Scanner-Agnostic MRI Harmonization via SSIM-Guided Disentanglement

TL;DR

The paper tackles cross-site MRI variability that hinders multicenter analyses. It presents a 3D image-based harmonization framework that disentangles anatomical content from scanner-site style using encoders and a generator, guided by a differentiable SSIM-based loss to preserve structure while aligning appearance. Key contributions include the SSIM loss with luminance, contrast, and structure components, a volumetric extension of a domain-translation network, and extensive validation across five public datasets showing improved anatomical fidelity (SSIM ~0.97), appearance consistency (Wasserstein distances and luminance SSIM ~0.98–0.99), and downstream gains in brain-age prediction and AD classification. The method enables robust cross-site harmonization, enhancing reproducibility and clinical utility in large-scale neuroimaging studies, with potential extensions to other modalities and covariate integration.

Abstract

The variability introduced by differences in MRI scanner models, acquisition protocols, and imaging sites hinders consistent analysis and generalizability across multicenter studies. We present a novel image-based harmonization framework for 3D T1-weighted brain MRI, which disentangles anatomical content from scanner- and site-specific variations. The model incorporates a differentiable loss based on the Structural Similarity Index (SSIM) to preserve biologically meaningful features while reducing inter-site variability. This loss enables separate evaluation of image luminance, contrast, and structural components. Training and validation were performed on multiple publicly available datasets spanning diverse scanners and sites, with testing on both healthy and clinical populations. Harmonization using multiple style targets, including style-agnostic references, produced consistent and high-quality outputs. Visual comparisons, voxel intensity distributions, and SSIM-based metrics demonstrated that harmonized images achieved strong alignment across acquisition settings while maintaining anatomical fidelity. Following harmonization, structural SSIM reached 0.97, luminance SSIM ranged from 0.98 to 0.99, and Wasserstein distances between mean voxel intensity distributions decreased substantially. Downstream tasks showed substantial improvements: mean absolute error for brain age prediction decreased from 5.36 to 3.30 years, and Alzheimer's disease classification AUC increased from 0.78 to 0.85. Overall, our framework enhances cross-site image consistency, preserves anatomical fidelity, and improves downstream model performance, providing a robust and generalizable solution for large-scale multicenter neuroimaging studies.

Figures (5)

• Figure 1: A high-level functional diagram of the model training procedure.
• Figure 2: Visualizations of image generation during inference. (a) Generation with target scanner-site attributes. (b) Generation using style-agnostic.
• Figure 3: Axial, coronal, and sagittal slices from 9 MR images, both from the original scans (C-M) and their harmonized versions according to scanner-site pair A (left column), scanner-site pair B (middle column), and a style-agnostic target (right column). Heatmaps indicate pixel-wise differences between the original and harmonized images.
• Figure 4: Comparison of voxel intensity mean distributions for each scanner-site pair in the test images, before (a) and after harmonization using the scanner-site pair B (b), the scanner-site pair A (c), and the unspecified scanner-site as the target scanner-site attributes.
• Figure 5: Heatmaps showing the pairwise Wasserstein distances between mean voxel intensity distributions for all scanner-site pairs in the healthy control test dataset. Panel (a) shows distances before harmonization, while panels (b), (c), and (d) show distances after harmonization to scanner-site pair B, scanner-site pair A, and the style-agnostic target, respectively.