Robustness of Diffusion Model-Based Methods to Distribution Shifts in Medical Imaging

Motivation: Diffusion model (DM)-based methods demonstrate competitive performance in solving medical imaging inverse problems. However, it remains uncertain whether these methods are vulnerable to distribution shifts, a common issue in traditional learning-based approaches. Goal(s): Our goal is to conduct experiments to evaluate how well DM-based methods handle distribution shifts in various scenarios. Approach: We utilize three different sampling methods based on a pretrained Diffusion model and apply them to four clinical tasks without fine-tuning. Results: Experiments indicate that DM-based methods efficiently address distribution shifts without requiring fine-tuning. They exhibit robustness to many distribution shifts, even when the test data deviates from the training data. Impact: DM-based methods reduce resource consumption and the need for extensive training datasets, potentially inspiring further development of DM-based techniques for enhancing in-vivo tasks with limited resources.