Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Diffusion models excel with large datasets for image generation, but medical data access is limited by privacy and cost.
• Fine-tuning general diffusion models for medical imaging is challenging due to domain distribution shifts.

Purpose of the Study:

• To introduce Latent Drift (LD), a novel method to address distribution shifts in diffusion models for medical imaging.
• To enable counterfactual image generation in medical domains, allowing exploration of parameter effects like age and disease.

Main Methods:

• Latent Drift (LD) is proposed as a technique adaptable to various fine-tuning methods or usable during inference.
• The method conditions diffusion models for specialized medical image tasks, including counterfactual generation.

Main Results:

• Evaluated on three public datasets (brain MRI, chest X-rays) for counterfactual generation.
• Demonstrated significant performance improvements across diverse scenarios and fine-tuning approaches.

Conclusions:

• Latent Drift effectively mitigates distribution shift issues in medical diffusion model applications.
• The method shows promise for generating high-fidelity counterfactual medical images, aiding research into patient parameter alterations.