Semantic segmentation of palpebral conjunctiva using predefined deep neural architectures for anemia detection

Abstract

Non-invasive detection of anemia is generally done by physical examination of regions like palpebral conjunctiva, fingernails, tongue and palmar creases. However, such examination is subject to large inter- and intra-observer bias. This problem can be alleviated by automating the anemia detection process through computerized analysis of images of these regions. The automated process includes sub-processes like preprocessing, segmentation of region-of-interest (ROI), ROI analysis or feature extraction, and classification. Of all these sub-processes, segmentation is the most crucial one as it helps in the precise extraction of ROI where the most crucial information for decision making lies. Recently, deep learning-based architectures have given exemplary performance in biomedical image segmentation. This paper is unique in a sense that it simulatneously analyzes performance of five deep learning-based architectures namely UNet, UNet++, FCN, PSPNet, and LinkNet. The experiments are performed on customly built dataset comprising of 2592 palpebral images of the pediatric population. The experimental results indicate that as compared to its counterparts, the LinkNet architecture performs the best. Its scores 94.17%, 90.14% and 93.78% for the accuracy, intersection-over-union (IoU), Dice score performance metrics, respectively. The study concludes that LinkNet architecture can be used for real-time segmentation of palpebral conjunctiva from images.