Development of an image-based Deep Learning model for Schizosaccharomyces pombe cell cycle phase classification

Abstract

Despite medical advances, cancer remains a life-threatening disease, accounting for millions of deaths per year. Aiming to develop novel cancer treatments, understanding the nature of cancerous cells and mechanisms of activation and progression at the microscale are material of study of many investigators. Special interest has been given to the actin cytoskeleton, whose morphology is altered in many cell functions -such as cell division- under tumorous processes. However, observing and manually identifying cells according to their intracellular architecture might be time-consuming for the investigator. In light of the above, this study intents to construct a Convolutional Neural Network -a type of Deep Learning model- for Schizosaccharomyces pombe -a common model of study- classification according to their actin cytoskeleton during the cell cycle. For this purpose, a dataset containing representative images of different actin phenotypes was used for the training and testing phases, as well as for the final validation of the constructed model. The outcoming results demonstrate the successful learning capacity of the algorithm, whose evaluation metrics depicted a nearly perfect model. Nonetheless, when facing unseen data, its reliability is questioned, since it fails to correctly identify a considerable proportion of the introduced unseen images. For this reason, the algorithm prediction cannot be considered as the absolute truth but as a complementary tool. In order to improve the predictive ability of the model and seeking for a better performance, a future dataset reconstruction and expansion, with a subsequent validation, should be performed.