Fluorescence Microscopy

A fluorescence microscope images the light that labelled molecules emit, not the light that the specimen absorbs or delays. The specimen is lit at one band of wavelengths, its fluorophores emit at longer wavelengths, and filters pass only that emission to the detector, so the labelled structure appears bright on a dark background and every other structure is invisible. The optics are those of the optical microscope, with the same numerical aperture, diffraction-limited point-spread function (PSF) and sampling rules. What changes for image analysis is the signal. The light is so faint that it is counted in photons, so the images are dominated by Poisson noise; the labels fade under illumination and can harm living cells; colours overlap from one channel into the next; and blur is often undone numerically by deconvolution. Bioimage analysis surveys the modalities, labels and analysis tasks in outline; this article goes deeper into the sensor.

How It Works

Fluorescence and fluorophores

Fluorescence is two events in sequence. A molecule, the fluorophore, absorbs a photon and is raised to an excited electronic state; it then relaxes to the ground state and emits a photon (Olarte et al., 2018). Before the emitting transition the molecule usually passes through other, mostly vibrational, energy levels, so the emitted photon has less energy and a longer wavelength than the absorbed one. This is the Stokes shift, and the energy difference stays in the sample, where it can cause photodamage through heat (Olarte et al., 2018). The quantum yield $\Phi$ is the fraction of excited fluorophores that return to the ground state by emitting a photon. In the linear regime, Olarte et al. write the number of photons emitted per second along an interaction length $l$ of the excitation beam as

$$\Theta_{\mathrm{em}} = \Phi\, \delta_{1p}\, N_0\, \frac{P}{h\nu}\, l ,$$

where $\delta_{1p}$ is the one-photon absorption cross section, $N_0$ the fluorophore density, $P$ the excitation power and $h\nu$ the energy of an excitation photon. The expression holds only well below saturation, when the number of emitted photons is much smaller than the number of fluorophores (Olarte et al., 2018). Emission is therefore proportional to both the excitation power and the number of fluorophores, which is what makes fluorescence intensities usable as measurements.

The labels themselves, antibodies, fluorescent proteins and DNA stains, are described in Bioimage analysis. Whatever the label, the image tells no more than how much light was emitted from each point, and the working assumption is that twice as many photons from one region mean around twice as many fluorescing molecules (Bankhead, 2022–2025). Most of the emitted light is never recorded. Photons leave in all directions, only a proportion enters the objective, and the detector misses some of those, so perhaps as much as 90% of the light is lost; because a similar proportion is lost everywhere, relative brightness is still preserved (Bankhead, 2022–2025).

Filters and channels

The microscope separates the weak emission from the strong excitation light by wavelength. In the scheme that Mualla et al. (2018, Figure 5.6(b)) draw, the light passes an excitation filter, a dichroic mirror and an emission filter: the excitation filter selects the part of the spectrum that excites the fluorophores, and the emission filter separates the emitted light from the excitation light. The result has high contrast compared with brightfield images (Mualla et al., 2018).

Each combination of excitation band and emission filter defines a channel, and a multichannel fluorescence image is a stack of such channels, each recorded from light filtered by wavelength (Bankhead, 2022–2025). Display colours are a choice of lookup table, not part of the data, and Bankhead advises against converting a merged display to RGB before analysis, which loses the raw values. Whole-slide scanners digitize fluorescent slides as tiles whose channels are stored as separate bands rather than as an RGB colour image, as whole-slide imaging describes. Mualla et al. (2018) point out two limits of the contrast itself: the dye can alter the specimen, and the image shows where the dye is, which need not cover the whole object, so shape information can be incomplete.

Widefield, confocal, light-sheet and other variants

The variants differ in which part of the specimen is excited and how the emitted light is collected, and those choices set the photon count, the out-of-focus background and the light dose.

• Widefield (epifluorescence). The whole field is illuminated at once and the camera records all pixels in parallel (Bankhead, 2022–2025). Images collect many photons even at high frame rates, which overcomes read noise, but excitation passes through the whole depth of the specimen, so in a thick specimen every plane contributes out-of-focus light and there is no optical sectioning (Bankhead, 2022–2025; Olarte et al., 2018).
• Laser scanning confocal. A focused spot is scanned point by point and a pinhole in a plane conjugate to the focus blocks light from other planes (Cremer & Masters, 2013; Olarte et al., 2018). The image is built one pixel at a time on a point detector such as a photomultiplier tube (PMT), so each pixel receives light for only a tiny fraction of the frame time, and confocal images tend to be slow, noisy or both (Bankhead, 2022–2025). A pinhole of about one Airy disc balances in-focus signal against sectioning; because the Airy disc scales with wavelength, one pinhole setting corresponds to different sizes in different channels (Bankhead, 2022–2025). Cremer and Masters (2013) give typical confocal resolution as a full width at half maximum of 200 nm laterally and 600 nm axially, with a lateral improvement over widefield of at most a factor of 1.4; the main gain is the optical section. Excitation still passes through the whole specimen along the beam, so bleaching and phototoxicity occur outside the plane that is recorded (Olarte et al., 2018).
• Spinning-disk confocal. Many pinholes in a rotating disk excite and detect well-separated points simultaneously onto a camera. Images are likely to be less sharp than with a laser scanning confocal, because some light scatters through other pinholes, and the pinhole size cannot be adjusted, but acceptable signal-to-noise ratios come at much higher frame rates (Bankhead, 2022–2025).
• Multiphoton. Excitation needs the simultaneous absorption of several longer-wavelength photons, which happens only where the intensity is highest, at the focus, so no pinhole is needed, less of the sample is damaged outside the focal plane, and imaging reaches several hundred micrometres deep (Bankhead, 2022–2025).
• Total internal reflection fluorescence (TIRF). An evanescent wave excites only the fluorophores right at the coverslip, recorded on a camera as in widefield; Bankhead (2022–2025) suits it to seeing to a depth of about 100 nm, while the source cited in Bioimage analysis gives a few hundred nanometres.
• Light-sheet. A thin sheet illuminates only the plane imaged by a perpendicular widefield detection arm, so the camera image is an optical section without a pinhole, fluorophores outside the sheet do not bleach, and the low light dose allows imaging for up to several days (Olarte et al., 2018). Absorption and scattering along the sheet produce stripes and shadows that hinder segmentation; double-sided illumination, pivoting the sheet and multiview fusion reduce them (Olarte et al., 2018).

Detectors follow the scan mode. Bankhead (2022–2025) pairs widefield with cameras, laser scanning with PMTs, whose quantum efficiency can be as low as 10–15%, and spinning disk with cameras for which an electron-multiplying CCD is often preferable, because spinning-disk images collect fewer photons and read noise would otherwise dominate; optical microscopy and image sensor cover the camera types themselves.

Data It Produces

A fluorescence acquisition produces the multidimensional arrays of channels, z-planes and time points that Bioimage analysis describes; light-sheet experiments add multiple views of the same specimen (Olarte et al., 2018).

The pixel values are not photon counts. They are proportional to the detected charge, often with an offset added, and the offset has to be estimated, from a background region or an image acquired without excitation, and subtracted before intensities from different images are compared under identical acquisition settings (Bankhead, 2022–2025). Sage et al. (2017) state the standard image-formation model: the fluorophore distribution $x$ is convolved with the PSF $h$ and degraded by noise,

$$y = Hx + n ,$$

where $Hx$ is the discretized convolution of $x$ with the PSF. The noise has a signal-dependent part, the Poisson-distributed number of photons arriving at a pixel, and a part usually modelled as additive Gaussian, which collects background, read-out and quantization noise (Sage et al., 2017). Under Poisson statistics $n$ depends on the data, which the additive notation hides. The model treats the microscope as shift-invariant, characterized by a single PSF; when the PSF varies across the field it no longer holds (Sage et al., 2017).

Noise

Bankhead’s (2022–2025) one-line lesson for the working microscopist is that reducing noise means detecting more photons. Photon emission is random, so even a perfect detector records Poisson-distributed counts, whose signal-to-noise ratio at an expected count $\lambda$ is $\sqrt{\lambda}$, and read noise dominates when few photons arrive; image noise gives the combined Poisson–Gaussian model, its signal-to-noise ratio and how its parameters are estimated. What this means for fluorescence intensities is shown by Bankhead’s example: true values of 4 and 8 photons give one-standard-deviation ranges that overlap (about 2–6 and 5–11), so a twofold difference cannot be told apart, while 100 and 200 photons do not overlap (Bankhead, 2022–2025).

Two consequences matter for analysis. First, because the noise depends on the local brightness, a structure of fixed brightness is easier to detect on a dark background than on a bright one, so counts of small structures from bright and dark regions of the same image are not directly comparable (Bankhead, 2022–2025). Second, isolated bright pixels in a dark image are expected from noise alone: with a mean of one photon per pixel, a 512 × 512 image should contain about 800 pixels with a value of 5 (Bankhead, 2022–2025). A real structure covers at least the area of the PSF, so single-pixel outliers can be removed when the pixels are smaller than the PSF (Bankhead, 2022–2025).

Scientific CMOS (sCMOS) cameras, whose pixels differ in offset, variance and gain as optical microscopy describes, appear to flicker even without light, and Mandracchia et al. (2020) note that the assumption of white noise fails at low photon counts for these sensors. In our reading, a noise model with one gain and one read-noise level for the whole image then misdescribes the darkest pixels; Mandracchia et al. review camera calibrations that estimate each pixel’s response, and propose a correction method of their own.

Photobleaching and Phototoxicity

Fluorophores do not emit forever. Under illumination, each fluorophore has a certain probability of irreversibly losing its ability to fluoresce, so the emitted fluorescence decreases over a time-lapse even when the amount of labelled molecule does not change (Miura, 2020). This photobleaching degrades visual quality, distorts measurements of molecular kinetics and degrades the segmentation of target objects. Miura lists the remedies before acquisition: the choice of fluorophore, anti-fading reagents, lower laser power, higher detector gain and longer intervals between frames. After acquisition, the loss can be compensated by bleach correction, for which Miura’s ImageJ plugin offers three algorithms:

• Simple ratio. Each frame, after background subtraction, is multiplied by the ratio of the first frame’s mean intensity to its own. The result depends heavily on the background estimate: on Miura’s sample, a yeast time-lapse with a measured background of about 68, background values of 64 and 72 turned the corrected curve into a decreasing and an increasing trend.
• Exponential fitting. A single exponential is fitted to the mean intensity over time and its inverse is applied. Miura notes that bleaching sometimes follows a double exponential, so the model must be justified or its fit checked. Because the fitted curve ignores short fluctuations, it preserves them in the corrected data; for the same reason, provided bleaching is known to be single-exponential, it does not mistake a real increase in fluorescence, such as the synthesis of new fluorescent protein, for an error to be corrected, as the simple ratio does.
• Histogram matching. Each frame’s histogram is matched to the first frame’s, which needs no background estimate and gave Miura’s most stable corrected curve, but assumes the distribution of fluorescence stays the same; when signal changes from diffuse to punctate, for example, it forces a wrong correction.

Correction restores the mean level, not the information: a corrected late frame is a rescaled low-count image with the noise of a low-count image, as the example below shows.

The light that excites fluorophores can also harm the specimen. Wäldchen et al. (2015) explain that high irradiation generates reactive oxygen species through excited-state reactions of endogenous and introduced chromophores, which can damage cellular components. Irradiating cell lines under conditions typical of single-molecule localization microscopy and following their survival for 20–24 hours, they found that sensitivity rose sharply at shorter wavelengths: cells tolerated about 1 kW cm⁻² at 640 nm for several minutes, but the maximum dose at 405 nm was only about 50 J cm⁻². Transfection alone increased sensitivity even without fluorophores. Citing earlier work in which yeast cells that looked healthy after low-dose irradiation failed to divide overnight, they conclude that a cell’s appearance directly after irradiation is not a meaningful marker of photodamage (Wäldchen et al., 2015). Olarte et al. (2018) summarize Stelzer’s argument from the solar flux at the Earth’s surface, below 1.4 kW m⁻², that irradiance on a living specimen should not exceed about 1 nW µm⁻², or 100 mW cm⁻². Wäldchen et al. speak of the limited photon budget of live-cell experiments; the trade-off it imposes between resolution, speed, light exposure and depth, and how learned restoration exploits it, is described in Bioimage analysis.

Spectral Overlap and Bleed-Through

Fluorophore emission spectra are broad and overlap, so light from one fluorophore is also recorded in the channel meant for another. McRae et al. (2019) call this bleed-through or cross-talk, note that it can produce false-positive detections, and describe it as almost unavoidable with three or more fluorophores; it complicates colocalization analysis and any quantitative comparison between channels.

The standard correction is linear unmixing. Each pixel’s measured signal across channels is modelled as a linear combination of the reference spectra of the fluorophores present, and the fitted weights are proportional to the fluorophores’ concentrations (McRae et al., 2019). Writing it in matrix form, our notation rather than theirs, the vector of channel intensities $\mathbf{y}$ at a pixel is $\mathbf{y} = S\,\mathbf{a}$, where the columns of $S$ are the fluorophores’ spectra as seen through each channel and $\mathbf{a}$ holds their abundances, estimated by least squares. McRae et al. list the limitations: the method needs at least as many detection channels as fluorophores and prior knowledge of the reference spectra, which are instrument-specific and hard to measure; spectra can change in the tissue environment; and superposition can fail under quenching, photobleaching and two-photon absorption. Non-negative matrix factorization drops the need for known spectra but can return several equally valid, different solutions, and McRae et al. propose a k-means clustering method that, assuming each labelled structure is discrete so that every pixel belongs to one fluorophore group, learns spectral signatures from the mixed image itself and can also remove background and autofluorescence.

A related source of channel-to-channel error is chromatic aberration, which shifts focus with wavelength and misaligns channels, as optical microscopy notes.

Deconvolution

Deconvolution inverts the blur numerically: given the image $y$, the PSF and a noise model, it estimates the fluorophore distribution $x$, instead of improving the PSF optically as confocal, multiphoton and most super-resolution methods do (Sage et al., 2017). Its indirect benefits, contrast enhancement, noise reduction and use at lower light levels, reduce bleaching and phototoxicity; badly tuned, it produces artifacts such as noise amplification, ringing and aliasing that can mislead interpretation (Sage et al., 2017). Deconvolution of micrographs goes back to Agard and Sedat in the 1980s, who combined a series of widefield images taken at different focal positions to remove out-of-focus information (Cremer & Masters, 2013; Sage et al., 2017). The PSF is measured from very small beads or computed from a model of the optics, and the Richardson–Lucy algorithm is the maximum-likelihood method for the Poisson noise of photon counts (Sage et al., 2017). Image deblurring derives these methods, their regularization and boundary handling, and continues the example below.

Practical Example: Bleaching and Photon Noise

The example simulates a widefield fluorescence time-lapse of a synthetic specimen whose ground truth is known. The specimen, on a 128 × 128 grid, contains a uniform disc standing in for a nucleus, three thin lines standing in for filaments, two isolated point sources, and a pair of point sources 4 pixels apart; it emits a stated number of photons per pixel per frame, above a constant background of 2 photons per pixel. The PSF is a Gaussian with $\sigma = 2$ pixels, the in-focus approximation that optical microscopy mentions; the pair’s separation, $2\sigma$, is where the sum of two such Gaussians no longer has a dip between them. The specimen bleaches mono-exponentially with a time constant of 20 frames over 60 frames, and each frame is a Poisson sample of its expected counts, with no read noise. It uses NumPy 2.3.5, SciPy 1.16.3 and scikit-image 0.25.2; all parameters were fixed before the first run, and the only change since is the removal of an import that the moved deconvolution part needed, which leaves the output unchanged. It measures the signal-to-noise ratio in the centre of the disc, fits the bleaching time constant from the frames and applies an exponential bleach correction. The point sources and filaments are not measured here; image deblurring deconvolves the first and last frames of the same simulation with the known PSF.

import numpy as np
from scipy.signal import fftconvolve
from skimage.draw import disk, line

# Ground truth: emission in photons per pixel per frame at t = 0
obj = np.zeros((128, 128))
obj[disk((40, 40), 12)] = 100.0                       # uniform "nucleus"
for r0, c0, r1, c1 in [(90, 10, 120, 60), (70, 70, 70, 120), (10, 80, 50, 120)]:
    obj[line(r0, c0, r1, c1)] = 150.0                 # thin "filaments"
for r, c in [(100, 90), (40, 95), (100, 100), (100, 104)]:
    obj[r, c] = 5000.0                                # point sources; last two are a pair 4 px apart
background = 2.0                                      # photons/px, does not bleach

y, x = np.mgrid[-12:13, -12:13]
psf = np.exp(-(x**2 + y**2) / (2 * 2.0**2))
psf /= psf.sum()                                      # Gaussian PSF, sigma = 2 px

blurred = fftconvolve(obj, psf, mode="same")
tau, T = 20.0, 60                                     # bleaching time constant (frames), frame count

def expected(t):
    return blurred * np.exp(-t / tau) + background

def series(seed):
    rng = np.random.default_rng(seed)
    return np.stack([rng.poisson(expected(t)) for t in range(T)]).astype(float)

rows, cols = np.mgrid[:128, :128]
core = (rows - 40) ** 2 + (cols - 40) ** 2 <= 6**2    # centre of the nucleus
bg_region = (slice(0, 10), slice(0, 30))              # contains no objects

frames = series(0)
print(" t  expected  mean    std    SNR  sqrt(expected)")
for t in (0, 10, 20, 40, 59):
    v = frames[t][core]
    e = expected(t)[core].mean()
    print(f"{t:2d}  {e:7.2f}  {v.mean():6.2f}  {v.std():5.2f}  {v.mean() / v.std():5.2f}  {np.sqrt(e):5.2f}")

bg = np.array([f[bg_region].mean() for f in frames])
signal = np.array([f[core].mean() for f in frames]) - bg
slope, _ = np.polyfit(np.arange(T), np.log(signal), 1)
tau_fit = -1 / slope
print(f"background estimate {bg.mean():.2f} photons/px, fitted tau = {tau_fit:.2f} frames (true {tau:.0f})")
for t in (0, 59):
    corr = (frames[t] - bg[t]) / np.exp(-t / tau_fit)
    v = corr[core]
    print(f"bleach-corrected t={t:2d}: mean {v.mean():6.2f}, SNR {v.mean() / v.std():5.2f}")

 t  expected  mean    std    SNR  sqrt(expected)
 0   101.95  101.66   9.44  10.77  10.10
10    62.62   62.51   7.82   7.99   7.91
20    38.77   38.73   6.16   6.29   6.23
40    15.53   16.01   4.36   3.67   3.94
59     7.23    6.80   2.57   2.65   2.69
background estimate 2.00 photons/px, fitted tau = 19.98 frames (true 20)
bleach-corrected t= 0: mean  99.69, SNR 10.56
bleach-corrected t=59: mean  93.10, SNR  1.89

The measured signal-to-noise ratio tracks $\sqrt{\lambda}$ as the expected count falls from about 100 to 7 photons per pixel: bleaching costs signal-to-noise ratio exactly as fewer photons would. The fit recovers the time constant almost exactly because the simulated bleaching is mono-exponential and the background constant, assumptions that Miura (2020) warns real data need not satisfy. The correction then returns the last frame’s mean to about the original level, but its signal-to-noise ratio falls to below 2, lower even than before correction: rescaling cannot add information, and subtracting the background removes its photons from the signal while leaving their noise in place.

Characteristics

Property	What sets it	Consequence for analysis
Contrast	Labels, excitation and emission filters	Only labelled structures are visible; background is dark but not zero
Signal level	Photons detected per pixel: labelling, excitation power, exposure, detector	Poisson noise with SNR $\sqrt{\lambda}$; read noise matters at low counts
Optical sectioning	Widefield, confocal pinhole, multiphoton, TIRF, light sheet	Out-of-focus haze in widefield; sections in the others
Signal over time	Photobleaching	Intensities fall during time-lapse; correction rescales but cannot restore SNR
Channel purity	Emission spectra, filters	Bleed-through between channels; unmixing needs reference spectra
Resolution	NA, wavelength, PSF, sampling	Diffraction-limited unless a super-resolution variant is used; deconvolution sharpens within limits

Calibration

The quantities that the sections above show must be measured rather than assumed are the detector offset, pixel by pixel for sCMOS cameras; the PSF, from beads or a model; the reference spectra for unmixing, on the instrument used; and the background level for bleach correction, from a region without fluorescence or a blank acquisition (Miura, 2020). Shading across the field and the pixel size are calibrated as for any light microscope, as optical microscopy describes.

Processing the Data

A typical chain begins with offset subtraction and shading correction, unmixing if channels overlap, and bleach correction for time-lapse data. Denoising or deconvolution follows when the analysis needs it; the general denoising methods are in image denoising, and the noise model that should guide the choice is Poisson with a Gaussian component, not the additive white Gaussian noise of most benchmarks. Light-sheet data first need views registered and fused and stripes removed (Olarte et al., 2018). The restored channels then go to cell segmentation, tracking and measurement.

Strengths

• Specific contrast: only the labelled molecule or structure is imaged, on a dark background, and several labels can be imaged at once as channels (Mualla et al., 2018; Bankhead, 2022–2025).
• Intensities proportional to fluorophore numbers within one experiment, which makes the images measurements rather than pictures (Bankhead, 2022–2025).
• Live imaging with genetically encoded labels, and with light-sheet illumination, of whole developing organisms over days (Olarte et al., 2018).

Limitations

• Photon-limited: every gain in speed, resolution or duration costs signal-to-noise ratio.
• Photobleaching makes intensities time-dependent, and phototoxicity can alter or kill the cells under study (Miura, 2020; Wäldchen et al., 2015).
• Bleed-through between channels, and labels that may perturb the specimen or cover it incompletely (McRae et al., 2019; Mualla et al., 2018).
• Diffraction-limited resolution and, in widefield, out-of-focus light; deconvolution can introduce artifacts (Sage et al., 2017).

Typical Applications

Cell biology relies on fluorescence to localize proteins, follow them in living cells and measure many features per cell in automated screens, which Bioimage analysis surveys. Developmental biology uses light-sheet microscopy to image embryos of Drosophila, zebrafish and other small organisms over long periods (Olarte et al., 2018). Mualla et al. (2018) give a viability assay as an example: propidium iodide enters only cells whose membranes have lost selective permeability and binds their nucleic acids, which greatly increases its fluorescence, so dead cells stand out from unstained living ones. They also describe confocal laser endomicroscopy, in which a fibre bundle carries laser light into the body and a fluorescent contrast agent outlines cellular structures, as used in the gastrointestinal tract. In pathology, fluorescence and multiplexed slides are scanned as described in whole-slide imaging.

History

Cremer and Masters (2013) trace the history of confocal microscopy. Its invention is often attributed to Marvin Minsky, who filed his patent on 7 November 1957, but it was partly predated by a paper of Hiroto Naora in 1951, and practical confocal fluorescence microscopy became possible only with lasers of short enough wavelengths in the 1960s; deconvolution-based reconstruction and confocal laser scanning fluorescence microscopy, introduced in the 1980s, made it practical to resolve detail along the optical axis. They also trace light-sheet illumination back to the lateral illumination of Siedentopf and Zsigmondy’s ultramicroscope, combined with laser excitation and orthogonal detection schemes of the 1990s. The selective plane illumination microscope of 2004 and the 2008 Nobel Prize for the green fluorescent protein are summarized in Bioimage analysis.

Super-resolution fluorescence microscopy, recognized by the 2014 Nobel Prize in Chemistry, goes below the diffraction limit by controlling the states of fluorophores, in stimulated emission depletion (STED) microscopy and in the single-molecule localization methods PALM and STORM; structured illumination microscopy, by contrast, improves lateral resolution by at most a factor of two as long as the emission stays linear in the excitation, and remains confined by the diffraction limit (Royal Swedish Academy of Sciences, 2014). Bioimage analysis summarizes these methods. Deconvolution also works within the limit.

Comparison with Other Sensors

• Brightfield, phase contrast and DIC image the whole specimen by absorption or phase without labels, as optical microscopy describes; fluorescence trades that generality for specificity and pays with photon-limited noise, bleaching and phototoxicity.
• Super-resolution fluorescence variants resolve below the diffraction limit; localization microscopy needs irradiances in the kW cm⁻² range, which is why Wäldchen et al. (2015) studied its phototoxicity.
• Electron microscopy resolves far finer detail but works in vacuum on dead specimens, as optical microscopy notes.
• Whole-slide scanners apply fluorescence to entire slides as tiled multichannel images for pathology, as whole-slide imaging describes.