Towards Fluorescent-Tag-Less Viral Titration: Automated Estimation of Cell-Size Distribution and Infection Level from Phase-Contrast Microscopy Using Deep Learning and Transfer Learning

Automated detection of infected insect cells is one of the crucial tasks in the field of recombinant protein production and vaccine development. The major challenge lies in manual segmentation of cells and quantifying cell size distribution is tedious and requires extensive effort. While such assessment of the infection levels is possible through fluorescent imaging, it requires tagging the expressed protein with a fluorescent marker. In contrast, we present a deep learning-based approach to obtain the cell size and thereby to detect the morphological change in infected cells from phase contrast images. First, We tested the performance of HOG (Histogram of oriented gradients)+ SVM(Support vector machine), Faster RCNN (Region-based Convolutional Neural Network) and YOLO [1] to detect cells and calculate cell size distribution for infected and uninfected cells. The results show that YOLO performs better in detecting cells and calculating cell size with a limited dataset using transfer learning. Further, we validate that the size distribution shows a significant difference in cell size for infected and non-infected cells. The YOLO-based cell size learning has the potential to be used for the classification of infection levels without fluorescent tagging to the protein of interest.