Polyp-DDPM: Diffusion-Based Semantic Polyp Synthesis for Enhanced Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Gastrointestinal (GI) polyp segmentation faces challenges due to limited medical data, high annotation costs, and privacy concerns.
• Existing methods struggle with generating diverse and high-quality synthetic data for training segmentation models.

Purpose of the Study:

• To introduce Polyp-DDPM, a novel diffusion-based model for generating realistic polyp images conditioned on segmentation masks.
• To enhance the performance of polyp segmentation models through improved data augmentation.

Main Methods:

• Utilized a diffusion model conditioned on binary segmentation masks to generate synthetic polyp images.
• Evaluated image quality using Fréchet Inception Distance (FID) and segmentation performance using Intersection over Union (IoU).

Main Results:

• Polyp-DDPM achieved a superior FID score of 78.47, outperforming baseline models.
• The method resulted in an IoU of 0.7156 for polyp segmentation, surpassing baseline synthetic data and approaching real data performance.
• Generated synthetic data comparable in quality and segmentation enhancement capabilities to real medical images.

Conclusions:

• Polyp-DDPM effectively generates high-quality, diverse synthetic polyp images, addressing data limitations in medical imaging.
• The approach significantly enhances polyp segmentation model performance, offering a valuable tool for medical data augmentation.
• Publicly available code and weights facilitate further research and application in GI tract polyp analysis.