Blind harmonization reinvented via 3D diffusion model

Motivation: Recently proposed blind harmonization, which enables harmonization training without source domain data, shows limitations in 2D constraints, image quality, and scalability across large domain gaps. Goal(s): We propose BlindHarmonyDiff, 3D blind harmonization framework using a diffusion model to overcome the limitations of previous methods. Approach: An edge-to-image diffusion model is trained on target domain images to reconstruct the original images from edge maps. Source domain image edges are then detected and input into this model to produce harmonized images. Results: BlindHarmonyDiff successfully harmonized source domain images, even those with large domain gaps, to the target domain and improved performance in downstream tasks. Impact: BlindHarmonyDiff is a novel blind harmonization technique that overcomes the limitations of previous methods by generating high-quality 3D harmonized images, effectively handling images with large domain gap. Its refinement module mitigates hallucinations from diffusion models, improving reliability and clinical applicability.