GAN based approaches for self-supervised segmentation: A comparative study

Abstract

Image segmentation is a fundamental image processing technique that involves di-viding an image into distinct regions or segments to enable the analysis and extrac-tion of valuable information. It finds applications in various fields including medi-cine, pattern recognition, computer vision, and security surveillance. Different types of image segmentation techniques can be employed based on specific application requirements, including thresholding, region-based segmentation, edge-based segmen-tation, clustering-based segmentation, active contour models (Snakes), graph-based segmentation, watershed segmentation, and deep learning-based segmentation. the latter has become a powerful tool in image segmentation. Deep learning (DL) models can be trained on large datasets of images, allowing them to learn intricate relation-ships between pixels and object classes. This is particularly beneficial for challenging segmentation tasks. However, one significant challenge is the scarcity of labeled training data. To address this issue, self-supervised approaches using generative mod-els like Generative Adversarial Networks (GANs) provide a solution. GANs can gener-ate synthetic training data, which is useful for tasks where acquiring labeled training data is difficult or expensive, such as in medical imaging or remote sensing applica-tions. Additionally, GANs can generate realistic segmentation masks, which are cru-cial for tasks like medical imaging. In this study, we conducted a comparative analy-sis of DL-based segmentation approaches for COVID-19 CT scans. The evaluated approaches were assessed using metrics such as Dice Score, Specificity, and Sensitivi-ty. These metrics provide quantitative measures of the segmentation performance, allowing for an objective evaluation and comparison of the different techniques.