In this paper, we propose DynamicDPS, a novel method to reduce hallucinations (structures that are not present in the real data), which is a serious problem in medical image reconstruction, especially in data-driven conditional models. DynamicDPS is a diffusion-based framework that integrates conditional and unconditional diffusion models to systematically reduce hallucinations while improving low-quality medical images. After an initial reconstruction using the conditional model, it is improved using an adaptive diffusion-based inverse problem solver, and the optimal starting point is selected sample-wise and Wolfe's linear search is applied to improve efficiency and image fidelity. Through extensive evaluations on synthetic and real MR scans, we show that DynamicDPS reduces hallucinations and improves tissue volume estimation by more than 15% while using only 5% of the sampling steps compared to conventional diffusion models. As a model-independent and learning-free method, it provides a powerful solution to hallucination reduction in medical images.
