An evaluation of image-based and statistical techniques for harmonizing brain volume measurements
Yuan-Chiao Lu1,2,3,4, Lianrui Zuo5, Yi-Yu Chou1,3
1The Henry M Jackson Foundation for the Advancement of Military Medicine, Inc., Bethesda, MD, United States.
Imaging Neuroscience (Cambridge, Mass.)
|August 13, 2025
Summary
This study compared three MRI harmonization methods for brain imaging. HACA3, a deep learning approach, showed the best performance in ensuring consistent brain volume measurements and detecting atrophy across different MRI scans.
Area of Science:
- Neuroimaging
- Medical image analysis
- Computational neuroscience
Background:
- Magnetic resonance imaging (MRI) brain image analysis is challenged by scanner and acquisition variability.
- Retrospective harmonization techniques are increasingly used to address these inconsistencies.
- Standardized brain volume measurements are crucial for accurate clinical interpretation and research.
Purpose of the Study:
- To evaluate and compare the effectiveness of three retrospective MRI harmonization methods: neuroCombat, DeepHarmony, and HACA3.
- To assess their performance in achieving consistent brain volume measurements across different T1-weighted MRI acquisitions (GRE and MPRAGE).
- To determine their ability to detect simulated brain atrophy changes.
Main Methods:
- Comparison of neuroCombat (statistical), DeepHarmony (supervised deep learning), and HACA3 (unsupervised deep learning) for MRI harmonization.
- Evaluation of measurement consistency using absolute volume difference percentage (AVDP) and coefficient of variation (CV).
- Assessment of agreement using mean difference and intra-class correlation (ICC) between GRE and MPRAGE images.
- Testing atrophy detection capabilities using simulated changes.
Main Results:
- All methods improved regional brain volume consistency compared to unharmonized data.
- HACA3 demonstrated the lowest measurement variation (AVDP <3%) and highest agreement (mean difference 0.12, ICC >0.9) between GRE and MPRAGE.
- HACA3 showed superior performance in detecting simulated atrophy, while neuroCombat's performance was dependent on training data availability.
- DeepHarmony offered significant improvements in several regions, outperforming neuroCombat in atrophy detection.
Conclusions:
- HACA3 is the most effective method for harmonizing MRI data, offering superior consistency and agreement.
- DeepHarmony is a viable alternative, showing significant improvements over unharmonized data.
- NeuroCombat provides improvements but exhibits higher variability and limitations in detecting subtle atrophy without training data.
