SU-NET Based colorectal polyp Segmentation from Colon Cancer Morphology Images

Abstract

Precise demarcation of glands from clinical histology images are pre-requirement for accurate medical diagnosis. Colorectal polyps that originate and expandsover the rectum or colon membrane are the decisive reason for colorectal Cancer(CRC). The early-stage recognition and of polyps and treatment can decrease the mortality rate. To lower the polyp miss-rate in colonoscopy, a Computer-Aided Medical Diagnosing(CAD) system with high accuracy is needed. In recent times, researchers develop deep learning models for accurate polyp detection from histomorphology images, but accuracy is still the most requisite factor for reliable results. In this paper, we propose to develop and test a Convolutional Neural Network(CNN) based U-shape network (SU-NET) model for semantic segmentation of colorectal polyps from colonoscopy images.SU-NET is an Encoder-Decoder-based architecture, inspired by the popular segmentation architectures SegNet and U-Net for improved colon polyp segmentation. In the proposed model the top most layers transfer the Pooling indices whereas the lower-level layers transfer the feature-maps to incorporate fine multiscale information for better colon polyp contour identification.We evaluated the proposed algorithm in contrast with various prominentdeep learning architectures across multi-modal biomedical image segmentation tasks to segment polyps from the colonoscopy and histopathology images.For evaluating the proposed model, an accredited and publicly available colonoscopy image dataset CVC-ColonDB is employed. The model achieves a recall of 91.3%, F1-Score of 90.81%, F2-Score of 86.39%, Precision of 89.21%, and the Dice similarity coefficient of 0.895 outshines the existing advanced deep learning CNN models.