Unpaired Multi-Site Brain MRI Harmonization with Image Style-Guided Latent Diffusion

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Multi-site brain MRI data exhibit heterogeneity due to variations in scanners and protocols.
• This heterogeneity poses challenges for consistent data analysis and interpretation.
• Existing image harmonization methods often require paired data or complex models.

Purpose of the Study:

• To develop an unpaired MRI harmonization framework (UMH) to address multi-site data heterogeneity.
• To leverage a novel image style-guided diffusion model for harmonization.
• To eliminate the need for paired data and reduce computational complexity.

Main Methods:

• UMH employs a two-stage approach: a coarse harmonizer using a conditional latent diffusion model and a fine harmonizer utilizing CLIP-derived style embeddings.
• The framework aligns multi-site MRIs to a unified domain while preserving anatomical information.
• CLIP embeddings capture semantic style differences, avoiding explicit style learning.

Main Results:

• UMH demonstrated superior performance compared to state-of-the-art methods on three multi-site datasets (4,123 MRIs).
• Evaluations included image-level comparisons, downstream classification, and brain tissue segmentation tasks.
• The method effectively harmonized MRI data without requiring paired samples.

Conclusions:

• The proposed UMH framework offers an effective solution for harmonizing heterogeneous multi-site brain MRI data.
• UMH reduces computational costs and eliminates the need for paired training data.
• This approach enhances the consistency and reliability of MRI analysis across different sites.