Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis

This repository contains the website for our CVPR 2025 paper on counterfactual medical image synthesis using diffusion models.

Paper Abstract

We present Latent Drifting, a novel approach for counterfactual medical image synthesis using diffusion models. Medical counterfactual generation is crucial for creating controlled variations in medical images that can help with diagnosis, treatment planning, and medical education. Our method addresses a critical limitation in current diffusion-based approaches that often struggle to maintain anatomical consistency while introducing clinically meaningful variations. By strategically manipulating the latent space trajectories during the diffusion process, Latent Drifting generates highly realistic counterfactual medical images that preserve patient-specific anatomical structures while modifying targeted pathological features.

Authors

• Yousef Yeganeh (Technical University of Munich, MCML)
• Azade Farshad (Technical University of Munich, MCML)
• Ioannis Charisiadis (Technical University of Munich)
• Marta Hasny (Technical University of Munich)
• Martin Hartenberger (Technical University of Munich)
• Björn Ommer (LMU Munich, MCML)
• Nassir Navab (Technical University of Munich, MCML)
• Ehsan Adeli (Stanford University)

Website

The paper website is accessible at index.html.

Citation

@inproceedings{yeganeh2025latentdrifting,
    title={Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis},
    author={Yeganeh, Yousef and Farshad, Azade and Charisiadis, Ioannis and Hasny, Marta and Hartenberger, Martin and Ommer, Björn and Navab, Nassir and Adeli, Ehsan},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2025}
}