An evaluation of image-based and statistical techniques for harmonizing brain volume measurements

Yuan-Chiao Lu, Lianrui Zuo, Yi‐Yu Chou, Blake E. Dewey, Samuel W. Remedios, Russell T. Shinohara, Sonya Steele, Govind Nair, Daniel S. Reich, Jerry L. Prince, Dzung L. PhamImaging 2025 · DOI

TL;DR

This paper evaluates three methods for harmonizing brain MRI volumes to reduce variations from different scanners and protocols. The unsupervised deep learning method HACA3 significantly outperformed statistical (neuroCombat) and supervised deep learning (DeepHarmony) approaches in achieving consistent brain volume measurements and detecting simulated atrophy, making it a promising tool for multi-site neuroimaging studies.

What Problem Does This Paper Address?

Analyzing brain volumes from MRI scans is crucial for understanding aging, disease, and treatment effects. However, variations in MRI scanner hardware, software, and acquisition settings across different sites and over time introduce significant noise and bias. This "batch effect" can lead to inconsistent measurements, hindering the reliability and comparability of results, especially in large, multi-center studies. Retrospective harmonization techniques aim to correct these variations after data acquisition, but their effectiveness varies. This paper addresses the need for a comprehensive evaluation of leading retrospective harmonization methods to identify the most effective approach for ensuring consistent and reliable brain volume measurements.

Methods

The study compares three harmonization techniques:

  • neuroCombat: A statistical batch correction tool that uses linear regression to adjust for "site-specific" effects while preserving biological signals. The study explored its performance with and without training data, and with different empirical Bayes options.
  • DeepHarmony: A supervised deep learning method that uses image-to-image translation to harmonize MR images. It requires paired image data from different acquisition protocols for training.
  • HACA3: An unsupervised deep learning image translation approach. Unlike DeepHarmony, HACA3 can be trained on data from different subjects and doesn't require paired data, offering greater flexibility.

The researchers also employed FreeSurfer for brain segmentation and volume extraction, and conducted a sensitivity analysis using simulated hippocampal atrophy to assess the methods' ability to detect subtle changes.

Datasets

The study utilized a dataset of T1-weighted MRI scans from 39 healthy subjects. Two types of T1-weighted images were acquired for each subject in a single session:

  • GRE (Gradient-Recalled Echo): 3D spoiled gradient-recalled echo images.
  • MPRAGE (Magnetization Prepared Rapid Gradient Echo): 3D magnetization prepared rapid gradient echo images.

For training and testing, the data was split. DeepHarmony used a training cohort of 12 scans and a testing cohort of 27 scans. HACA3 was trained on a diverse set of 21 external imaging datasets (not including the study's GRE and MPRAGE cohorts) and tested on the 27 scans from the study's testing cohort.

Key Insights and Ideas

  • Unsupervised Deep Learning Reigns Supreme: HACA3, an unsupervised deep learning method, consistently outperformed both statistical (neuroCombat) and supervised deep learning (DeepHarmony) methods across most metrics. This suggests that unsupervised approaches, which don't rely on paired data, may be more robust and generalizable for harmonization.
  • The Power of MPRAGE as Reference: For both DeepHarmony and HACA3, using MPRAGE as the reference contrast yielded more consistent volumetric measurements compared to using GRE as a reference. This highlights the importance of selecting an appropriate reference image for image-based harmonization.
  • neuroCombat's Training Data Dependency: The study revealed that neuroCombat's effectiveness, particularly in detecting simulated atrophy, was significantly dependent on the availability of training data. Without training data, it struggled to differentiate between images with and without atrophy, indicating a potential limitation in real-world scenarios where training data might be scarce.
  • Trade-offs in Harmonization: While all methods improved consistency over unharmonized data, HACA3 achieved the lowest measurement variation and highest agreement, suggesting it preserves biological variability better. DeepHarmony also showed improvements but to a lesser extent, while neuroCombat, though effective at removing bias, exhibited higher variability.

Key Results

  • Measurement Consistency: HACA3 demonstrated the lowest absolute volume difference percentage (AVDP) across all brain regions (<3%) and the highest agreement between GRE and MPRAGE measurements (smallest mean difference in CV, narrowest confidence intervals, and ICC >0.9).
  • Atrophy Detection: In simulated atrophy experiments, HACA3 consistently achieved the smallest AVDPs. DeepHarmony showed significant improvements in some regions, while neuroCombat exhibited higher variability.
  • neuroCombat Limitations: Without training data, neuroCombat could not differentiate between images with and without atrophy. However, with training data, it effectively detected hippocampal atrophy.
  • Combined Methods: Combining DeepHarmony with neuroCombat did not significantly improve harmonization. However, combining HACA3 with neuroCombat showed a significant improvement in cortical GM volume difference.

Limitations and Future Work

  • Dataset Size: The study was performed on a relatively small dataset, and the authors acknowledge that larger sample sizes might benefit neuroCombat and DeepHarmony.
  • Geometric Distortions: The study did not explicitly consider geometric distortions, which can affect volumetric analyses, especially in structural imaging. Future work could investigate how image-based harmonization algorithms handle these differences.
  • Contrast Change Magnitude: The GRE and MPRAGE contrast change used in this study was significant. Future research should explore harmonization performance with more subtle contrast differences.
  • Generalizability: While HACA3 performed well on the independent dataset, its generalizability across all unseen imaging protocols and demographic/clinical variabilities needs further investigation. The authors suggest users evaluate compatibility and fine-tune models when necessary.
  • Longitudinal Harmonization: The study focused on cross-sectional harmonization. Further research is needed to optimize harmonization for longitudinal data, especially when temporal changes are minimal.
  • Covariate Incorporation: While the study did not find significant improvements by incorporating biological covariates into neuroCombat, this might be beneficial in studies with less matched subject populations.

Summary updated

Abstract

Volumetric analysis of magnetic resonance brain images is often complicated by variations in scanner hardware, software, and acquisition settings. Over the past several years, there has been an increase in the use of retrospective harmonization techniques for addressing these variations. This research evaluates three image harmonization methods-neuroCombat (a statistical batch correction tool), DeepHarmony (a supervised deep learning method based on image-to-image translation), and HACA3 (an unsupervised deep learning image translation approach). The study focuses on their effectiveness in achieving consistent brain volume measurements across differing T1-weighted acquisitions (GRE and MPRAGE) and their ability to detect simulated atrophy changes in the same acquisitions. While all three methods notably enhance the consistency of regional brain volumes compared with unharmonized images, HACA3 demonstrated the lowest measurement variation in terms of absolute volume difference percentage (AVDP) across all brain regions (<3%). It also demonstrated the highest agreement between the coefficient of variation (CV) measurements of GRE and MPRAGE images, evidenced by the smallest mean difference (0.12) and the narrowest 95% confidence intervals ([-1.04, 1.28]), alongside achieving the highest intra-class correlation (ICC) values across all regions (ICC >0.9). In the atrophy simulation experiments, HACA3 consistently achieved the smallest AVDPs across most unchanged brain regions, while DeepHarmony showed significant improvements in several regions, and neuroCombat exhibited higher variability. Additionally, using neuroCombat with training data effectively detected hippocampal atrophy, whereas without training, neuroCombat could not differentiate between images with and without atrophy, highlighting a potential limitation in its ability to detect subtle brain volume changes when training data are unavailable. In most metrics, HACA3 was found to be the most effective for harmonizing MRI data, followed by DeepHarmony, with neuroCombat showing more measurement variability but still offering improvements over unharmonized data.
BibTeX
@inproceedings{lu2025evaluation,
  author = {Yuan-Chiao Lu and Lianrui Zuo and Yi‐Yu Chou and Blake E. Dewey and Samuel W. Remedios and Russell T. Shinohara and Sonya Steele and Govind Nair and Daniel S. Reich and Jerry L. Prince and Dzung L. Pham},
  title = {{An evaluation of image-based and statistical techniques for harmonizing brain volume measurements}},
  booktitle = {Imaging},
  year = {2025},
  doi = {10.1162/imag.a.73},
}
Figure