Oh3
1Department of Biomedical Engineering, Hankuk University of Foreign Studies, Yongin, the Republic of Korea.
Neuroimage
|July 6, 2025
Summary
We developed Physics-Constrained Deep Neural Network for multisite and multiscanner Harmonization (PhyCHarm) to improve magnetic resonance imaging (MRI) scan consistency. PhyCHarm effectively harmonizes images across different scanners and sites, enhancing data reliability for research.
Area of Science:
- Medical Imaging
- Neuroimaging
- Machine Learning
Background:
- Magnetic Resonance Imaging (MRI) data exhibit variability due to differing scan parameters and scanner specifications.
- Image harmonization is essential to minimize these discrepancies and ensure data consistency across multicenter studies.
- Existing harmonization methods often lack robustness and may not fully account for underlying physics.
Purpose of the Study:
- To develop and evaluate a novel MR physics-based harmonization framework named Physics-Constrained Deep Neural Network for multisite and multiscanner Harmonization (PhyCHarm).
- To assess the performance of PhyCHarm in generating quantitative T1- and M0-maps and harmonizing T1-weighted images.
- To compare PhyCHarm against other deep learning-based harmonization techniques.
Main Methods:
- PhyCHarm utilizes two deep neural networks: a Quantitative Maps Generator for T1- and M0-maps and a Harmonization Network.
- The Quantitative Maps Generator was trained on a 3T MP2RAGE dataset (n=50).
- The Harmonization Network was trained on a traveling 3T T1w dataset (n=9) and evaluated using SSIM, PSNR, and volumetric analysis.
Main Results:
- PhyCHarm demonstrated superior performance in generating quantitative maps, evidenced by high SSIM, PSNR, and NRMSE.
- The Harmonization Network achieved higher SSIM and PSNR compared to U-Net, Pix2Pix, CALAMITI, and HarmonizingFlows.
- PhyCHarm resulted in a greater reduction in gray and white matter volume differences post-harmonization than the compared methods.
Conclusions:
- PhyCHarm represents a significant advancement in MRI harmonization by integrating physics-based constraints into a deep learning framework.
- The framework shows promise for improving data consistency in multisite and multiscanner neuroimaging studies.
- The integration of physics constraints, even in a supervised setting, paves the way for more sophisticated unsupervised harmonization techniques.
