WORKING DRAFT authorea.com/54156

# Photon counting imaging and centroiding with an electron-bombarded CCD using single molecule localisation software

Abstract

Photon event centroiding in photon counting imaging and single-molecule localisation in super-resolution fluorescence microscopy share many traits. Although photon event centroiding has traditionally been performed with simple single-iteration algorithms, we recently reported that iterative fitting algorithms originally developed for single-molecule localisation fluorescence microscopy work very well when applied to centroiding photon events imaged with an MCP-intensified CMOS camera. Here, we have applied these algorithms for centroiding of photon events from an electron-bombarded CCD (EBCCD). We find that centroiding algorithms based on iterative fitting of the photon events yield excellent results and allow fitting of overlapping photon events, a feature not reported before and an important aspect to facilitate an increased count rate and shorter acquisition times.

Keywords: Photon counting imaging, single-molecule localisation, electron-bombarded CCD

OCIS codes: (040.3780) Detectors: Low light level, (030.5260) Coherence and statistical optics: Photon counting, (100.6640) Image processing: Superresolution, (110.0180) Imaging systems: Microscopy, (170.2520) Medical optics and biotechnology: Fluorescence microscopy

# Introduction

The detection of single photons is a technique used in many fields of science and technology, including fluorescence microscopy and spectroscopy, bioluminescence studies, optical tomography, DNA sequencing, lidar, quantum information science and encryption, and optical communications both on earth and in space.(Hadfield 2009, Buller 2010, Eisaman 2011, Seitz 2011) Photon counting imaging is a well-established low light level imaging technique where an image is assembled from individually detected photons. In conventional photon counting imaging, photon events on the phosphor screen of a microchannel plate (MCP)-based image intensifier are imaged with a charge-coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) camera at high frame rates, and many frames are accumulated to build up an image.(Hirvonen 2014a) Photon counting imaging is also possible with electron-bombarded (EB) sensors, where single photoelectrons liberated from the photocathode are accelerated by a high voltage directly into the CCD or CMOS sensor (Barbier 2011) to produce a photon event.(Spring 1998) These are smaller and less bright than MCP-intensified photon events, due to a generally lower gain of electron-bombarded sensors, with a narrow, voltage-dependent pulse height distribution.(Hirvonen 2014) EBCCD or EBCMOS-based photon counting imaging avoids distortion of the image due to the coupling of the intensifier to the camera, and image lag due to the phosphor decay time, and there is no need for spectral matching of the camera sensitivity and the phosphor.

A characteristic feature of the photon counting imaging technique is the possibility of calculating the true position of a photon event that covers several pixels with subpixel accuracy - a processes termed centroiding.(Suhling 2002, Suhling 1999, Boksenberg 1985) The original centroiding algorithms, based on a simple center-of-mass calculation,(Boksenberg 1985) were developed for implementation in hardware. With the advent of more powerful computers, it became possible to implement increasingly complex centroiding algorithms based on fitting the photon event in software. However, some of the algorithms employed in photon counting imaging are still simple, one-iteration algorithms.(Postma 2011)

In the past decade, photoswitchable and photoactivatable fluorescent probes (Fernández-Suárez 2008) have allowed the same centroiding principle to be employed in circumventing the diffraction limit in fluorescence microscopy. Single-molecule localisation fluorescence microscopy techniques are based on the activation of a small subpopulation of the fluorescent proteins or fluorophores used to stain the sample. They are imaged and subsequently deactivated before the process is repeated with a different subset of fluorophores.(Betzig 2006, Rust 2006, Hess 2006) The centroid positions of the fluorescent probes are calculated in each frame, typically by fitting a three-dimensional Gaussian function to the fluorophore’s point spread function, and the final image is formed by summing many frames. Single-molecule localisation fluorescence microscopy is now a well-established technique, and much effort has been put into the development and optimisation of many different types of centroiding algorithms, including iterative fitting algorithms.(Small 2014)

We recently reported that single-molecule localisation algorithms produce excellent results when applied to centroiding single photon events imaged with an MCP-intensified CMOS camera.(Hirvonen 2015) Here, we extend this work and apply super-resolution software for centroiding photon events detected with an EBCCD camera. Moreover, multi-emitter fitting analysis was used for separating overlapping photon events, an important aspect not reported before, which allows an increased count rate and shorter acquisition times.

# Method

Photon counting imaging was performed with a dual mode cooled Hamamatsu C1790-13 EBCCD, with $$512\times512$$ pixels and $$24\times24$$ $$\mu$$m pixel size. The EBCCD was cooled to -15$$^\circ$$C, and HiPic 7.1.0 software was used for image acquisition with 10 $$\mu$$s exposure time and super-high amplifier gain. The EBCCD was attached to the output port of an inverted Nikon Eclipse TE2000-U microscope, as schematically illustrated in Fig \ref{fig1}a. For transmission imaging of a 1951 USAF resolution test chart (Fig \ref{fig1}b), the microscope was used with a 4$$\times$$ 0.13NA air objective (Nikon) and a halogen lamp. For epifluorescence imaging, a cell sample (FluoCells Prepared Slide #1, Molecular Probes) was excited with a pulsed 467 nm diode laser (Hamamatsu PLP-10) and imaged with a 100$$\times$$ 1.4NA oil objective (Nikon). The illumination intensity was adjusted such that single photon events could be observed (Fig \ref{fig1}c,d).

The frames containing single photon events were processed with ThunderSTORM (Ovesný 2014) superresolution imaging plug-in for ImageJ. Due to memory limitations, the USAF test chart data was processed in 6$$\times$$5,000 and the cell data in 3$$\times$$2,000 image stacks. The software first detects the events from the noise background, and an approximate localisation algorithm locates the center pixel of each event. A sub-pixel localisation algorithm then calculates the center of the events with greater resolution.

The software camera parameters were set to 80.0 nm pixel size and 36 photoelectrons per A/D count. The base level varied between image stacks due to fluctuations in the EBCCD temperature, and was set to the average minimum grey value for the image stack in the range of 100-140 A/D counts. A wavelet (b-spline) image filter was applied with order of 3 and scale of 2.0. For the approximate localisation of the events, the centroid of connected components method was used with a peak intensity threshold (PIT) of 2*std(Wave.F1) for the USAF test chart data, and a PIT of 1.5*std(Wave.F1) for cell data, with the watershed algorithm enabled for all data.

All sub-pixel localisation methods offered by ThunderSTORM (Maximum Likelihood (ML) and Least squares (LS) fitting with both Gaussian (G) and Integrated Gaussian (IG) point-spread function (PSF), Centroid of local neighborhood, Radial Symmetry) were tested, with fitting parameters optimised for maximum photon count and minimum likelihood of false photon event recognition.

For the best results, ML fitting was used with a Gaussian PSF, with standard deviation (SD) set to 1.0 pixels. For fast processing with adequate results the PSF fitting radius was set to 2 pixels, and for optimal photon detection and separation of overlapping events the radius was set to 7 pixels. Multiple-emitter fitting analysis (MFA) was tested with a maximum of 2 molecules per fitting region with a model selection threshold (p-value) of 10$$^{-6}$$. When MFA was enabled, ThunderSTORM’s “remove duplicates” post-processing tool was applied with a distance threshold of 160 nm, and “intensity$$>$$4000” filter was applied to the USAF data and “intensity$$>$$3000” filter to the cell data.

Results are also shown for LS fitting method with an IG PSF with a 3 pixel fitting radius and 1.6 pixel SD, and a radial symmetry localisation method with 2 pixel estimation radius.