Deep learning-based polyp segmentation network with a dual encoder-decoder

Abstract

Colorectal cancer (CRC) constitutes one of the most serious forms of cancers across the world. Current screening procedures through imaging with colonoscopy are costly, laborious, and time consuming. Computer-aided diagnosis (CAD) and the integration of deep learning within the field have provided the opportunity for improvements with respect to human error. The recent developments with convolutional neural networks (CNNs) have demonstrated the capacity for real time detection and semantic segmentation of early signs of CRC. These early signs of CRC are referred to as polyps, and thus the identification of these potentially cancerous tumors improves patient outcomes and mortality rates associated with CRC. In an effort to compensate for some of the issues presented for current screening procedures, a novel polyp segmentation network, PSNet [1], is proposed for the semantic segmentation of polyps. PSNet compensates for current issues affecting polyp segmentation networks such as boundary pixel definitions, as well as model generalization and overfitting issues. PSNet provides state-of-the-art (SOTA) performance with respect to two major semantic segmentation performance metrics, mean intersection-over-union (mIoU) and mean Dice (mDice) on 5 major publicly available datasets, with a combined mIoU and mDice score of 0.863 and 0.797 across all sets, respectively. A new configuration of these 5 publicly available datasets is also proposed to improve upon model generalization and help demonstrate some of the issues associated with current standards of polyp datasets and performance evaluation. The mIoU and mDice score for this new configured dataset was 0.941 and 0.897, respectively, and therefore improved even further on the reported SOTA results.