Label-free detection of individual virus-infected cells using deep learning

Abstract Numerous applications in research and medicine rely on reliable identification and quantification of virus-infected cells. Current methods either apply reporter viruses, that often differ from clinical isolates (e. g. cell tropism, immune evasion) or staining approaches, that prevent live-cell experiments and may introduce biases through manual counting. We present a deep learning model for the label-free identification of virus-infected cells on light microscopy images (VAIruScope). To overcome limitations, our pipeline enables an automated quantification of virus-infected cells based on the recognition of cytopathic effects. The method was applied to different cell models and four clinically relevant prototype viruses representing RNA- (influenza A virus), DNA- (human cytomegalovirus, herpes simplex virus-1) and retroviruses (human immunodeficiency virus-1). VAIruScope identified infected cells achieving classification accuracies of up to 96 %. As proof-of-concept, the method was validated using electron microscopy for a wild-type HSV-1. VAIruScope may be applicable to live-cell imaging to investigate infection dynamics.