A deep learning approach to high space-bandwidth product phase microscopy with coded illumination (Conference Presentation)

We investigate quantitative phase imaging techniques based on oblique illumination including differential phase contrast microscopy (DPC) and Fourier Ptychography Microscopy (FPM). DPC uses partially coherent, asymmetric illumination to achieve 2X resolution improvement but has small field of view (FOV). FPM achieves both wide FOV and high resolution but requires a large number of measurements. Achieving high space-bandwidth product (SBP) imaging in real-time remains challenging. Our goal is to develop a data-driven approach to enable highly multiplexed illumination to substantially improve the acquisition speed for high-SBP quantitative phase imaging. To do so, we abandon the traditional sampling strategy and phase retrieval algorithms. Instead we design a convolutional neural network (CNN) that uses only 4 brightfield and 3 darkfield images under asymmetrically coded illuminations as input and predicts high-SBP phase images. Particularly, instead of restoring a deterministic image, our CNN predicts pixel-wise probability distributions (Laplace) that is characterized by the location and scale. The predicted location map corresponds to the desired high-resolution phase image; in addition, the scale map provides per-pixel confidence of the prediction. Additionally, we show the potential of transfer learning that with minor extra training, the CNN can be optimized for different cell types. Experimental results demonstrate that the proposed method is robust against experimental imperfections, e.g. aberrations, misalignment, and reconstructs high-SBP phase images with significantly reduced acquisition and processing times.