Harmonization in Magnetic Resonance Imaging: A Survey of Acquisition, Image-level, and Feature-level Methods

Magnetic resonance imaging (MRI) has greatly advanced neuroscience research and clinical diagnostics. However, imaging data collected across different scanners, acquisition protocols, or imaging sites often exhibit substantial heterogeneity, known as batch effects or site effects. These non-biological sources of variability can obscure true biological signals, reduce reproducibility and statistical power, and severely impair the generalizability of learning-based models across datasets. Image harmonization is grounded in the central hypothesis that site-related biases can be eliminated or mitigated while preserving meaningful biological information, thereby improving data comparability and consistency. This review provides a comprehensive overview of key concepts, methodological advances, publicly available datasets, and evaluation metrics in the field of MRI harmonization. We systematically cover the full imaging pipeline and categorize harmonization approaches into prospective acquisition and reconstruction, retrospective image-level and feature-level methods, and traveling-subject-based techniques. By synthesizing existing methods and evidence, we revisit the central hypothesis of image harmonization and show that, although site invariance can be achieved with current techniques, further evaluation is required to verify the preservation of biological information. To this end, we summarize the remaining challenges and highlight key directions for future research, including the need for standardized validation benchmarks, improved evaluation strategies, and tighter integration of harmonization methods across the imaging pipeline.

💡 Research Summary

This review provides a comprehensive overview of MRI harmonization methods aimed at mitigating non‑biological variability—commonly referred to as batch or site effects—that arises when data are collected across different scanners, protocols, or sites. The authors organize the field into three overarching categories: prospective (acquisition‑level) harmonization, retrospective image‑level harmonization, and retrospective feature‑level harmonization, and they also discuss traveling‑subject designs that serve as a prospective validation tool.

In the prospective domain, the paper highlights vendor‑agnostic pulse‑sequence frameworks such as Pulseq, gammaSTAR, and R‑THawk, which allow researchers to define RF pulses and gradient waveforms in a hardware‑independent format. Empirical evidence is presented showing that Pulseq‑based diffusion sequences dramatically reduce inter‑scanner variability in fractional anisotropy (FA) measurements (35‑50 % reduction in standard error) compared with vendor‑provided sequences. The authors also discuss harmonized image‑reconstruction techniques—including multi‑coil combination, distortion correction, and standardized reconstruction pipelines—that address variability at the raw‑data stage, particularly for echo‑planar imaging used in diffusion and functional MRI. Traveling‑subject studies, where the same individuals are scanned at multiple sites, are described as a powerful means of directly quantifying and correcting site‑specific biases.

Retrospective harmonization dominates current practice because large public multi‑site datasets are readily available. Image‑level approaches are divided into statistical methods (e.g., ComBat, CovBat) and deep‑learning image‑to‑image translation models (GANs, CycleGANs, VAEs). While these techniques can align intensity distributions, contrast, and signal‑to‑noise ratios across sites, the authors caution that generative models may inadvertently alter anatomical structures or introduce artifacts, underscoring the need for rigorous biological fidelity checks.

Feature‑level harmonization operates on derived quantitative measures such as regional brain volumes, cortical thickness, diffusion tensor metrics, functional connectivity matrices, and radiomic features. Linear mixed‑effects models, Bayesian frameworks, and deep‑learning regressors are surveyed. This category benefits from the ability to incorporate biological covariates (age, sex, diagnosis) directly into the correction model, reducing the risk of anatomical distortion. However, its performance is tightly coupled to the upstream feature‑extraction pipeline, limiting flexibility and reusability.

The review catalogs major public datasets (ABIDE, ADNI, HCP, UK Biobank, etc.) and outlines common evaluation metrics: mean squared error, intraclass correlation coefficient, variance ratios, and downstream task performance (classification accuracy, prediction error). A critical gap identified is the lack of standardized metrics that explicitly quantify preservation of true biological signal after harmonization; most studies rely on indirect statistical tests or visual inspection.

Key insights include: (1) current methods can achieve a degree of site invariance but may over‑correct, attenuating subtle disease‑related effects; (2) deep‑learning image‑level methods are powerful yet demand large, diverse training data and robust generalization strategies; (3) feature‑level methods are more interpretable but inherit errors from the feature extraction stage. The authors argue for an integrated harmonization pipeline that spans acquisition, reconstruction, preprocessing, and analysis, coupled with community‑wide benchmark datasets and multi‑dimensional evaluation frameworks.

Future directions emphasized are: development of unified, end‑to‑end harmonization frameworks; creation of standardized, multimodal benchmark suites that include ground‑truth biological labels; incorporation of explainable AI techniques to verify that biological information is retained; and exploration of privacy‑preserving federated learning approaches that enable harmonization model training without sharing raw images.

In sum, the paper synthesizes the state‑of‑the‑art in MRI harmonization, acknowledges substantial progress in reducing site‑related variability, and calls for systematic validation standards and pipeline‑wide integration to ensure that harmonized data remain biologically meaningful and clinically useful.