GH-DDM: the generalized hybrid denoising diffusion model for medical image generation

Deep-learning-based medical imaging plays a pivotal role in modern healthcare while suffering from the data scarcity bottleneck, since obtaining sufficient high-quality data in the medical imaging area is difficult and expensive. To alleviate this problem, existing methods adopt the convolutional Generative Adversarial Networks to generate diverse images while having limits come from (1) lack of global physiological structural capture ability; (2) inherent mode collapse problem; and (3) lack the generalization to universal tasks. Therefore, this paper presents the Generalized Hybrid Denoising Diffusion Model (GH-DDM) for medical image generation, which leverages the strong abilities of transformers into diffusion models to model long-range interactions and spatial relationships between anatomical structures, and further proposes several key modifications to make our model easy to generalize to various kinds of generation tasks. Extensive experiments are conducted over multiple medical datasets of diverse modalities, including computed tomography (CT) scans, X-rays, retinal OCT, etc. The visualization and quantitative results demonstrate the efficacy and good generalization of our model for generating a wide array of high-quality medical images and benefit downstream medical tasks.