Deep learning assisted Fourier ptychography for cells and tissue analysis

Abstract

Fourier Ptychography (FP) is a quantitative phase imaging technique that overcomes the trade-off between lateral resolution and field of view. Typically, low numerical aperture (NA) microscope objectives are used, and super-resolution is obtained by illuminating specimens at different angles, thus generating a larger synthetic NA. These features are pivotal in both cells and tissue slide analysis. FP provides phase contrast images by applying phase retrieval algorithms and the quantitative information can be used to describe and characterize stain-free cells and tissues. To make FP microscopy viable for clinical practice, the issues arising from misalignments of the optical system or the presence of scattering structures to be imaged should be considered though. These non-ideal recording conditions result in phase artefacts in the recovered highresolution FP reconstruction. To tackle this problem, we propose a blind multi-look FP (ML-FP) algorithm that directly minimizes the artefacts while ensuring correct phase retrieval and can be used by unskilled operators in clinical practice. Here, we show how ML-FP allows the analysis of cells seeded onto micropatterned substrates (for mechanobiology applications) and tissues (for physiology and diagnostic applications). In order to improve robustness in the presence of misalignments, we use the ML-FP outcome as a ground truth and train a GAN architecture to emulate the phase retrieval process. The GAN receives as input the complex amplitude at the first iteration of the FP phase retrieval algorithm and returns in real-time the high-resolution FP complex amplitude.