Abstract
Automated diagnostic systems (ADS) have shown significant potential in theearly detection of polyps during endoscopic examinations, thereby reducing theincidence of colorectal cancer. However, due to high annotation costs andstrict privacy concerns, acquiring high-quality endoscopic images poses aconsiderable challenge in the development of ADS. Despite recent advancementsin generating synthetic images for dataset expansion, existing endoscopic imagegeneration algorithms failed to accurately generate the details of polypboundary regions and typically required medical priors to specify plausiblelocations and shapes of polyps, which limited the realism and diversity of thegenerated images. To address these limitations, we present Polyp-Gen, the firstfull-automatic diffusion-based endoscopic image generation framework.Specifically, we devise a spatial-aware diffusion training scheme with alesion-guided loss to enhance the structural context of polyp boundary regions.Moreover, to capture medical priors for the localization of potential polypareas, we introduce a hierarchical retrieval-based sampling strategy to matchsimilar fine-grained spatial features. In this way, our Polyp-Gen can generaterealistic and diverse endoscopic images for building reliable ADS. Extensiveexperiments demonstrate the state-of-the-art generation quality, and thesynthetic images can improve the downstream polyp detection task. Additionally,our Polyp-Gen has shown remarkable zero-shot generalizability on otherdatasets. The source code is available athttps://github.com/CUHK-AIM-Group/Polyp-Gen.