Related Experiment Video
Updated: Apr 25, 2026
14:38
Creating Objects and Object Categories for Studying Perception and Perceptual Learning
Published on: November 2, 2012
13.6K
Cycle Diffusion Model for Counterfactual Image Generation
Fangrui Huang1, Alan Wang1, Binxu Li1
1Stanford University, Stanford CA 94305, USA.
Predictive Intelligence in Medicine. PRIME (Workshop)
|April 24, 2026
Summary
Cycle Diffusion Models (CDM) improve medical image synthesis by ensuring generated images match original data. This enhances realism and accuracy for applications like data augmentation and disease modeling.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep generative models excel at medical image synthesis but struggle with conditioning faithfulness and realism.
- Direct and counterfactual generation of high-quality synthetic medical images remains a significant challenge.
Purpose of the Study:
- To introduce a novel cycle training framework to fine-tune diffusion models for enhanced medical image generation.
- To improve conditioning adherence and synthetic image realism in diffusion models.
Main Methods:
- Developed the Cycle Diffusion Model (CDM), a framework incorporating cycle constraints to enforce consistency between generated and original images.
- Fine-tuned diffusion models using the CDM approach on a combined 3D brain MRI dataset (ABCD, HCP, ADNI, PPMI).
Main Results:
- CDM significantly improved conditioning accuracy compared to baseline diffusion models.
- Enhanced synthetic image realism and quality, as evidenced by improved FID and SSIM scores.
- Demonstrated reliable direct and counterfactual generation capabilities.
Conclusions:
- The cycle strategy in CDM is an effective method for refining diffusion-based medical image generation.
- CDM shows promise for applications in medical data augmentation, counterfactual analysis, and disease progression modeling.
