Can Programmable Gradient Information Enhance Polyp Segmentation?

Colorectal cancer ranks as the third leading cause of cancer-related deaths globally, with early detection and removal of polyps crucial in reducing mortality. Despite colonoscopy being an effective diagnostic tool, up to 28% of polyps are missed due to operator variability. Automated polyp segmentation algorithms provide a solution to enhance diagnostic accuracy. However, challenges such as boundary pixel misdetection, varying polyp sizes, and poor model generalization remain prevalent. In this paper, we propose a novel polyp segmentation approach utilizing the YOLOv9 model and its training strategy called Programmable Gradient Information. The study evaluates whether these innovations outperform state-of-the-art polyp segmentation models and introduce a new direction for future architectures. This study addresses the challenges and significantly outperforms state-of-the-art models across five widely-used public datasets. The YOLOv9 based PGI approach exhibits remarkable generalization ability, and demonstrates superior performance, achieving mDice scores of 0.932 on CVC-300, 0.866 on CVC-ColonDB, 0.889 on ETIS-Larib, and 0.919 on KvasirSEG. The model's high inference speed of 42.3 FPS, make it suitable for real-time applications. These results highlight its potential to improve early colorectal cancer detection.