Scanner-agnostic MRI harmonization via SSIM-guided disentanglement

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 Measure (SSIM) to preserve biologically meaningful features while reducing inter-site variability. This formulation allows luminance, contrast, and structural components to be modeled separately during optimization. Training and validation were performed on multiple publicly available datasets spanning diverse scanners and sites, with testing on both healthy individuals and populations with pathological conditions. The proposed approach was evaluated across multiple target settings, including scanner-site-specific targets and a style-agnostic target, and compared with representative image-based harmonization benchmark methods.

Across these target settings, harmonization produced consistent and high-quality outputs. Visual comparisons, voxel intensity distributions, and SSIM-based metrics demonstrated that harmonized images achieved improved alignment across acquisition settings while preserving anatomical fidelity. In the style-agnostic setting, within-subject original-harmonized comparisons showed high anatomical preservation, with the structural component of SSIM reaching 0.975 ± 0.007. Appearance consistency also improved, with Wasserstein distances between mean voxel intensity distributions decreasing from 8.45 ± 5.35 before harmonization to 1.77 ± 0.62, and luminance similarity increasing from 0.952 ± 0.037 to 0.982 ± 0.017. Downstream analyses further confirmed the effectiveness of the proposed approach. For brain age prediction, mean absolute error decreased from 4.08 ± 1.16 to 2.81 ± 0.55 years following style-agnostic harmonization. For Alzheimer's disease classification, the area under the ROC curve improved from 0.857 ± 0.038 to 0.899 ± 0.024. Compared with the considered benchmark methods, the proposed framework showed stronger image-level harmonization and more consistent downstream improvements under the adopted evaluation protocol.

Overall, the proposed framework enhances cross-site image consistency, preserves anatomically relevant information, and improves downstream predictive performance, providing a robust and generalizable solution for large-scale multicenter neuroimaging studies.