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

TL;DR

This survey addresses the pervasive problem of non-biological variability in multi-site MRI data by surveying acquisition-level harmonization and retrospective image- and feature-level methods, including traveling-subject frameworks. It emphasizes that while existing techniques can achieve site invariance, robust preservation of biological signals requires standardized validation benchmarks and integrated assessment across the imaging pipeline. The authors categorize methods into prospective acquisition, retrospective image-level and feature-level approaches, and traveling-subject–based strategies, detailing representative techniques such as vendor-agnostic pulse sequences, CycleGAN-style image translation, MISPEL-style disentanglement, ComBat-based statistical corrections, cVAE extensions, and normalizing-flow models. The review highlights key challenges, including limited cross-modality generalization, potential over-correction or hallucinations, regulatory and deployment barriers, and the need for unified evaluation protocols to enable reliable clinical translation and large-scale mega-analyses.

Abstract

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.

Figures (6)

• Figure 1: Overview of harmonization strategies across the entire MRI pipeline, covering image acquisition, reconstruction, preprocessing, feature extraction and analysis, and including representative methods discussed in this review. dMRI, diffusion-weighted MR imaging; T1WI, T1-weighted imaging; T2WI, T2-weighted imaging; RBV, regional brain volume; CT, cortical thickness; FA, fractional anisotropy; TS, Traveling subject.
• Figure 2: Harmonized acquisition and reconstruction workflow proposed in RN12. The pulse sequence was implemented using Pulseq to ensure identical configurations across scanners and vendors. All post-processing steps including image reconstruction and quantitative parameter fitting were performed offline using a consistent pipeline.
• Figure 3: Four representative categories of image-level deep learning-based harmonization methods: (a) adversarial learning and style transfer, (b) anatomy-contrast disentanglement, (c) multi-contrast prior learning and (d) source-free distribution modeling. A: anatomy; S: style; C: content.
• Figure 4: Schematic illustration of the feature-level deep learning harmonization method DeepResBat RN5. The covariates effect of the original features were first removed by subtraction and used as the input to a VAE. Then, the VAE output was added back to the removed covariate components to obtain harmonized features. ROI: region of interest.
• Figure 5: Two representative deep learning-based harmonization strategies using traveling subject data: (a) end-to-end mapping and (b) anatomy-contrast disentanglement methods. The availability of paired training data provides additional supervision related to site or subject identity, which enhances the learning of site-invariant representations. A: anatomy; S: style.