Early Stage Detection of Colorectal Cancer using Segmentation of Polyps

Colorectal Cancer(CRC) is a major health problem around the world, accounting for a high number of cancer related deaths. It is important to detect the polyp at the initial stage and remove it to prevent them from becoming cancerous. Colonoscopy is the standard examination for CRC, but pathologists face difficulties in detecting the polyps in the colonoscopy image due to the small size and the color contrast between the polyp and its background. Expert pathologists are very less in numbers compared with the cancer patients.therefore, an automated system is required to assist pathologists in the detection of polyps in early stage. Deep learning models help the pathologists to detect polyps at an early stage automatically that may be missed otherwise due to their small size, low contrast, and presence of extremely small polyps in a single image. We propose light weight model Transformer ResU- Net3plus (TransResU-Net3+) for automated segmentation of polyps in early stage of cancer. Proposed method consists of residual blocks using ResNet-50 as the backbone and also uses transformer self-attention and dilated convolutions. We have applied the proposed method on a publicly available dataset kvasir-seg and achieved an Intersection Over Union (IOU) of 0.892 and outperforms over existing state of the art methods on kvasir-seg dataset.