High Throughput End-to-End Design of Quantitative Phase Microscopy based on Customizable Nano-Optics

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

Imaging cytometry provides biomedically important morphological features of cells or tissues via high- speed acquisition hardware and fast image processing algorithms. Imaging cytometry has found applications ranging from rare cellular event detection to drug screening. Despite significant advances in faster and more multiplexed imaging sensors, the imaging cytometry’s throughput is often limited by the speed of electronic hardware. In order to improve throughput further, images can be acquired in an “optically compressed” form such that more information can be transferred beyond the existing electronic hardware bottleneck. We and other groups, have previously demonstrated sparsity-exploiting compressive imaging with random under-sampling 4. Recent advances in machine learning, however, suggest that superior optimized compressive imaging schemes may be derived by treating the front-end optics as part of the down-stream image processing pipeline in a neural network algorithm 5. Along similar lines, in the Wadduwage Lab we are currently developing a general learning- based microscopy technology, called differentiable microscopy (δμ). In δμ, we consider the front-end optics and the decompression algorithm as a differentiable auto-encoder, with the assumption that we can find a lower dimensional feature space, specialized for a class of images. This lower dimensional signal can be acquired with lower bandwidth detectors with the front-end optics working as an image compressor. A promising approach to implement optical compression for δμ, is through the use of optical diffractive networks (ODNs). ODNs have been demonstrated for: non-compressive imaging, all-optical image classification, and hybrid optical and electronic image classification. Most current ODN demonstrations, however, are for terahertz wavelength applications while most biomedical imaging tasks require visible light. For ODN to operate efficiently with visible light, optimal wavefront manipulation will require lithography with precision substantially better than visible wavelength. Implosion Fabrication (ImpFab), a low-cost, on-demand facile 3D fabrication technology with tens of nanometer resolution has recently been invented in the Boyden Lab. As a first demonstration of δμ relevant for biomedical applications, we will design and construct high throughput image cytometers with quantitative phase contrast that are powered by visible ODNs fabricated with ImpFab. We expect approximately two orders of magnitude voxel throughput improvement over the state-of-the-art.