Harmonizing Multi-Site Multi-Sequence Brain MRI via Semantic-Guided Conditional Diffusion

Mengqi Wu, Minhui Yu, Hongtu Zhu, Pew-Thian Yap, Mingxia Liu

Keywords: Brain MRI Harmonization, Style Translation, Diffusion

TL;DR: We propose a semantic-guided conditional diffusion model for unpaired multi-sequence MRI harmonization that preserves anatomy, handles diverse imaging styles, and outperforms existing methods.

Abstract: Training robust AI models for brain MRI analysis typically requires large datasets, prompting many studies to aggregate multi-site data. However, this introduces unwanted variations due to differences in scanners and/or acquisition protocols. These non-biological variations (known as site effects) can significantly compromise the performance and generalizability of downstream deep learning models. While image-level harmonization has emerged as a promising solution, existing methods frequently demand paired data (e.g., scans of the same subject at different sites) or costly encoder-decoder networks to disentangle anatomical content from predefined imaging style (e.g., intensity and contrast), which struggle to comprehensively capture diverse image styles. Moreover, existing methods cannot readily adapt across different MRI sequences, limiting their scalability. This paper proposes a semantic-guided conditional diffusion (SGCD) framework for unpaired 3D multi-sequence MRI harmonization. SGCD first trains a conditional diffusion model (CDM) to align multi-site, multi-sequence MRIs into a unified, sequence-specific domain, reducing global site-related variations. It then fine-tunes the CDM for target-specific harmonization using a style loss derived from BiomedCLIP, trained on medical imaging data. By capturing differences in disentangled semantic image style between the harmonized and target MRIs, this loss enables effective harmonization that preserves anatomical structure and does not require paired training data. We evaluate SGCD on 4,163 T1-/T2-weighted MRIs from three multi-site datasets, with results suggesting its superiority over several state-of-the-art methods across voxel-level comparison, downstream classification, and brain tissue segmentation tasks.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 7827