Tortuosity Controllable Retinal Fundus Image Generation using Diffusion Model

Deep learning has shown great promise in retinal image generation and augmentation, offering opportunities to enhance diagnostic performance by synthesizing high-quality, diverse datasets. However, few studies have focused on developing reliable models that allow control over image characteristics, such as vessel tortuosity. To address this gap, we propose a novel diffusion-based controllable image generation framework. The approach integrates a quantitative tortuosity control module to generate more tortuous vessel masks, simulating disease cases where highly tortuous vessels are commonly observed. These masks serve as input prompts to guide a generation process. Experiments and comparative studies against state-of-the-art methods validate the quality and anatomical fidelity of the generated images, showing superior controllability and realism. This work demonstrates the potential of controllable generative models to overcome data limitations in retinal fundus imaging and support the development of deep learning-based models.