Cell And Sub-Cellular Segmentation In Quantitative Phase Imaging Using U-Net

Abstract

The ability to automatically segment images, especially microscopy images of cells, opensnew opportunities in cancer research or other practical applications. Recent advancements in deeplearning enabled for effective single-cell segmentation, however, automatic segmentation of subcellularregions is still challenging. This work describes an implementation of a U-net neural networkfor label-free segmentation of sub-cellular regions on images of adherent prostate cancer cells,specifically PC-3 and 22Rv1. Using the best performing approach, out of all that have been tested,we have managed to distinguish between objects and background with average dice coefficients of0.83, 0.78 and 0.63 for whole cells, nuclei and nucleoli respectively