Optical high-throughput microscopy with single-pixel detectors

Abstract

Optical high-throughput microscopy promises to deliver rich information on biological cells in large populations (≥ 1000 cells/second) for science and bioengineering studies. Recently, single-pixel imaging techniques have attracted widespread attention for their ultrafast pixel readout rate — tens of thousands to hundreds of thousands of different cellular behaviors can be observed and monitored in a fraction of the time needed by conventional technologies. In particular, optical time-stretch imaging offers high signal transmission bandwidth of the optical fiber and high optical power with the fiber-based amplifier, beating the fastest camera-based technologies. However, ultrahigh speed imaging inherently generates an enormous stream of data, inflating the budget for the digitizers. Even with adequate funding, the technique is restricted to intensity and phase imaging at micrometer resolution, as it requires both illumination at infrared wavelength and detection of a coherent light-field. Popular cell monitoring approaches in life sciences, such as fluorescence staining and subcellular molecule tracking, are yet to adopt this novel technology. This thesis aims to extend the “depth” and “breadth” of single-pixel high-throughput microscopy technology, i.e. to relax the digitizer bandwidth constraint without sacrificing image resolution, and to harness the technology in various imaging modalities. First, the pixel super-resolution method is applied to lower the digitizer bandwidth. With active beam scanning control, it is experimentally demonstrated that the pixel readout rate can be lowered to 20 GPixel/s. In the presence of significant pixel drifting due to asynchronous sampling, the bandwidth requirement is relaxed further to 5 GPixel/s without any active scanning control. Next, the feasibility of ultrafast microscopy at a limited photon budget — a typical property of a fluorescence-stained specimen — is discussed. A pseudo-classical model is constructed to predict the speed-resolution-sensitivity tradeoff at ultrafast raster-scanning of the laser beam. It is discovered that the laser sweep rate is limited to approximately 100 kHz due to severe Poisson noise in the detected signal. Even with sufficiently high illumination flux at a 10 MHz sweep rate, the image resolution can degrade to approximately 10 μm, as the temporal fluorescence decay is coupled to the point spread function of the system. Equipped with the knowledge of the photon budget constraint in ultrafast raster-scanning systems, alternative single-pixel imaging approaches are explored. Structured illumination with a two-dimensional spectral disperser is proposed as the enabling technique for the non-scanning imaging at nanosecond shutter speed — tens to hundreds of optically integrated measurements can be captured to reconstruct images at the effective frame rate beyond 10 kHz. The performance characteristics of the spectral disperser, especially spatial distortion and astigmatism, are thoroughly evaluated to find a set of optimal designs. Signal reduction of non-scanning high-throughput microscopy is also explored by applying the theory of compressive sensing. Taking advantage of data redundancy, it is shown that 95% of single-pixel measurements can be discarded on-the-fly in frequency-multiplexed fluorescence imaging systems without losing image quality. The subsampling concept is also applied to scanning holographic imaging, in which the acquisition speed of the three-dimensional object can be increased up to 25× with a low-density non-raster scanning trajectory.