Viruses represent the most abundant biological entities on Earth, with global scientific estimates suggesting that approximately ten to the thirty-first power viral-like particles exist across the biosphere. This extraordinary figure exceeds a mole of atoms, surpasses the estimated number of stars in the observable universe, and dwarfs the total cellular count of the human body. However, assessing viral abundance and morphology directly in complex ecosystems presents persistent technical bottlenecks. To overcome these long-standing analytical barriers, researchers have introduced EpiVirQuant, an open-source computer vision framework designed to automate the enumeration and morphological sizing of viral-like particles in aquatic systems using epifluorescence microscopy images, computational image processing, and a novel tunable point-spread function for dynamic blind deconvolution.
Historically, transmission electron microscopy served as the pioneering technique for detecting and characterizing viral-like particles in environmental surveys. While transmission electron microscopy continues to provide high-resolution structural observations of viral capsids, it proved impractical for general-use ecological enumeration because the methodology is labor-intensive, demands expensive infrastructure, exhibits significant operational variability, and consistently underestimates actual viral abundance. Epifluorescence microscopy subsequently emerged as a more cost-effective alternative utilizing portable equipment, serving as the primary benchmark for viral quantification across aquatic environments, soils, and sediments for forty-seven years. Throughout this extended history, epifluorescence protocols incorporated diverse fluorophores, including 4′,6-diamidino-2-phenylindole, Yo-Pro, SYBR Green I, and SYBR Gold. Despite offering high target specificity, straightforward protocols, and multi-fluorophore compatibility, epifluorescence microscopy remains burdened by significant operational challenges, specifically the requirement for manual counting by trained biologists, the complete absence of reliable morphological sizing, and the lateral diffraction limit of optical systems.
Standard optical microscopy is subject to an inescapable diffraction barrier that restricts lateral spatial resolution to approximately two hundred and fifty nanometers. When physical diffraction interacts with inherent optical defects in microscope lenses, the resulting imagery exhibits smearing, blurred circular spots, and obscured fine structural details. These optical distortions present severe impediments for viral studies because the physical dimensions of most viral-like particles fall well below this diffraction boundary. In optical physics, this blurring phenomenon is mathematically modeled as an underlying transformation that maps an ideal point source onto a distorted image plane through a point-spread function. Image deconvolution algorithms can computationally reconstruct the underlying uncorrupted image from blurred data only after an accurate point-spread function is specified, necessitating blind deconvolution methods.
To test their computational deconvolution method on authentic environmental samples, the investigators gathered aquatic specimens from Great Salt Lake at Antelope Island State Park near Layton, Utah. Water samples measuring five hundred milliliters were concentrated and filtered through zero-point-twenty-two micrometer polyvinylidene fluoride membranes. The retentate was fixed with electron-microscopy grade glutaraldehyde to a final concentration of two percent volume by volume, flash-frozen in liquid nitrogen, and preserved at minus eighty degrees Celsius. A working stock of SYBR Gold dye was prepared by centrifuging concentrated stain at thirty-five hundred times gravity for five minutes, diluting the supernatant at a one-to-ten ratio with autoclaved, membrane-filtered molecular-grade water, and passing the solution through a zero-point-twenty-two micrometer filter before storage at minus twenty degrees Celsius. Sample aliquots in nuclease-free microcentrifuge tubes were stained with four microliters of working dye, incubated on a dark eighty-degree hot block for fifteen minutes, supplemented with two microliters of ascorbic acid antifade, and mounted onto slides.
Slide preparation and optical calibration relied on spiked fluorescent microspheres from the Molecular Probes PS-Speck Microscope Point Source Kit. Exhibiting excitation and emission maxima at three hundred and sixty and four hundred and forty nanometers, these blue microspheres feature a certified diameter of one hundred and seventy-five plus or minus five nanometers, providing uniform, sub-diffraction calibration standards. Microscope slides cleaned with ethanol were treated with ten percent poly-lysine solution, dried at sixty degrees Celsius, and spiked with approximately three-point-five microliters of diluted microspheres. Imaging was conducted using an ECHO Revolve microscope equipped with a one-hundred-times oil immersion objective across two channels: a green fluorescein isothiocyanate channel at four hundred and ninety-five nanometers for SYBR Gold-stained viral-like particles, and a blue channel at three hundred and sixty nanometers for microspheres. Twenty-six paired image matrices of twenty-forty-eight by twenty-four-forty-eight pixels were acquired, with a horizontal scale bar establishing an initial conversion coefficient of thirty-four-point-two nanometers per pixel.
The computational architecture of EpiVirQuant operates through four sequential analytical stages, beginning with optimization box selection from a blue-channel calibration image. Coarse object segmentation is initially performed using Otsu’s thresholding algorithm, which divides pixel intensities into foreground and background classes by maximizing the ratio of between-class variance to within-class variance. A subsequent morphological erosion iteration using a disk structuring element of three-pixel radius resolves closely spaced doublet microspheres to ensure viable candidate selection. The algorithm measures object centroids and identifies candidate pairs separated by less than thirty pixels, or approximately ten hundred and twenty-six nanometers. In an exemplar calibration image, four pairs were detected, and the pair exhibiting a minimum centroid proximity of approximately four hundred and forty-nine nanometers was retained. Applying a fourteen-pixel padding coefficient established a bounding box of thirty-six by fifty pixels, yielding a maximum filter size of thirty-six pixels for rapid localized point-spread function derivation.
In the second stage, blind deconvolution is executed without prior equipment calibration by generating a tunable point-spread function. Rather than adopting standard static functions, the researchers formulated a tunable function grounded in a normalized cardinal sine expressed through Euler’s reflection formula using the gamma function. This mathematical expression avoids translational coordinate shifts that typically arise from point discontinuities at the origin. Parameterized by variables governing periodicity and vertical stretch, the function is exponentiated and swept across predefined filter sizes and parameter ranges. Deconvolution optimization employs a modified Richardson-Lucy algorithm assuming Poisson image noise, preserving total photon energy while maintaining non-negativity constraints across pixels. The parameter search is objectively guided by two complementary quantitative texture metrics: minimizing discrete Shannon information entropy to reduce statistical uncertainty, and maximizing gray-level co-occurrence matrix energy to enhance structural sharpness. The optimal parameter triplet corresponds to the absolute minimum difference between these two performance metrics.
Experimental analyses revealed that the choice of blue-channel calibration image significantly affects the derived point-spread function and subsequent size estimations. Across twenty-six calibration fields, six achieved mean microsphere sizes below two hundred nanometers, with image GSL-fifteen yielding the closest approximation to target values at one hundred and eighty-two-point-forty-nine plus or minus thirty-six-point-zero-one nanometers. Conversely, image GSL-twenty-three was discarded because its measurement variance exceeded the actual standard bead diameter. When benchmarked against alternative blind deconvolution operators, including five-by-five Gaussian functions and normalized uniform matrices of ones, the tunable point-spread function exhibited markedly superior convergence and numerical stability. The tunable function reached the target microsphere diameter of one hundred and seventy-five nanometers in one hundred and fourteen iterations, whereas Gaussian and uniform matrices suffered catastrophic artifact amplification, rapidly inflating detected objects to an artificial magnitude of one million while the tunable operator stabilized cleanly between twelve hundred and twenty and thirteen hundred and twenty-two objects.
Equipped with the optimized point-spread function, EpiVirQuant executes its third and fourth stages by establishing a global size-correction factor before segmenting environmental viral particles. In step three, blue-channel calibration images are deconvolved across Richardson-Lucy iterations, and segmented object axes are measured under Otsu thresholding. In an exemplar evaluation spanning fifteen hundred and sixty-three detected microspheres, the observed mean diameter was one hundred and ninety-six-point-thirty-seven plus or minus forty-six nanometers, establishing a size-correction coefficient of zero-point-eighty-eight-seventy-eight, representing an eleven-point-twenty-two percent reduction to match the true physical standard. In step four, this empirical scaling factor is applied to deconvolved green-channel viral images. Automated enumerations closely tracked expert manual counts, identifying twenty-four hundred and seventy-five viral-like particles compared to twenty-four hundred and seventy-two counted manually across twenty-six fields. Notably, the deconvolution resolved overlapping Airy-disk diffraction halos between closely spaced virions, preventing false-negative merger errors and significantly improving lateral morphological accuracy.
To eliminate non-viral artifacts and optical anomalies, the researchers incorporated an automated candidate removal step based on particle diameter and linear eccentricity. Furthermore, linear eccentricity filtering removed objects exceeding zero-point-nine-hundred-and-seventy-five, which correspond to rod-shaped bacteria, and artifacts with an eccentricity of zero, such as two-by-two pixel squares lacking geometric resolution. Following outlier removal, the final census settled at twenty-four hundred and sixty-six viral-like particles, with a mean diameter of one hundred and seventy-four-point-fifty-three nanometers, consistent with aquatic viral populations. Processing required approximately sixty to eighty-five seconds per image on sixteen Intel Xeon cores. The authors concluded that EpiVirQuant provides an effective proof of concept for high-throughput viral analysis, with future enhancements aimed at direct unseparated imaging of microbial mats, cellular sizing, and deep learning integration.
Subject of Research: Automated enumeration and sizing of viral-like particles in epifluorescence microscopy via blind deconvolution
Article Title: Resolving and quantifying viral-like particles via blind deconvolution
Article References: Figueroa III, J. L., Hollenack, S. M., Bellanger-Perry, M., Fulghum, B., Visscher, P. T., & White III, R. A. (2026). Resolving and quantifying viral-like particles via blind deconvolution. BMC Methods, 3(1), Article 10. https://doi.org/10.1186/s44330-026-00060-z
Image Credits: AI Generated
DOI: 10.1186/s44330-026-00060-z
Keywords: EpiVirQuant, viral-like particles, epifluorescence microscopy, blind deconvolution, point-spread function, Richardson-Lucy algorithm, virus quantification, microbial ecology, Resolving, quantifying, viral-like, particles
News Source: Kristina Jarvis. (November 13, 2026). Automated Deconvolution Framework Enhances Epifluorescence Microscopy to Enumerate Environmental Viruses. Scienmag.
