2018/08/10 by Michael Kellman, Emrah Bostan, Kellman, Michael R. +5 · 1 citation
Computer Science · Physics and Astronomy · #Advanced X-ray Imaging Techniques #Computer Vision and Pattern Recognition (cs.CV) #Digital Holography and Microscopy #FOS: Computer and information sciences #FOS: Electrical engineering #Optical measurement and interference techniques #Signal Processing (eess.SP) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1808.03571
openalex publication_date 2018/08/10 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Coded-illumination can enable quantitative phase microscopy of transparent\nsamples with minimal hardware requirements. Intensity images are captured with\ndifferent source patterns and a non-linear phase retrieval optimization\nreconstructs the image. The non-linear nature of the processing makes\noptimizing the illumination pattern designs complicated. Traditional techniques\nfor experimental design (e.g. condition number optimization, spectral analysis)\nconsider only linear measurement formation models and linear reconstructions.\nDeep neural networks (DNNs) can efficiently represent the non-linear process\nand can be optimized over via training in an end-to-end framework. However,\nDNNs typically require a large amount of training examples and parameters to\nproperly learn the phase retrieval process, without making use of the known\nphysical models. Here, we aim to use both our knowledge of the physics and the\npower of machine learning together. We develop a new data-driven approach to\noptimizing coded-illumination patterns for a LED array microscope for a given\nphase reconstruction algorithm. Our method incorporates both the physics of the\nmeasurement scheme and the non-linearity of the reconstruction algorithm into\nthe design problem. This enables efficient parameterization, which allows us to\nuse only a small number of training examples to learn designs that generalize\nwell in the experimental setting without retraining. We show experimental\nresults for both a well-characterized phase target and mouse fibroblast cells\nusing coded-illumination patterns optimized for a sparsity-based phase\nreconstruction algorithm. Our learned design results using 2 measurements\ndemonstrate similar accuracy to Fourier Ptychography with 69 measurements.\n