A Multi-Scale Patch-Based Deep Learning System for Polyp Segmentation

Abstract

Colorectal cancer (CRC) is one of the deadliest forms of cancer and is on the rise. Accurate segmentation of the precursor lesion, the polyp, can ensure the survival rate. Hence, there is a research trend to develop medical diagnosis support tool to assist clinicians. This study focuses on the development of deep learning-based CNN model for automated segmentation of polyp. As polyp varies frame to frame in terms of size, color, shape, and texture, segmentation is still an unsolved problem and a very challenging task. We have proposed a multi-scale patch-based CNN model for automatic segmentation of the polyp region. Local and global patches are extracted from each pixel of the input image and fed into two similar CNNs of which one is responsible for the extraction of local features and the other for extraction of global features that are being concatenated for accurately pixel label annotation of the polyp region. As in colonoscopy frames, there are some regions with the same intensity/texture as the polyp regions, so in the predicted segmentation map, some non-polyp regions are also considered as polyp regions, which are further refined by post-processing operation. The proposed model is evaluated on CVC-Clinic DB. The experimental result shows that our proposed method outperforms other baseline CNNs and state-of-the-art methods.