Computational methodologies for reliable tomographic phase imaging flow cytometry

Abstract

In recent years, tomographic phase microscopy has gained credits as one of the most powerful imaging modality for label-free single-cell analysis in 3D. The conventional tomographic imaging systems probe the sample from different directions, experimentally fixed a priori, and the tomographic reconstruction is performed by well-known techniques. The recent demonstration of such technology in flow cytometry condition, in which cells rotate in a microfluidic channel, permits to achieve high-throughput analysis, but it is needed of a reliable and robust computational processing pipeline. In fact, no a priori information about the rotation angles of cells are available, hence suitable algorithms are designed to recover them. Moreover, the number of such rotating angles is low if compared to the number of orientation directions explored in conventional systems, thus requiring more sophisticated algorithm to reconstruct cells tomograms. Finally, due to the high-throughput modality, the huge amount of data to manage becomes one of the main computational problems to face with. Here we show an efficient computational processing pipeline to achieve reliable tomographic reconstructions based on (i) a fast quantitative phase maps recovery method based on deep learning end-to-end reconstruction, (ii) the modelling of the cells' 3D pose in microfluidic flow through holographic tracking, (iii) the use of a suitable tomographic reconstruction algorithm, (iv) a new strategy to encode single-cell phase contrast tomograms by using the 3D version of Zernike polynomials, thus allowing an efficient data storage.