Deep Semantic Segmentation of Cell Painted Nuclei Images using UNet++

Fluorescence microscopy based cell painting technique profiles the morphological characteristics of specific cell organelles with high resolution. However, photo toxicity, photo bleaching and advanced instrumentation limits its utility for comprehensively annotating the cell structure. Generating cell painted organelles from simple and least invasive transmitted light microscopy provides a surrogate for clinical applications. In this study, the employability of semantic segmentation model for delineating the nuclei from composite image using UNet++ is investigated. For this, a public dataset consisting of 3456 composite images are considered from Broad Bioimage Benchmark collection. The binary masks of endoplasmic reticulum (ER), nuclei and cytoplasm are generated for pixel wise labelling. The composite images and their labelled dataset are fed to the UNet++ model for segmenting the cell painted nuclei. The performance of the deep semantic network is analysed for 50 epochs and the segmentation results are validated with mean intersection-over-union (IoU). The UNet++model provides an accuracy of 95.8% with a minimum loss of 0.1. Mean IoU of 0.91 is achieved for the prediction of nuclei from the composite images. The obtained results suggest that this study could be employed for predicting the subcellular components from transmitted light microscopy without the use of fluorescent labelling