An Early-Stage Colorectal Cancer Detection from Colonoscopy Images Using Enhanced Res-UNET

The segmentation of the polyps during colonoscopy is one of the most crucial factors for successful colon cancer diagnosis. Due to the wide variety of sizes and shapes of these polyps, this process can be challenging. The introduction of deep learning models in biomedical image analysis has shown a great impact on disease detection and segmentation. Most of the deep learning models concentrated on larger size polyps only, smaller size polyps may not segmented accurately during the procedure and can cause cancer in 5–7 years. This can have a significant impact on the early-stage detection of colon cancer. This chapter proposes an enhanced Res-UNET-based deep learning–based model that can improve the performance of segmentation on small-size polyps. We evaluated the proposed model on ETIS-LaribPolypDB, CVC-ColonDB, CVC-ClinicDB, Kvasir-SEG colonoscopy image datasets. The results show that the proposed enhanced Res-UNET model achieves the top Dice segmentation accuracy and top Hausdorff distance over the recent state-of-the-art models.