Color space and colorectal polyps segmentation: An empirical study

Abstract

Colorectal polyps are a prevalent precursor to colorectal carcinoma. In recent years, there has been an increasing interest among deep learning researchers in developing automated neoplasm detection systems to serve as assistive tools for clinicians in detecting diminutive and inconspicuous polyps. Precise segmentation of these neoplasms in medical images is essential for early detection and intervention. While current research efforts focus on enhancing segmentation performance and achieving new state-of-the-art results, a comprehensive analysis of the various factors influencing the performance of neoplasm segmentation models remains to be conducted. In this study, we investigate the impact of color space on the performance of colorectal neoplasm segmentation networks. We employ three pre-trained semantic segmentation architectures: U-NET, DeepLabV3 and Pyramid Attention Network (PAN) to elucidate the relationship between the color space of input images and model performance. We examine this relationship using four color spaces: 1) RGB, 2) HSV, 3) HSL and 4) CIEL*A*B*. Four publicly available datasets: Kvasir-SEG CVC-ClinicDB, CVC-ColonDB and ETIS-LaribPolypDB, are utilized for training and testing. Our findings indicate that the choice of color space can significantly influence the performance of colorectal neoplasm segmentation networks.