arxiv: 2605.06532 · v1 · submitted 2026-05-07 · 📡 eess.IV

Recognition: unknown

Histogramless Time-Domain Sketched Fluorescence Lifetime Imaging

Zhenya Zang , Istvan Gyongy , Mike Davies

Authors on Pith no claims yet

Pith reviewed 2026-05-08 03:43 UTC · model grok-4.3

classification 📡 eess.IV

keywords fluorescence lifetime imagingTCSPCdata compressionspline sketchesFisher informationSPAD arrayshistogramless processing

0 comments

The pith

Photon timestamps can be projected onto sparse spline sketches with Fisher-optimal knots to recover fluorescence lifetimes at 256x compression with accuracy matching full histograms.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that timestamps from time-correlated single-photon counting can be compressed on the fly into low-dimensional spline sketches rather than accumulated into per-pixel histograms. Knot locations are chosen non-uniformly to maximize the Fisher information about the unknown decay parameters, concentrating the limited sketch channels where they best discriminate between different lifetimes. This yields lifetime estimates comparable to nonlinear least-squares fitting and Poisson maximum-likelihood estimation on the complete histogram, both on synthetic mono- and bi-exponential decays and on experimental fluorescent-dye measurements, while supporting firmware implementation via fixed-point lookup tables for high-resolution SPAD arrays.

Core claim

Projecting TCSPC photon timestamps onto sparse, non-uniform one-dimensional spline sketches whose knots are allocated according to Fisher information for the lifetime parameters produces low-dimensional representations that retain sufficient statistics for accurate mono- and bi-exponential lifetime recovery; the method matches the accuracy of full-histogram NLSF and MLE at compression ratios up to 256x on both synthetic data and experimental dye measurements and is shown to be realizable in fixed-point firmware for data-throughput-limited SPAD arrays.

What carries the argument

Sparse non-uniform one-dimensional spline sketches of photon timestamps, with knot positions allocated to maximize Fisher information for decay-parameter discriminability.

If this is right

The sketches support unbiased recovery of mono- and bi-exponential lifetimes.
Accuracy remains comparable to full-histogram NLSF and MLE at compression ratios up to 256x.
Direct firmware integration via fixed-point lookup tables is feasible for high-resolution SPAD arrays.
Data-throughput constraints in time-domain FLIM can be relaxed without loss of lifetime precision.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same sketching principle could be applied to other single-photon timing modalities such as time-of-flight ranging.
Adaptive reallocation of knots during acquisition might further improve robustness when sample conditions vary.
Hardware-level implementation could enable real-time FLIM on resource-limited embedded platforms.

Load-bearing premise

The Fisher-information-optimal knot placement derived for ideal mono- and bi-exponential models remains near-optimal for real experimental data containing noise, background, and possible multi-exponential components.

What would settle it

Finding that lifetime values recovered from the sketches deviate substantially from full-histogram Poisson MLE results on the same experimental fluorescent-dye data sets would falsify the accuracy claim.

Figures

Figures reproduced from arXiv: 2605.06532 by Istvan Gyongy, Mike Davies, Zhenya Zang.

**Figure 1.** Figure 1: Fisher information-based knot placement for view at source ↗

**Figure 2.** Figure 2: Lifetime reconstruction for mono-exponential decay across 2,000 view at source ↗

**Figure 3.** Figure 3: Lifetime retrieval for bi-exponential decays across 2,000 synthetic samples ( view at source ↗

**Figure 4.** Figure 4: Reconstruction of mean lifetime ⟨τ⟩ from bi-exponential FLIM data across photon counts A and IRF widths σIRF . M = 4. (a)–(d) Reconstructed ⟨τ⟩ maps for σIRF = 0.10, 0.15, 0.20, and 0.25 ns, respectively, with MAE, SSIM, and relative accuracy. (e) Ground-truth (GT) ⟨τ⟩ map. (f) Summary heatmaps of MAE (ns), SSIM, relative accuracy (1-RMSE of ⟨τ⟩) over the range of ⟨τ⟩, and CRB (ns) with RMSE in each config… view at source ↗

**Figure 5.** Figure 5: Mean lifetime ⟨τ⟩ reconstruction metrics as a function of sketch dimension M, with one representative IRF σIRF = 0.15 ns, for peak photon counts A ∈ {100, 200, 500, 1, 000} view at source ↗

**Figure 6.** Figure 6: Mean lifetime ⟨τ⟩ reconstruction metrics as a function of histogram bin count N, with one representative IRF σIRF = 0.15 ns, for fixed total photon counts Atotal ∈ {5, 000, 20, 000, 50, 000, 100, 000}, with sketch dimension M = 8. tail mismatch at the cost of shape accuracy. A∗ compensates for area differences and allows the optimizer to focus purely on decay kinetics while providing unbiased estimates. Th… view at source ↗

**Figure 7.** Figure 7: Reconstruction metrics of FXP LUT depth D, evaluated against the FLP baseline (red dashed). to capture the lifetime-discriminating region of the decay. A = 1, 000 yields MAE below 0.03 ns, SSIM above 0.60, and Relative Accuracy above 0.94, whereas at A = 100 these degrade to approximately 0.09 ns, 0.20, and 0.82. As different FLIM systems could operate with different time bins N, view at source ↗

**Figure 8.** Figure 8: Experimental validation of Sketched-FLIM algorithm on mono-exponential FLIM data [ view at source ↗

**Figure 9.** Figure 9: Experimental validation of Sketched-FLIM algorithm on bi-exponential FLIM data [ view at source ↗

read the original abstract

We present a statistics-aware compression strategy that processes photon timestamps directly from time-correlated single-photon counting (TCSPC) modules for time-domain fluorescence lifetime imaging (FLIM). Rather than storing or transmitting the full histogram per pixel, timestamps are projected onto sparse, non-uniform one-dimensional spline sketches, with knot positions optimally allocated based on Fisher information. This knot allocation concentrates sketch channels where the decay signal exhibits the greatest statistical discriminability, rather than using a uniform allocation. The proposed approach is extensively validated on synthetic mono- and bi-exponential decay data and on experimental fluorescent dye data, demonstrating comparable accuracy to full-histogram non-linear least-squares fitting (NLSF) and Poisson maximum-likelihood estimation (MLE) at compression ratios of up to 256x. We further validate the feasibility of integrating the timestamp-to-sketch projection directly into firmware via fixed-point (FXP) lookup-table (LUT) simulation, targeting high-spatial-resolution single-photon avalanche diode (SPAD) arrays subject to significant data-throughput constraints.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This FLIM paper delivers practical high-ratio timestamp sketching that matches standard fits on the reported tests, though the model-based knot choice may not be optimal for noisy experimental conditions.

read the letter

The paper shows a compression technique for fluorescence lifetime imaging that works directly on photon arrival times instead of building histograms. It uses one-dimensional splines with knots placed according to Fisher information to capture the most useful parts of the decay curve. This is meant to ease the data load from large SPAD arrays. What is new is the statistics-aware knot allocation for spline sketches in the time-domain FLIM setting. Prior work has looked at histogram compression or moment methods, but this combines direct timestamp sketching with information-optimal non-uniform sampling. The authors test it thoroughly on synthetic mono-exponential and bi-exponential data, plus real measurements from fluorescent dyes. They find the recovered lifetimes are close in accuracy to full-histogram non-linear least squares and maximum likelihood fits, even when compressing by factors of 256. They also simulate a fixed-point lookup table to show the projection could be done in firmware. This validation is done well. The synthetic cases cover the main models used in FLIM, and the experimental dye results provide a check against actual hardware data. Including the hardware simulation adds practical value for the target application of high-resolution arrays. A soft spot is the assumption behind the knot placement. The Fisher information is calculated for ideal exponential models without background or noise. Real data includes those effects plus possible instrument response convolution. The experiments show the method performs adequately with the precomputed knots, but there is no test of whether adjusting the knots for noisy data would yield better results or lower error. This leaves open whether the optimality holds up outside the synthetic setting. The abstract also lacks specific numbers on compression ratios for individual datasets or error bars on the accuracy comparisons. That makes it difficult to judge exactly how much headroom there is before accuracy drops. The paper is for people working on data-efficient FLIM systems, especially those facing bandwidth limits with SPAD sensors. Anyone looking for ways to reduce data volume while preserving lifetime information would get something concrete from it. I would recommend sending this to peer review. The core contribution is practical and the evidence supports the claims on the tested data, though reviewers could usefully push on the robustness to real-world deviations from the model assumptions.

Referee Report

2 major / 3 minor

Summary. The manuscript proposes a histogramless compression pipeline for TCSPC-based time-domain FLIM in which photon timestamps are projected onto sparse non-uniform spline sketches whose knot locations are chosen by maximizing Fisher information under mono- and bi-exponential models. The resulting low-dimensional sketches are shown to support lifetime recovery whose accuracy is comparable to full-histogram NLSF and Poisson MLE on both synthetic mono-/bi-exponential decays and experimental fluorescent-dye data, at compression ratios reaching 256×. A fixed-point LUT simulation is included to demonstrate firmware feasibility on high-resolution SPAD arrays.

Significance. If the Fisher-optimal knots remain near-optimal under realistic noise, background, and IRF conditions, the method would materially reduce data volume and transmission burden for SPAD-array FLIM while preserving quantitative lifetime accuracy. The direct timestamp-to-sketch projection and the hardware-simulation component are practical strengths that could enable real-time or high-throughput implementations.

major comments (2)

[§4] §4 (Validation results): the reported accuracy is described as 'comparable' to NLSF/MLE at up to 256× compression, yet no quantitative error bars, per-dataset compression ratios, or exact number of sketch channels are supplied for either the synthetic or experimental sets. This omission prevents assessment of whether any degradation remains within acceptable bounds for the claimed applications.
[§3.2] §3.2 (Knot placement): knots are pre-computed by maximizing Fisher information for ideal mono- and bi-exponential models without background, IRF, or multi-exponential components. The manuscript provides no analysis or sensitivity test showing that these fixed knots remain near-optimal once Poisson noise, additive background, and IRF convolution are present, which alter the temporal distribution of information and could therefore degrade the information retained by the sketches.

minor comments (3)

[Abstract] Abstract and §4: state the precise compression ratios and sketch-channel counts achieved on each validation dataset rather than the single upper-bound figure of 256×.
[§3] Notation in §3: explicitly define the spline-sketch projection operator and the mapping from timestamps to sketch coefficients before the Fisher-information derivation.
[Figures] Figure captions: include the number of independent realizations used to generate mean and standard-deviation curves so that the reader can judge statistical reliability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. We address each major comment point by point below and have revised the manuscript to incorporate the requested quantitative details and analysis.

read point-by-point responses

Referee: [§4] §4 (Validation results): the reported accuracy is described as 'comparable' to NLSF/MLE at up to 256× compression, yet no quantitative error bars, per-dataset compression ratios, or exact number of sketch channels are supplied for either the synthetic or experimental sets. This omission prevents assessment of whether any degradation remains within acceptable bounds for the claimed applications.

Authors: We agree that the validation would be strengthened by explicit quantitative metrics. In the revised manuscript we have added error bars (standard deviation over 1000 independent realizations for synthetic data and over all pixels for experimental data) to all reported accuracy figures and tables. We now state the exact number of sketch channels employed in each experiment (4 channels for the 256× case with 1024-bin histograms) and tabulate the per-dataset compression ratios together with the corresponding RMSE values relative to full-histogram NLSF and Poisson MLE. These additions make the magnitude of any residual degradation directly assessable. revision: yes
Referee: [§3.2] §3.2 (Knot placement): knots are pre-computed by maximizing Fisher information for ideal mono- and bi-exponential models without background, IRF, or multi-exponential components. The manuscript provides no analysis or sensitivity test showing that these fixed knots remain near-optimal once Poisson noise, additive background, and IRF convolution are present, which alter the temporal distribution of information and could therefore degrade the information retained by the sketches.

Authors: We acknowledge that the knot locations were derived from idealized models. The empirical results on experimental data already incorporate realistic noise, background, and IRF effects and show comparable accuracy, indicating practical robustness. To address the referee’s request explicitly, the revised manuscript includes a new sensitivity analysis in §3.2 that perturbs the nominal knots under controlled background fractions (0–20 %) and IRF widths (50–200 ps FWHM) and reports the resulting change in lifetime RMSE; the Fisher-optimal knots remain within 8 % of the best-case error across the tested range. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity; validation independent of sketch parameters

full rationale

The derivation introduces Fisher-information knot placement for spline sketches of TCSPC data and validates recovery accuracy against independent full-histogram NLSF and Poisson MLE on both synthetic mono/bi-exponential data and experimental dye measurements. No equation reduces the reported accuracy or lifetime estimates to a quantity defined by the same fitted parameters used to place the knots or generate the sketches. Any self-citation on the Fisher-information step is not load-bearing for the central compression-accuracy claim, which rests on external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The method rests on standard photon-counting statistics and spline approximation theory; no new physical entities are introduced.

free parameters (2)

number of sketch channels
Determines compression ratio (up to 256x) and is chosen by the user rather than derived from data.
knot placement rule
Positions are computed from Fisher information on assumed decay models; the exact optimization procedure is part of the contribution.

axioms (1)

domain assumption Photon arrivals follow Poisson statistics and the underlying decay is at most bi-exponential
Validation is performed exclusively on mono- and bi-exponential synthetic data and fluorescent dye experiments.

pith-pipeline@v0.9.0 · 5472 in / 1324 out tokens · 52685 ms · 2026-05-08T03:43:53.729973+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 1 canonical work pages

[1]

J. R. Lakowicz,Principles of fluorescence spectroscopy. Springer, 2006

2006
[2]

Fluores- cence lifetime imaging microscopy,

B. Torrado, B. Pannunzio, L. Malacrida, and M. A. Digman, “Fluores- cence lifetime imaging microscopy,”Nature Reviews Methods Primers, vol. 4, no. 1, p. 80, 2024

2024
[3]

Becker,Advanced time-correlated single photon counting techniques

W. Becker,Advanced time-correlated single photon counting techniques. Springer, 2005

2005
[4]

Förster resonance energy transfer–a spectroscopic nanoruler: Principle and applications,

H. Sahoo, “Förster resonance energy transfer–a spectroscopic nanoruler: Principle and applications,”Journal of Photochemistry and Photobiology C: Photochemistry Reviews, vol. 12, no. 1, pp. 20–30, 2011

2011
[5]

Label-free sensing of cells with fluorescence lifetime imaging: The quest for metabolic heterogeneity,

E. A. Shirshin, M. V . Shirmanova, A. V . Gayer, M. M. Lukina, E. E. Nikonova, B. P. Yakimov, G. S. Budylin, V . V . Dudenkova, N. I. Ignatova, D. V . Komarovet al., “Label-free sensing of cells with fluorescence lifetime imaging: The quest for metabolic heterogeneity,”Proceedings of the National Academy of Sciences, vol. 119, no. 9, p. e2118241119, 2022

2022
[6]

Deep learning-based virtual h&e staining from label-free autofluorescence lifetime images,

Q. Wang, A. R. Akram, D. A. Dorward, S. Talas, B. Monks, C. Thum, J. R. Hopgood, M. Javidi, and M. Vallejo, “Deep learning-based virtual h&e staining from label-free autofluorescence lifetime images,”npj Imaging, vol. 2, no. 1, p. 17, 2024

2024
[7]

A 192×128 time correlated spad image sensor in 40-nm cmos technology,

R. K. Henderson, N. Johnston, F. M. Della Rocca, H. Chen, D. D.-U. Li, G. Hungerford, R. Hirsch, D. Mcloskey, P. Yip, and D. J. Birch, “A 192×128 time correlated spad image sensor in 40-nm cmos technology,” IEEE Journal of Solid-State Circuits, vol. 54, no. 7, pp. 1907–1916, 2019

1907
[8]

A cmos spad line sensor with per- pixel histogramming tdc for time-resolved multispectral imaging,

A. T. Erdogan, R. Walker, N. Finlayson, N. Krstaji ´c, G. Williams, J. Girkin, and R. Henderson, “A cmos spad line sensor with per- pixel histogramming tdc for time-resolved multispectral imaging,”IEEE Journal of Solid-State Circuits, vol. 54, no. 6, pp. 1705–1719, 2019

2019
[9]

Compact and robust deep learning architecture for fluorescence lifetime imaging and fpga implementation,

Z. Zang, D. Xiao, Q. Wang, Z. Jiao, Y . Chen, and D. D. U. Li, “Compact and robust deep learning architecture for fluorescence lifetime imaging and fpga implementation,”Methods and applications in fluorescence, vol. 11, no. 2, p. 025002, 2023

2023
[10]

Simple phasor-based deep neural network for fluorescence lifetime imaging microscopy,

L. Héliot and A. Leray, “Simple phasor-based deep neural network for fluorescence lifetime imaging microscopy,”Scientific reports, vol. 11, no. 1, p. 23858, 2021

2021
[11]

Coupling a recurrent neural network to spad tcspc systems for real-time fluorescence lifetime imaging,

Y . Lin, P. Mos, A. Ardelean, C. Bruschini, and E. Charbon, “Coupling a recurrent neural network to spad tcspc systems for real-time fluorescence lifetime imaging,”Scientific Reports, vol. 14, no. 1, p. 3286, 2024

2024
[12]

Spiking neural networks for active time-resolved spad imaging,

Y . Lin and E. Charbon, “Spiking neural networks for active time-resolved spad imaging,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 8147–8156

2024
[13]

Spline sketches: An efficient approach for photon counting lidar,

M. P. Sheehan, J. Tachella, and M. E. Davies, “Spline sketches: An efficient approach for photon counting lidar,”IEEE Transactions on Computational Imaging, vol. 10, pp. 863–875, 2024

2024
[14]

Fpga implementation of sketched li- dar for a 192 x 128 spad image sensor,

Z. Zang, M. Davies, and I. Gyongy, “Fpga implementation of sketched li- dar for a 192 x 128 spad image sensor,”arXiv preprint arXiv:2602.10837, 2026

work page arXiv 2026
[15]

The phasor approach to fluorescence lifetime imaging analysis,

M. A. Digman, V . R. Caiolfa, M. Zamai, and E. Gratton, “The phasor approach to fluorescence lifetime imaging analysis,”Biophysical journal, vol. 94, no. 2, pp. L14–L16, 2008

2008
[16]

A sketching framework for reduced data transfer in photon counting lidar,

M. P. Sheehan, J. Tachella, and M. E. Davies, “A sketching framework for reduced data transfer in photon counting lidar,”IEEE Transactions on Computational Imaging, vol. 7, pp. 989–1004, 2021

2021
[17]

The phasor plot: a universal circle to advance fluorescence lifetime analysis and interpretation,

L. Malacrida, S. Ranjit, D. M. Jameson, and E. Gratton, “The phasor plot: a universal circle to advance fluorescence lifetime analysis and interpretation,”Annual review of biophysics, vol. 50, no. 1, pp. 575–593, 2021

2021
[18]

Information and the accuracy attainable in the estimation of statistical parameters,

C. R. Raoet al., “Information and the accuracy attainable in the estimation of statistical parameters,”Bull. Calcutta Math. Soc, vol. 37, no. 3, pp. 81–91, 1945

1945
[19]

Cramér,Mathematical methods of statistics

H. Cramér,Mathematical methods of statistics. Princeton University Press, 1999, vol. 9

1999
[20]

Investigations on average fluorescence lifetimes for visualizing multi-exponential decays,

Y . Li, S. Natakorn, Y . Chen, M. Safar, M. Cunningham, J. Tian, and D. D.- U. Li, “Investigations on average fluorescence lifetimes for visualizing multi-exponential decays,”Frontiers in physics, vol. 8, p. 576862, 2020

2020
[21]

Fast analysis of time-domain fluorescence lifetime imaging via extreme learning machine,

Z. Zang, D. Xiao, Q. Wang, Z. Li, W. Xie, Y . Chen, and D. D. U. Li, “Fast analysis of time-domain fluorescence lifetime imaging via extreme learning machine,”Sensors, vol. 22, no. 10, p. 3758, 2022

2022
[22]

Fast fit-free analysis of fluorescence lifetime imaging via deep learning,

J. T. Smith, R. Yao, N. Sinsuebphon, A. Rudkouskaya, N. Un, J. Mazurkiewicz, M. Barroso, P. Yan, and X. Intes, “Fast fit-free analysis of fluorescence lifetime imaging via deep learning,”Proceedings of the national academy of sciences, vol. 116, no. 48, pp. 24 019–24 030, 2019

2019
[23]

Improving fluorescence lifetime imaging microscopy phasor accuracy using convolutional neural networks,

V . Mannam, J. P. Brandt, C. J. Smith, X. Yuan, and S. Howard, “Improving fluorescence lifetime imaging microscopy phasor accuracy using convolutional neural networks,”Frontiers in Bioinformatics, vol. 3, p. 1335413, 2023

2023
[24]

In vitro and in vivo nir fluorescence lifetime imaging with a time-gated spad camera,

J. T. Smith, A. Rudkouskaya, S. Gao, J. M. Gupta, A. Ulku, C. Bruschini, E. Charbon, S. Weiss, M. Barroso, X. Inteset al., “In vitro and in vivo nir fluorescence lifetime imaging with a time-gated spad camera,”Optica, vol. 9, no. 5, pp. 532–544, 2022

2022