Colon polyp elucidation through U-Net Architecture

Abstract

It is well understood that detection and classification of colon polyps are important for diagnosis and prevention of CRC. This paper examines the use of the U-Net architecture, a CNN particularly developed for the biomedical image segmentation in the identification of colon polyps from colonoscopy images. Another model, the U-Net which is known for its performance even when training and validating with a small set of images and for generating highly accurate segmentation maps was trained and validated with a set of colonoscopy images with annotations. In our method, we utilize the decoder path of the U-Net for getting the precise localization through the upsampling layers, and the encoder path for collecting context through the down-sampling layers. This technique enables one to draw a clear boundary of the polyps, which in turn makes segmentation or categorization easier. The results indicate that the UNet design is effective in terms of IoU and Dice coefficient, solidifying its robustness and applicability in differentiating polyps of varying sizes and shapes. The model’s applicability in clinical practice, specifically in real-time polyp detection during colonoscopy operations, is underscored by its ability to accurately produce high-resolution segmentation maps out of comparatively small training sets.