Related Experiment Video
Updated: Jan 7, 2026
07:13
Author Spotlight: An Efficient and Robust Software for Automated Fusion of Multiple Preclinical Imaging Modalities
Published on: October 27, 2023
1.6K
Evaluation of Image-Level Harmonization Methods for Multi-Center MR Neuroimaging.
Brandon C Ho1, Donghoon Kim1, Ashwin Kumar1
1Department of Radiology, Stanford University, Stanford, California, USA.
Journal of Magnetic Resonance Imaging : JMRI
|January 5, 2026
Summary
Deep learning harmonization (HACA3) improved MRI consistency across vendors more than statistical methods. However, harmonizing T2-FLAIR images remains challenging, indicating limitations in current multi-contrast MRI harmonization tools.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Neuroimaging
Background:
- Multi-center imaging studies generate valuable data for pathology identification and deep learning model training.
- Scanner and site variations can confound analyses, necessitating image harmonization.
- Standardizing MRI data is crucial for reliable large-scale studies.
Purpose of the Study:
- To assess scanner-induced differences in T1w and T2-FLAIR MRI scans within the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
- To evaluate the effectiveness of available image-level harmonization tools for MRI data.
- To compare deep learning-based harmonization with traditional statistical methods.
Main Methods:
- Retrospective analysis of 1143 ADNI3 subjects across different scanner vendors (GE, Philips, Siemens).
- Comparison of gray/white matter contrast ratio (G/W ratio), white matter hyperintensity (WMH) volume, and image similarity metrics (FID, LPIPS) before and after harmonization.
- Application of statistical (ComBat) and deep learning (HACA3) harmonization algorithms.
Main Results:
- Significant baseline differences in G/W ratio and WMH volume were observed between scanner vendors.
- Both ComBat and HACA3 improved G/W ratio consistency, with HACA3 showing superior performance.
- HACA3 achieved the best image similarity across datasets and normalized WMH volume differences.
- Harmonization of T2-FLAIR images, especially from GE scanners, showed improvement but still presented challenges.
Conclusions:
- Deep learning-based HACA3 harmonization significantly outperformed statistical ComBat, enhancing MR contrast consistency and feature similarity across vendors.
- While effective for T1w and some T2-FLAIR data, current harmonization tools face limitations in fully standardizing multi-contrast MRI data.
- Further development is needed for robust harmonization of complex MRI contrasts like T2-FLAIR across diverse scanners.
