Continuous Cell Cycle Representation Using Ordinal Regression and Siamese Network from Quantitative Phase Images

The cell cycle plays a crucial role in governing the growth and development of living organisms. Aberrations in this mechanism can be indicators of cancer development. Hence, accurate evaluation of the cell cycle stage is essential for predicting potential diseases and abnormalities.Previous studies mainly focused on directly classifying discrete cycle stages. Our goal is the prediction of the continuous cell cycle: we propose a method to infer its value for single-cell image patches from quantitative phase image time-lapse. Our model can predict a discrete cell cycle phase and a pseudo-time value representing the current cycle position from cell images.We use a convolutional neural network with two outputs as our baseline. In addition, we exploit information from cell tracks by using a siamese-type network that penalizes large differences between predictions for consecutive images of a cell. Finally, we propose simple metrics to evaluate how continuous the resulting representation is.