AI-aided holographic flow cytometry for label-free identification of ovarian cancer cells in the presence of unbalanced datasets

Abstract

Liquid biopsy is a valuable emerging alternative to tissue biopsy with great potential in the noninvasive early diagnostics of cancer. Liquid biopsy based on single cell analysis can be a powerful approach to identify circulating tumor cells (CTCs) in the bloodstream and could provide new opportunities to be implemented in routine screening programs. Since CTCs are very rare, the accurate classification based on high-throughput and highly informative microscopy methods should minimize the false negative rates. Here, we show that holographic flow cytometry is a valuable instrument to obtain quantitative phase-contrast maps as input data for artificial intelligence (AI)-based classifiers. We tackle the problem of discriminating between A2780 ovarian cancer cells and THP1 monocyte cells based on the phase-contrast images obtained in flow cytometry mode. We compare conventional machine learning analysis and deep learning architectures in the non-ideal case of having a dataset with unbalanced populations for the AI training step. The results show the capacity of AI-aided holographic flow cytometry to discriminate between the two cell lines and highlight the important role played by the phase-contrast signature of the cells to guarantee accurate classification.