Abstract
Little is known about the feasibility of virtual staining for industry applications such as high-throughput screening (HTS). We provide a thorough analysis of the usability of image-to-image translation for the virtual staining of label-free bright-field microscopy images of live cells, using a pool of more than 1.6 million images across six lung, six ovarian and six breast cell lines consisting of paired bright-field, cytoplasm, nuclei and DNA-damage stains. To our knowledge this is the first time an analysis of virtual staining has been performed on three levels; pixel-based, biological-feature based, and determining if virtual staining can reproduce drug-effect. Our results reveal that while virtually stained nuclei and cytoplasm images often consistently and faithfully reproduce the information found in fluorescence microscopy, virtually stained images of DNA-damage are usually less accurate.
Samuel Tonks
PhD Student in Computer Science
My research interests include generative modelling, uncertainty quantification and healthcare applications.