How Many Photons Does It Take to Tell Fluorescence from Reflection?

Many real surfaces, including coral, leaves, dyed fabric and biological tissue, do two things to incoming light at once. They reflect it at the same wavelength, and they fluoresce: they absorb short-wavelength light and re-emit it at longer wavelengths. A camera records the sum of the two. One line of work separates them computationally by shining light whose spectrum oscillates rapidly across wavelength. Reflection copies those oscillations, while fluorescence averages over its broad absorption band and washes them out. These methods were designed and evaluated as if measurements were noise-free, or corrupted only by constant Gaussian noise. In practice the interesting applications are photon-starved, where every spectral band counts only a handful of photons and the noise is Poisson. Examples include live-cell imaging limited by phototoxicity, deep-sea biofluorescence surveys, and snapshot hyperspectral capture. In that regime nobody knows how many photons a given separation accuracy costs. Nobody knows how fast the light should oscillate, or when it would be better to simply put a filter in front of the lens.

This note proposes to answer those questions with the tools that physicists use to state sensitivity limits for counting experiments: Fisher information, the Cramér-Rao bound, and the split of error into bias and variance. The central prediction is that there is an optimal modulation frequency for any photon budget, rather than "the faster the better." Faster oscillation suppresses fluorescence leakage into the reflection estimate. It also forces finer spectral sampling, which spreads the same photons over more, noisier bands. The bound locates that optimum, predicts the photon count below which a filter-based capture wins, and lets us design the illumination spectrum directly.

Core Claim

We claim three things about separating reflection and fluorescence under spectrally modulated light at a fixed photon budget.

First, the achievable error is well predicted by a Poisson Fisher-information bound plus a leakage bias term. The condition number of the mixing matrix does not predict it well, yet that is the quantity prior work used to judge illuminants.

Second, the error-minimizing modulation frequency is interior and rises with photon budget.

Third, illumination spectra chosen to minimize the bound, paired with a Poisson maximum-likelihood estimator, reach a target accuracy with measurably fewer photons than the phase-shifted sinusoids of prior work.

The novelty lies in the bound-driven design and the predicted crossovers, not in using a Poisson likelihood. Poisson regression for fluorescence unmixing already exists.

Where This Sits

The separation method comes from three papers. Fu et al. showed that two phase-shifted sinusoidal spectral illuminations separate the two components by simple arithmetic. They then recovered absorption spectra by sparse coding against more than 500 known dye spectra, with an absorption RMSE of 0.012 over 183 materials. Their Nyquist argument gives a minimum modulation frequency but says nothing about photons.

Zheng et al. replaced the programmable source with a spiky HID car headlamp and a single hyperspectral image. They inverted a low-dimensional linear model with a pseudoinverse, reporting a condition number of 12.4 against infinity for smooth white light. We test whether that condition number actually predicts error once counts are low.

Koyamatsu et al. moved to RGB cameras under narrow-band LEDs with a per-pixel prior over fluorescent colors under Gaussian noise. They also showed that blind source separation fails on this problem, with an ICA RMS error around 38.0 against below 5.0 for their physics-based methods. All three motivate the proposal by leaving noise statistics unmodeled.

The statistical side supplies the method. Heide et al. replaced an invalid Poisson model for single-photon detectors with an exact multinomial likelihood. They found that information throughput peaks near one detection per pulse, an interior optimum set by the sensor physics. Our predicted optimal modulation frequency is the spectral analogue of that result. The same Fisher-information habit underlies optimal estimators in counting-limited science, such as the near-optimal bispectrum estimators of Creminelli et al..

Two papers constrain the design. SL-PRU already unmixes overlapping fluorophores with a Poisson likelihood plus sparse and low-rank spatial penalties. Adding a Poisson likelihood to spectral separation is therefore not new enough on its own, and SL-PRU becomes our strongest baseline rather than our contribution. Zhang et al.'s FMD dataset shows that confocal and two-photon fluorescence data are almost purely Poisson, while wide-field data carry more Gaussian read noise. That tells us the likelihood must be Poisson-Gaussian, not pure Poisson, for camera-based capture.

Research Program

We write the expected count in spectral band $k$ under illumination pattern $j$ as

$$\mu_{jk} = \tau\Big[I_j(\lambda_k)\,R(\lambda_k) + E(\lambda_k)\sum_{\lambda} A(\lambda)\,I_j(\lambda)\Big] + d_k,$$

where the symbols are:

  • $\tau$ is the exposure scale that sets the photon budget;
  • $I_j$ is the illumination spectrum of pattern $j$;
  • $R$ is the reflectance;
  • $A$ and $E$ are the fluorophore's absorption and emission spectra;
  • $d_k$ is the dark and background level.

The observed count is Poisson with mean $\mu_{jk}$, plus Gaussian read noise calibrated per sensor. Reflectance and emission are each parameterized in low-dimensional bases, as Zheng et al. did, giving a parameter vector $\theta$. For pure Poisson counts the Fisher information is

$$F(\theta) = \sum_{j,k} \frac{1}{\mu_{jk}}\,\nabla_\theta \mu_{jk}\,\nabla_\theta \mu_{jk}^{\top},$$

and read noise enters by adding its variance to $\mu_{jk}$ in the denominator. Two features of this expression drive everything. The fluorescence term adds to $\mu_{jk}$ in every band, so bright fluorescence inflates the variance of the reflection estimate. Finer spectral sampling lowers $\mu_{jk}$ per band.

For estimators that treat fluorescence as perfectly flat across phase shifts, like Fu et al.'s two-shot arithmetic, there is an additional leakage bias. It is proportional to $|\hat{A}(\omega)|$, the magnitude of the Fourier transform of the absorption spectrum at modulation frequency $\omega$. The predicted error is therefore this bias squared plus the trace of $F^{-1}$. The variance term shrinks as $1/\tau$, because Fisher information for Poisson counts scales linearly with expected counts. The bias term does not shrink with $\tau$. This is why the optimal $\omega$ should rise with photon budget.

We will build four things.

The first is the bound itself. It will be computed over the McNamara and Boswell spectra that Fu and Koyamatsu used. It is evaluated as a function of photon budget, modulation frequency, modulation depth, and the ratio of absorption bandwidth to modulation period.

The second is a Poisson-Gaussian maximum-likelihood estimator that models $A$ explicitly, so it has no leakage bias. It carries an optional spatial low-rank penalty borrowed from SL-PRU, so that the spatial prior can be switched on and off.

The third is an illumination design procedure. It chooses the programmable spectra $I_j$ to minimize the expected trace of $F^{-1}$, averaged over a prior on materials drawn from the dye database, subject to a total photon and power constraint.

The fourth is a crossover analysis against a filter-based capture that splits the same photon budget. That capture takes one shot with an emission barrier filter for fluorescence and one shot through a matched notch or short-pass arrangement for reflectance.

An extension to single-photon detectors is left as a follow-on. It would replace the Poisson likelihood with Heide et al.'s dead-time multinomial model.

Baselines

The naive reference point is prior work run exactly as published at reduced exposure. This means Fu et al.'s two-shot sinusoidal arithmetic and Zheng et al.'s pseudoinverse under their spiky source, both given the same total photon count as our method. The only reported numbers for these methods are Fu et al.'s absorption RMSE of 0.012 over 183 materials and Zheng et al.'s condition number of 12.4. Both are for effectively high-signal conditions. There is no reported result found on Lacuna for reflection-fluorescence separation error as a function of photon count, so the low-light curves of these baselines will be produced by us. Beating them means lower reflectance and emission RMSE at every photon level below the point where the methods converge.

The strongest existing competitor is SL-PRU from the sparse and low-rank Poisson regression paper, adapted by adding the known illumination-times-reflectance term as an extra endmember family. It is the strongest because it already uses the correct Poisson likelihood and a spatial prior shown to help at low counts. Its only reported gains are relative improvements in segmented cell circularity over non-negative least squares, 16% for Streptococcus and 18.4% for Veillonella, and it reports no reflection-separation number.

To keep the comparison fair, SL-PRU and our estimator will see identical raw captures, identical total photon budgets, and the same spatial window and regularization search. Hyperparameters will be tuned on a shared held-out set of simulated scenes, with an equal number of trials per method. SL-PRU will be run under Fu et al.'s sinusoidal illumination, which is the natural choice for a practitioner. Beating it requires the designed illumination to reach SL-PRU's reflectance RMSE at a target level with at least 30% fewer photons, across the dye database rather than on hand-picked materials.

The hardware baseline is a time-multiplexed filter capture with equal total exposure, with the photon split optimized by us so that it is not handicapped.

Three ablations isolate the ingredients. The first compares Poisson maximum likelihood with Fu's sinusoid against the same estimator with designed spectra, which isolates illumination design. The second compares Gaussian-weighted least squares after an Anscombe transform against Poisson maximum likelihood on the same data, which isolates the likelihood. The third toggles the spatial penalty, which isolates the spatial prior.

Why not just…?

Why not just use a barrier filter and take a separate reflectance shot? This is a real threat. It is the reason the crossover analysis is a core deliverable rather than a footnote.

Filters discard the reflected excitation light. They also need known fluorophore bands, which fails when several unknown dyes overlap, as the Fu et al. line of work notes. At very low photon budgets, however, a filter avoids fluorescence shot noise entirely in the reflectance shot. The Fisher analysis predicts it can win there.

We will report the photon count at which computational separation starts to win, both predicted by the bound and measured. If that count is above the budgets of the target applications, the honest conclusion is that filters are the right tool there, and the bound tells practitioners so.

Why not just add a reflection term to SL-PRU and stop? This may capture most of the gain. If it does, the illumination design contribution collapses.

The ablation that swaps Fu's sinusoid for designed spectra, under the same Poisson estimator, settles this. If the photon saving is under about 10%, we will present the work as a predictive sensitivity analysis rather than a design method. In that framing the bound tells users which light source and modulation to buy, and which exposure to use.

Why not just variance-stabilize and run the old pseudoinverse, or denoise first? On pure fluorescence denoising, FMD shows that VST followed by BM3D reaches 32.71 dB against 35.40 dB for Noise2Noise. The Poisson-aware and learned options win there, but that benchmark contains no separation task.

Our specific concern is that the Anscombe transform is biased at counts of a few photons, and that denoising before separation mixes the two components' noise. The likelihood ablation measures both effects directly.

A learned separator would need paired reflection and fluorescence ground truth. No dataset in the file supplies such pairs, so we do not treat it as a primary baseline.

Proposed Experiments

Simulation comes first. We will render scenes from the McNamara and Boswell dye spectra with measured reflectances. Noise will follow Poisson-Gaussian statistics whose read-noise-to-shot-noise ratios are matched to FMD's confocal and wide-field regimes, including its five averaging levels as reference exposure points. Photon budgets will sweep from about one to ten thousand counts per band. We report reflectance RMSE, emission RMSE, and absorption RMSE using Fu et al.'s sparse-coding step, each against the bound.

The physical study uses a programmable spectral light source and a hyperspectral camera on fluorescent charts and coral or leaf samples. Ground truth comes from long-exposure averaging, following FMD's 50-frame protocol, cross-checked by filter measurements.

Three decisive checks decide the claims.

First, across illuminants including Zheng et al.'s HID-like spectrum, does the bound rank empirical low-light error correctly where the condition number does not? We will measure this by Spearman correlation across at least ten illuminants and five photon levels.

Second, does the measured optimal modulation frequency rise with photon budget as the bias-variance split predicts? A flat or monotone-in-frequency optimum would falsify the core claim.

Third, is the measured crossover with the filter baseline within a factor of two of the predicted crossover? Agreement would show that the bound is a usable planning tool. Disagreement would indicate that the noise model or the leakage model is missing something, which we would then diagnose from the residuals.