Saliency and ballness driven deep learning framework for cell segmentation in bright field microscopic images

Abstract

Cell segmentation is the most significant task in microscopic image analysis as it facilitates differential cell counting and analysis of sub-cellular structures for diagnosing cytopathological diseases. Bright-field microscopy is considered the gold standard among different types of optical microscopes used for cell analysis due to its simplicity and cost-effectiveness. However, automatic cell segmentation in bright field microscopy is challenging due to imaging artifacts, poor contrast, overlapping cells, and wide variability of cells. Also, the availability of labeled bright-field images is limited, further constraining the research in developing supervised models for automated cell segmentation. In this research, we propose a novel cell segmentation framework termed Saliency and Ballness driven U-shaped Network (SBU-net) to overcome these challenges. The proposed architecture comprises a novel data-driven feature fusion module that enhances the perceivable structure of cells using its saliency and ballness features. This, together with an encoder–decoder model having dilated convolutions and a novel combination loss function, captured the global context of cell structures and produced accurate cell segmentation results. SBU-net is evaluated using two publicly available bright-field datasets of T cells and pancreatic cancer cells. The model is subjected to 5-fold cross-validation and outperformed state-of-the-art models by producing mean Intersection over Union (IoU) scores of 0.804, 0.829, and mean Dice of 0.891, 0.906, respectively. The architecture was also tested on a fluorescent dataset to see how well it could generalize, and it came out with a mean IoU of 0.892 and a mean Dice of 0.948, outperforming other models reported in the literature.