Motivation: The technical variability caused by the batch differences could hinder the ability to aggregate neuroimaging data across sites negatively impacting the reliability of study-level results. Goal(s): We propose a denoising diffusion probabilistic model-based method which handles multiple-domain harmonization task by harmonizing domain variant features while retaining domain invariant biological features. Approach: We have demonstrated the diffusion model's superior capability in harmonizing images from multiple domains. We also incorporated 3D fusion network to overcome the in-between-slice difference issue led by 2D harmonization work. Results: Our proposed method yields harmonization results with consistent anatomy preservation and 12.7% FID score improvement compared to other GAN-based methods. Impact: This work showed efficacy of using diffusion model to tackle neuroimaging harmonization problem with the preservation of anatomical and biological details. It is specifically evaluated to harmonize the imaging texture heterogeneity present in the large cohorts of multi-center dataset.
Lan et al. (Tue,) studied this question.