Reproduce “Diffusion Probabilistic Models beat GANs on Medical Images” on OpenResearch

About this paper

Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han and 3 more

Medfusion, a conditional latent Denoising Diffusion Probabilistic Model, consistently outperforms state-of-the-art Generative Adversarial Networks (GANs) in generating synthetic medical images across fundus, histopathology, and chest radiograph modalities. The model achieves significantly lower Fréchet Inception Distance (FID) scores (e.g., 11.63 vs. 20.43 for AIROGS) and produces more diverse, realistic synthetic data with fewer characteristic artifacts than GANs.

Key insights

• Denoising Diffusion Probabilistic Models (DDPMs) consistently outperform Generative Adversarial Networks (GANs) for synthetic medical image generation across diverse modalities, addressing issues like mode collapse and artifact generation.
• Optimizing the autoencoder within the latent diffusion framework, specifically by increasing channels from 4 to 8, significantly improves the reconstruction quality of fine details critical in medical images.
• Systematic comparison of diffusion models against established GANs across 2D medical imaging modalities demonstrates that DDPMs offer superior fidelity and diversity in generated data.

Summary by alphaXiv.