Reproduce “Paired Image Generation with Diffusion-Guided Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haoxuan Zhang, Wenju Cui, Yuzhu Cao, Tao Tan, Jie Liu, Yunsong Peng, Jian Zheng
A diffusion-guided diffusion model generates paired medical images and their corresponding annotations simultaneously, addressing data scarcity for segmenting concealed mass lesions in Digital Breast Tomosynthesis (DBT) images. The method achieved an 8.18 percentage point improvement in Dice score for downstream segmentation compared to baseline, and a 15.66 lower FID score than unconditional diffusion models.
Key insights
- The joint denoising of paired images can be effectively implemented by mutually guiding conditional diffusion processes, eliminating the need for external conditional inputs.
- Using a predicted clean image (e.g., denoised mask) as a guiding signal for the generation of its paired image (e.g., DBT slice) improves generation quality and consistency, particularly for subtle features like lesions.
- The internal guidance mechanism enables the network to focus on and accurately generate critical lesion regions, overcoming challenges faced by purely unconditional generation in concealed areas.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.