Improved deconvolution of very weak confocal signals

Introduction

Deconvolution is an established method for sharpening fluorescence images and removing background noise (Biggs, 2010; Sage et al., 2017). The usual input to a deconvolution algorithm is a Z-stack of optical sections generated by widefield or confocal microscopy. Because the benefits of deconvolution are fully realized when the signals are strong, the creators of deconvolution software recommend capturing a large number of photons while keeping the pixel sizes and Z-step intervals relatively small.

Those conditions are hard to meet with live cell imaging if Z-stacks are being collected at regular intervals to create a 3D time-lapse (4D) data set (De Mey et al., 2008). Intracellular structures are dynamic, so the images need to be taken rapidly. Moreover, the number of captured photons is severely constrained by the need to avoid photodamage to the cells and fluorophores (Carlton et al., 2010; Pawley, 2006). Such issues are prominent in our 4D confocal microscopy studies of secretory compartments in yeast cells (Bevis et al., 2002; Losev et al., 2006; Papanikou et al., 2015). We maximize the scan speed, minimize the intensities of the excitation lasers, and set the pixel sizes and Z-step intervals at the traditionally defined Nyquist limit to achieve a tolerable light exposure while ensuring accurate representation of the imaged structures (Day et al., 2016; Pawley, 2006). The resulting data sets typically comprise thousands of optical sections and have a low signal-to-noise ratio (SNR).

Even though the sampling characteristics of our 4D data are not ideal for deconvolution, the Huygens deconvolution software from Scientific Volume Imaging (SVI) can facilitate the analysis. A number of other freeware and commercial software packages are also available for deconvolution (Biggs, 2010; Sage et al., 2017), but in our experience, those programs are unsuitable for processing of multi-channel 4D confocal data due to some combination of cumbersome user interfaces, lack of compatibility with relevant file formats, and inadequate noise removal. Huygens is unique in that it readily deconvolves our data sets (Day et al., 2016; Papanikou et al., 2015). This software is widely used in the cell biology research community. Importantly, in addition to removing noise, Huygens smooths the uneven shapes and intensities obtained with low-SNR data to generate images that are easy to view and quantify.

Huygens works well for certain low-SNR fluorescence images, but when the fluorescence signals are very weak, Huygens may perform poorly (Arigovindan et al., 2013). We encountered this problem when imaging low-abundance proteins associated with yeast organelles. In such experiments, a pixel in the signal-containing portion of a confocal section may capture as few as 1–2 photons. To enable deconvolution of images with a very low SNR, Agard and colleagues developed deconvolution software called ER-Decon, which employs a novel regularization method tailored to fluorescence data (Arigovindan et al., 2013). However, ER-Decon has incompletely defined parameters, and it proved to be challenging to use. We therefore sought a method for processing very weak fluorescence signals with Huygens.

Methods

Results and discussion

We generated two small 4D data sets to illustrate confocal imaging of organelles in live Saccharomyces cerevisiae cells. The parameters were adjusted to capture either weak signals using a line accumulation of 8x, where each line in the image was scanned eight times and the results were summed (Video S1), or very weak signals using a line accumulation of 1x, where each line in the image was scanned only once (Video S2). Projections of representative Z-stacks from the two movies are shown in Figure 1. The organelles were dynamic, so the labeling patterns in the two movies are not identical, but the movies were brief and were captured sequentially, so the labeling patterns are similar enough to allow for comparison. With 8x line accumulation, the raw projections are noisy and display fluorescent structures with uneven shapes and intensities (Video S1 and Figure 1A). Deconvolution with Huygens efficiently removed the background noise and smoothed the structures. With 1x line accumulation, the data quality is even lower, but fluorescent structures can still be discerned in the raw projections (Video S2 and Figure 1B). In this case, deconvolution with Huygens erased almost all of the fluorescent structures. The processing employed standard settings in the Huygens software, including deconvolution with the Classic Maximum Likelihood Estimation algorithm. Although a larger percentage of the fluorescent structures in very weak data sets could be preserved by greatly reducing the number of deconvolution iterations or by using different SNR or background settings, the preserved structures often had distorted shapes (not shown). Similar loss or poor preservation of very weak fluorescent structures was seen with the Good’s roughness Maximum Likelihood Estimation algorithm, which is recommended for use with noisy confocal data (not shown). Based on these observations, we have continued to use standard settings in Huygens. Our data sets often lie between the two extremes depicted in Figure 1, and when movies are generated after deconvolution, the fluorescent structures blink because a given structure is erased in some movie frames but not in others (see Video S2). Such movies cannot be productively analyzed.

Figure 1. Improved deconvolution of 4D live cell data with a Gaussian blur prefilter.

Gene replacement in Saccharomyces cerevisiae was used to label late Golgi cisternae with Sec7-mCherry (red) and prevacuolar endosomes with Vps8-GFP (green) (Papanikou et al., 2015). Cells were imaged by 4D confocal microscopy. In consecutive movies, line accumulation was set to (A) 8x or (B) 1x. The data were average projected either with no processing, or after deconvolution with Huygens, or after prefiltering with a 2D Gaussian blur using a radius of 0.75 pixels followed by deconvolution. Fluorescence data are superimposed on differential interference contrast images of the cells (blue). Shown are representative frames from Video 1 (8x) and Video 2 (1x). The fluorescence patterns in (A) and (B) are similar, but not identical because the labeled structures changed during the interval between the two movies. Scale bar, 2 µm.

Figure 2. Improved deconvolution of non-punctate fluorescence signals with a Gaussian blur prefilter.

Gene replacement in Saccharomyces cerevisiae was used to label ER membranes with Hmg1-GFP (Koning et al., 1996). A confocal Z-stack was captured with line accumulation set to (A) 8x or (B) 1x. The data were average projected either with no processing, or after deconvolution with Huygens, or after prefiltering with a 2D Gaussian blur using a radius of 0.75 pixels followed by deconvolution. Fluorescence data are superimposed on differential interference contrast images of the cells (gray). Scale bar, 2 µm.

Figure 3. Improved deconvolution of simulated data with a Gaussian blur prefilter.

Simulated confocal Z-stacks of fluorescent point sources were created as described in Methods, either (A) without background noise or (B) with background noise. The data were processed and average projected as in Figure 1.

Figure 4. Effect of a Gaussian blur prefilter on deconvolution of widefield fluorescence data.

These images of fluorescent yeast Zip1 filaments correspond to Figure 4 of Arigovindan et al. (2013). The two exposure levels represent strong (100%) or weak (0.25%) signals, respectively. Where indicated, the data were subjected either to a Gaussian blur with a radius of 1.00 pixel, or to deconvolution with Huygens, or to a Gaussian blur prefilter followed by deconvolution. The theoretical point spread function was based on imaging parameters supplied with ER-Decon.

In the course of testing several types and combinations of image filters (Day et al., 2016), we discovered that for very weak confocal signals, the key step was to prefilter the optical sections in ImageJ with a Gaussian blur. That prefilter dramatically improved the results obtained after deconvolution (Video S2 and Figure 1B). Fluorescent structures were no longer erased, and instead were preserved and smoothed while the background noise was largely eliminated. Most of the structures visualized by this method were biologically relevant because they persisted between movie frames (Video S2). Essentially identical results were obtained with 2D and 3D Gaussian blurs (not shown), so we use a 2D Gaussian blur because the processing is faster. This prefiltering step enables us to generate useful 4D movies from data sets that contain very weak confocal signals.

Application of the Gaussian blur prefilter requires the data to be in a suitable format. Our images are collected with a high-sensitivity detector in photon counting mode, and the pixel values are in 8-bit format. For very weak signals, typical pixel values are 0, 1, or 2 because a pixel rarely captures more than 2 photons. To obtain a meaningful blur, the numbers are scaled up to allow for intermediate integer values. We convert the images to 16-bit format and multiply by 256, resulting in typical pixel values of 0, 256, and 512. A Gaussian blur then generates a range of intermediate values, effectively spreading the individual photon signals over multiple pixels.

An important question is how to determine whether a confocal data set is suitable for processing with a Gaussian blur prefilter. Ideally, this prefilter would be used routinely, because even if the average signal intensity is strong, some structures may have very weak signals. The concern with routine application of a Gaussian blur prefilter is that blurring might be propagated to the deconvolved images. Indeed, when the Gaussian blur prefilter was applied to signals strong enough to be preserved during normal deconvolution, we saw some blurring of the fluorescent structures (Video S1 and Figure 1A). However, this effect was minor with suitable parameters for the prefilter (see below). Our results indicate that a Gaussian blur prefilter can be used to image structures with a range of signal intensities, resulting in preservation of very weak signals without significant degradation of stronger signals.

Because the labeled structures in our 4D data sets were punctate, we tested whether a Gaussian blur prefilter would also improve deconvolution of other shapes. For this purpose, GFP was fused to a yeast endoplasmic reticulum (ER) protein that localizes mainly to the nuclear envelope (Koning et al., 1996). A single confocal Z-stack was captured at a low excitation laser setting. As shown in Figure 2A, which employed 8x line accumulation, the labeled protein appeared as prominent nuclear envelope rings with weaker labeling of peripheral ER membranes. Deconvolution of the raw data preserved the nuclear envelope rings. When a Gaussian blur was applied before deconvolution, additional signals outside the nuclear envelope were preserved. Figure 2B shows a parallel analysis with 1x line accumulation. In this case, the fluorescence signals were completely erased by deconvolution of the raw data, but application of a Gaussian blur before deconvolution preserved the nuclear envelope rings. We conclude that for various types of fluorescence patterns, a Gaussian blur prefilter preserves very weak confocal signals during deconvolution with Huygens.

To explore the Gaussian blur effect systematically, and to confirm that it was not limited to the particular configuration of our confocal microscopy setup, we used simulated data. 3D confocal imaging was simulated for an array of 64 faintly fluorescent point-like objects, each of which was represented by about 10–25 photons spread over multiple optical sections. Figure 3A shows projections of this simulated Z-stack before and after processing. After deconvolution with Huygens, only 7 objects were preserved, but after a Gaussian blur prefilter followed by deconvolution, all 64 objects were preserved. The total signal intensities for the objects were largely unchanged after either a Gaussian blur alone or a Gaussian blur followed by deconvolution (Supplementary Figure 1). A setting of 0.75 pixels for the radius (sigma) parameter of the prefilter preserved signals while causing very little blur in the final images, and similar results were obtained with a radius of 1.00 pixels (Supplementary Figure 2). We find empirically that radius values of 0.75 – 1.00 pixels work well for both real and simulated fluorescence data obtained under a variety of imaging conditions. When the simulation was repeated with added background noise, a Gaussian blur prefilter followed by deconvolution removed most of the noise while preserving all of the objects (Figure 3B). In this case, deconvolution in the absence of a prefilter completely erased the objects. The voxels in those simulations were 80x80x250 nm to mimic traditional Nyquist imaging with our confocal system (Pawley, 2006), but similar results were obtained with voxels of 40x40x120 nm (Supplementary Figure 3) chosen to meet the more stringent Nyquist criteria recommended by SVI. Thus, a Gaussian blur prefilter preserves weak confocal signals during deconvolution under multiple real and simulated conditions.

The paper describing the ER-Decon software showed that Huygens could give unsatisfactory results with low-SNR widefield images (Arigovindan et al., 2013). We processed low-SNR widefield microscopy data from that study with a Gaussian blur prefilter before deconvolution. The improvement was only moderate because Huygens did not erase the structures, but when the signal was weak, the prefilter did increase contrast between labeled structures and the background (Figure 4, right panels), yielding results similar to those obtained with ER-Decon (compare to Figure 4 in Arigovindan et al., 2013). The combined observations demonstrate that a Gaussian blur prefilter consistently improves deconvolution of low-SNR fluorescence data.

The reason for this beneficial effect of the prefilter is not fully understood. Gaussian blurs suppress high-frequency noise. That approach reduces pixel-to-pixel intensity variations, and it can facilitate analysis methods such as edge detection and particle tracking (Cheezum et al., 2001; Russ & Neal, 2015). A different mechanism presumably underlies the ability of a Gaussian blur to prevent loss of very weak signals during deconvolution. Huygens apparently “expects” a gradually varying distribution of the signal intensities within a set of nearby voxels, and the Gaussian blur prefilter converts the data to a form suitable for the Huygens algorithm.

Is a Gaussian blur prefilter before deconvolution an acceptable procedure? Processing of images before deconvolution is not generally recommended, but a Gaussian blur is relatively safe. This filter causes a simple and well behaved transformation of the data, and it preserves the total intensity of a fluorescent structure (Burger & Burge, 2008). Gaussian blurs have previously been employed during deconvolution to suppress noise buildup (Agard et al., 1989). A Gaussian blur prefilter was actually proposed by the founder of SVI as a method that can reduce noise sensitivity during deconvolution of confocal data (van Kempen et al., 1997). Therefore, it seems reasonable to apply this prefilter to very weak confocal signals for the novel purpose of avoiding complete erasure of biologically meaningful structures. When the signals are stronger, the Gaussian blur prefilter has a barely detectable effect on the final image, so there seems to be little risk in applying this prefilter routinely.

It could be argued that the Gaussian blur prefilter merely sidesteps a software flaw, in which case a better option would be to fix the Huygens algorithm. However, Huygens is optimized for processing images that exceed a minimum signal strength, and our confocal data sometimes fall below this threshold. Other deconvolution algorithms may perform differently. The available evidence specifically shows that the Gaussian blur prefilter is useful with Huygens. This straightforward method allows us to take advantage of the flexibility, noise removal capability, and smoothing properties of the Huygens software to process very weak fluorescence signals.

Data availability

Dataset 1: TIFF files for the experimental and simulated image data are provided in the compressed folder Original Image Files.zip. The following files are included: 4D_movie_1x.tif and 4D_movie_8x.tif are the 4D confocal data sets used for Figure 1 and Supplementary Figure 1, and for Video S1 and Video S2; Hmg1_1x.tif and Hmg1_8x.tif are the confocal image stacks used for Figure 2; simulation_80x80x250.tif is the simulated confocal image stack used for Figure 3A and Supplementary Figure 2; simulation_80x80x250_plus_noise.tif is the simulated confocal image stack used for Figure 3B; simulation_40x40x120.tif is the simulated confocal image stack used for Supplementary Figure 3; and Zip1_0.25%.tif and Zip1_100%.tif are the widefield image stacks used for Figure 4. doi, 10.5256/f1000research.11773.d163336 (Day et al., 2017).