Publication

Abstract : Colonoscopy is a medical procedure performed to detect the anomalies in the colon and rectum. A thin flexible wire embedded with the camera is inserted to directly visualize the colon. Direct visualization enables early detection and removal of the polyps in the colon. Polyps exist in different shapes and sizes. The physicians find it very challenging to diagnose small polyps in colonoscopy video. Delay in polyp removal leads to colorectal cancer and leads to cancer-related death. In this work UNET architecture with spatial attention layer is proposed to improve the precision of segmenting polyp regions in colonoscopy video. The CNN models proposed in the literature for polyp segmentation are basically trained using common loss functions such as dice and binary cross entropy loss. Under the training of these loss functions the model learns poorly in segmenting small sized polyps. This can be solved by using focal Tversky loss. Experiments are conducted on a publicly available dataset. Results show that UNET with spatial attention layer trained with focal Tversky loss performs better compared to standard UNet model trained with common loss functions.