Patch-level Anatomy- and Style-Guided 3D Diffusion for Multi-site T1 MRI Harmonization

Barnabé Hache, Vincent Roca, Grégory Kuchcinski, Dorian Manouvriez, Renaud Lopes

Keywords: MRI harmonization, diffusion models, patch-based synthesis, classifier-free guidance, anatomical representation

TL;DR: We introduce a patch-level 3D diffusion framework, conditioned on anatomical priors and learnable site embeddings, that harmonizes multi-site T1 MRI to a target scanner, reducing site-related variance while preserving fine anatomical detail.

Abstract: Pooling brain T1-weighted MRI from multiple sites increases statistical power but introduces non-biological scanner effects that bias analyses. Generative diffusion models generalize well and train stably, making them attractive for MRI harmonization. However, most prior 3D works use latent-space diffusion for efficiency, and the compression–decoding step can introduce blurring that undermines preservation of fine anatomical detail. We present a patch-level 3D diffusion framework that harmonizes volumes to a chosen target site while preserving individual anatomy. The model is jointly conditioned on an explicit anatomical prior derived from each volume and on a learnable site embedding injected via cross-attention; inference uses manifold-constrained classifier-free guidance to steer site style without displacing anatomy. Working on overlapping high-resolution patches avoids decoder-related blurring and preserves fine structural detail. We train on 1,000 scans pooled from seven public cohorts and evaluate on the SRPBS traveling-subject dataset. Compared to preprocessed images and three published baselines, our method yields modest improvements in image similarity (SSIM 0.864 → 0.874) and substantially reduces variance attributable to site (gray matter: 72.9\% → 50.7\%; white matter: 59.9\% → 21.4\%), while subject-level variability remains high, indicating preservation of biological differences. The code and model weights will be made available.

Primary Subject Area: Image Synthesis

Secondary Subject Area: Generative Models

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 205