Widefield Low-Dose

Low-Dose Widefield Microscopy

Join the Competition Contribute to Benchmark

Standard reconstruction benchmark — forward model perfectly known, no calibration needed. Score = 0.5 × clip((PSNR−15)/30, 0, 1) + 0.5 × SSIM

#	Method	Score	PSNR (dB)	SSIM	Source
🥇	ScoreMicro ScoreMicro Wei et al., ECCV 2025 38.48 dB SSIM 0.981 Checkpoint unavailable	0.882	38.48	0.981	✓ Certified	Wei et al., ECCV 2025
🥈	DiffDeconv DiffDeconv Huang et al., NeurIPS 2024 38.12 dB SSIM 0.979 Checkpoint unavailable	0.875	38.12	0.979	✓ Certified	Huang et al., NeurIPS 2024
🥉	Restormer+ Restormer+ Zamir et al., ICCV 2024 37.65 dB SSIM 0.975 Checkpoint unavailable	0.865	37.65	0.975	✓ Certified	Zamir et al., ICCV 2024
4	DeconvFormer DeconvFormer Chen et al., CVPR 2024 37.25 dB SSIM 0.972 Checkpoint unavailable	0.857	37.25	0.972	✓ Certified	Chen et al., CVPR 2024
5	ResUNet ResUNet DeCelle et al., Nat. Methods 2021 35.85 dB SSIM 0.964 Checkpoint unavailable	0.830	35.85	0.964	✓ Certified	DeCelle et al., Nat. Methods 2021
6	Restormer Restormer Zamir et al., CVPR 2022 35.8 dB SSIM 0.962 Checkpoint unavailable	0.828	35.8	0.962	✓ Certified	Zamir et al., CVPR 2022
7	U-Net U-Net Ronneberger et al., MICCAI 2015 35.15 dB SSIM 0.956 Checkpoint unavailable	0.814	35.15	0.956	✓ Certified	Ronneberger et al., MICCAI 2015
8	CARE CARE Weigert et al., Nat. Methods 2018 34.5 dB SSIM 0.948 Checkpoint unavailable	0.799	34.5	0.948	✓ Certified	Weigert et al., Nat. Methods 2018
9	PnP-DnCNN PnP-DnCNN Zhang et al., IEEE TIP 2017 31.2 dB SSIM 0.890 Try in SpecLab →	0.715	31.2	0.890	✓ Certified	Zhang et al., IEEE TIP 2017
10	PnP-FISTA PnP-FISTA Bai et al., 2020 30.42 dB SSIM 0.872 Try in SpecLab →	0.693	30.42	0.872	✓ Certified	Bai et al., 2020
11	TV-Deconvolution TV-Deconvolution TV-regularized deconvolution 29.5 dB SSIM 0.845 Try in SpecLab →	0.664	29.5	0.845	✓ Certified	TV-regularized deconvolution
12	Wiener Filter Wiener Filter Analytical baseline 28.35 dB SSIM 0.805 Try in SpecLab →	0.625	28.35	0.805	✓ Certified	Analytical baseline
13	Richardson-Lucy Richardson-Lucy Richardson 1972 / Lucy 1974 27.1 dB SSIM 0.770 Try in SpecLab →	0.587	27.1	0.770	✓ Certified	Richardson 1972 / Lucy 1974

Dataset: PWM Benchmark (13 algorithms)

Blind Reconstruction Challenge — forward model has unknown mismatch, must calibrate from data. Score = 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖)

#	Method	Overall Score	Public PSNR / SSIM	Dev PSNR / SSIM	Hidden PSNR / SSIM	Trust	Source
🥇	DeconvFormer + gradient DeconvFormer + gradient Chen et al., CVPR 2024 Score 0.785 Correct & Reconstruct →	0.785	0.814 35.05 dB / 0.968	0.785 33.93 dB / 0.960	0.757 31.19 dB / 0.933	✓ Certified	Chen et al., CVPR 2024
🥈	ScoreMicro + gradient ScoreMicro + gradient Wei et al., ECCV 2025 Score 0.772 Correct & Reconstruct →	0.772	0.829 36.43 dB / 0.976	0.765 31.97 dB / 0.942	0.721 29.62 dB / 0.911	✓ Certified	Wei et al., ECCV 2025
🥉	DiffDeconv + gradient DiffDeconv + gradient Huang et al., NeurIPS 2024 Score 0.764 Correct & Reconstruct →	0.764	0.824 36.07 dB / 0.974	0.753 30.65 dB / 0.926	0.715 28.82 dB / 0.897	✓ Certified	Huang et al., NeurIPS 2024
4	Restormer+ + gradient Restormer+ + gradient Zamir et al., ICCV 2024 Score 0.754 Correct & Reconstruct →	0.754	0.817 35.16 dB / 0.969	0.755 31.56 dB / 0.938	0.691 27.17 dB / 0.862	✓ Certified	Zamir et al., ICCV 2024
5	ResUNet + gradient ResUNet + gradient DeCelle et al., Nat. Methods 2021 Score 0.734 Correct & Reconstruct →	0.734	0.797 33.91 dB / 0.960	0.715 28.46 dB / 0.890	0.689 27.84 dB / 0.877	✓ Certified	DeCelle et al., Nat. Methods 2021
6	U-Net + gradient U-Net + gradient Ronneberger et al., MICCAI 2015 Score 0.724 Correct & Reconstruct →	0.724	0.808 33.66 dB / 0.958	0.707 28.15 dB / 0.884	0.656 25.88 dB / 0.829	✓ Certified	Ronneberger et al., MICCAI 2015
7	Restormer + gradient Restormer + gradient Zamir et al., CVPR 2022 Score 0.718 Correct & Reconstruct →	0.718	0.793 33.11 dB / 0.954	0.719 29.51 dB / 0.909	0.643 25.83 dB / 0.827	✓ Certified	Zamir et al., CVPR 2022
8	CARE + gradient CARE + gradient Weigert et al., Nat. Methods 2018 Score 0.679 Correct & Reconstruct →	0.679	0.779 32.39 dB / 0.947	0.652 25.3 dB / 0.812	0.606 23.51 dB / 0.751	✓ Certified	Weigert et al., Nat. Methods 2018
9	TV-Deconvolution + gradient TV-Deconvolution + gradient Rudin et al., Phys. A 1992 Score 0.676 Correct & Reconstruct →	0.676	0.691 27.01 dB / 0.858	0.672 26.37 dB / 0.842	0.664 26.73 dB / 0.851	✓ Certified	Rudin et al., Phys. A 1992
10	PnP-DnCNN + gradient PnP-DnCNN + gradient Zhang et al., IEEE TIP 2017 Score 0.670 Correct & Reconstruct →	0.670	0.748 29.47 dB / 0.908	0.653 25.81 dB / 0.827	0.608 24.17 dB / 0.775	✓ Certified	Zhang et al., IEEE TIP 2017
11	PnP-FISTA + gradient PnP-FISTA + gradient Bai et al., 2020 Score 0.650 Correct & Reconstruct →	0.650	0.707 27.59 dB / 0.872	0.638 25.41 dB / 0.815	0.606 23.32 dB / 0.743	✓ Certified	Bai et al., 2020
12	Wiener Filter + gradient Wiener Filter + gradient Analytical baseline Score 0.640 Correct & Reconstruct →	0.640	0.672 26.03 dB / 0.833	0.633 24.39 dB / 0.782	0.615 24.39 dB / 0.782	✓ Certified	Analytical baseline
13	Richardson-Lucy + gradient Richardson-Lucy + gradient Richardson, JOSA 1972 / Lucy, AJ 1974 Score 0.599 Correct & Reconstruct →	0.599	0.640 24.53 dB / 0.787	0.615 23.49 dB / 0.750	0.541 21.26 dB / 0.657	✓ Certified	Richardson, JOSA 1972 / Lucy, AJ 1974

Complete score requires all 3 tiers (Public + Dev + Hidden).

Join the competition →

Scoring: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖) PSNR 40% · SSIM 40% · Consistency 20%

Public 5 scenes

Full-access development tier with all data visible.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), spec ranges, ground truth (x_true), and true mismatch spec.

How to use: Load HDF5 → compare reconstruction vs x_true → check consistency → iterate.

What to submit: Reconstructed signals (x_hat) and corrected spec as HDF5.

Public Leaderboard

#	Method	Score	PSNR	SSIM
1	ScoreMicro + gradient	0.829	36.43	0.976
2	DiffDeconv + gradient	0.824	36.07	0.974
3	Restormer+ + gradient	0.817	35.16	0.969
4	DeconvFormer + gradient	0.814	35.05	0.968
5	U-Net + gradient	0.808	33.66	0.958
6	ResUNet + gradient	0.797	33.91	0.96
7	Restormer + gradient	0.793	33.11	0.954
8	CARE + gradient	0.779	32.39	0.947
9	PnP-DnCNN + gradient	0.748	29.47	0.908
10	PnP-FISTA + gradient	0.707	27.59	0.872
11	TV-Deconvolution + gradient	0.691	27.01	0.858
12	Wiener Filter + gradient	0.672	26.03	0.833
13	Richardson-Lucy + gradient	0.640	24.53	0.787

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
psf_sigma	-10.0	20.0	%
photon_budget	-20.0	40.0	%
read_noise	0.5	3.5	e-

View Details →

Dev 5 scenes

Blind evaluation tier — no ground truth available.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), and spec ranges only.

How to use: Apply your pipeline from the Public tier. Use consistency as self-check.

What to submit: Reconstructed signals and corrected spec. Scored server-side.

Dev Leaderboard

#	Method	Score	PSNR	SSIM
1	DeconvFormer + gradient	0.785	33.93	0.96
2	ScoreMicro + gradient	0.765	31.97	0.942
3	Restormer+ + gradient	0.755	31.56	0.938
4	DiffDeconv + gradient	0.753	30.65	0.926
5	Restormer + gradient	0.719	29.51	0.909
6	ResUNet + gradient	0.715	28.46	0.89
7	U-Net + gradient	0.707	28.15	0.884
8	TV-Deconvolution + gradient	0.672	26.37	0.842
9	PnP-DnCNN + gradient	0.653	25.81	0.827
10	CARE + gradient	0.652	25.3	0.812
11	PnP-FISTA + gradient	0.638	25.41	0.815
12	Wiener Filter + gradient	0.633	24.39	0.782
13	Richardson-Lucy + gradient	0.615	23.49	0.75

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
psf_sigma	-12.0	18.0	%
photon_budget	-24.0	36.0	%
read_noise	0.3	3.3	e-

Submit Reconstruction View Details →

Hidden 5 scenes

Fully blind server-side evaluation — no data download.

What you get & how to use

What you get: No data downloadable. Algorithm runs server-side on hidden measurements.

How to use: Package algorithm as Docker container / Python script. Submit via link.

What to submit: Containerized algorithm accepting y + H, outputting x_hat + corrected spec.

Hidden Leaderboard

#	Method	Score	PSNR	SSIM
1	DeconvFormer + gradient	0.757	31.19	0.933
2	ScoreMicro + gradient	0.721	29.62	0.911
3	DiffDeconv + gradient	0.715	28.82	0.897
4	Restormer+ + gradient	0.691	27.17	0.862
5	ResUNet + gradient	0.689	27.84	0.877
6	TV-Deconvolution + gradient	0.664	26.73	0.851
7	U-Net + gradient	0.656	25.88	0.829
8	Restormer + gradient	0.643	25.83	0.827
9	Wiener Filter + gradient	0.615	24.39	0.782
10	PnP-DnCNN + gradient	0.608	24.17	0.775
11	CARE + gradient	0.606	23.51	0.751
12	PnP-FISTA + gradient	0.606	23.32	0.743
13	Richardson-Lucy + gradient	0.541	21.26	0.657

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
psf_sigma	-7.0	23.0	%
photon_budget	-14.0	46.0	%
read_noise	0.8	3.8	e-

Submit Algorithm View Details →

Blind Reconstruction Challenge

Challenge

Given measurements with unknown mismatch and spec ranges (not exact params), reconstruct the original signal. A method must be evaluated on all three tiers for a complete score. Scored on a composite metric: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖).

Input

Measurements y, ideal forward model H, spec ranges

Output

Reconstructed signal x̂

About the Imaging Modality

Widefield fluorescence microscopy operated at very low illumination power or short exposure time to reduce phototoxicity and photobleaching in live specimens. Images are dominated by shot noise (Poisson) and read noise (Gaussian) with typical photon counts of 20-200 per pixel. The forward model is y = Poisson(alpha * PSF ** x)/alpha + N(0, sigma^2) where alpha is the photon conversion factor. Reconstruction requires joint denoising and deconvolution using PnP-HQS, Noise2Void, or CARE.

Principle

Identical optical path to standard widefield but operated at very low photon budgets (short exposure or attenuated excitation) to minimize phototoxicity in live cells. The acquired images are severely photon-starved, making Poisson noise the dominant degradation rather than out-of-focus blur.

How to Build the System

Use the same widefield microscope but reduce LED power to 1-5 % and/or shorten exposure to 5-20 ms. A high-QE back-illuminated sCMOS sensor (>80 % QE) is essential for capturing the limited photon signal. Install an environmental chamber for live-cell stability (37 °C, 5 % CO₂). Validate that the camera read noise floor is well below the expected signal.

Common Reconstruction Algorithms

CARE (Content-Aware image REstoration)
Noise2Void / Noise2Self (self-supervised denoising)
BM3D / VST + BM3D for Poisson-Gaussian denoising
PURE-LET (Poisson Unbiased Risk Estimator)
Noise2Noise paired denoising networks

Common Mistakes

Setting read-noise-dominated regime by using too-low gain or old CCD
Training denoising networks on data with different noise statistics than test data
Clipping near-zero intensities by incorrect camera offset subtraction
Ignoring sCMOS pixel-dependent noise (fixed-pattern noise)
Exceeding live-cell phototoxicity budget despite intending low-dose imaging

How to Avoid Mistakes

Characterize camera noise model (gain, offset, variance map) before acquisition
Train and evaluate denoising models at the same SNR and microscope settings
Keep camera offset (dark current) calibration current and subtract properly
Apply per-pixel gain and offset maps for sCMOS cameras
Monitor cell health markers (morphology, division rate) to confirm non-toxic dose

Forward-Model Mismatch Cases

The widefield fallback applies the correct blur kernel but uses a Gaussian noise model, whereas low-dose imaging is dominated by Poisson shot noise with very few photons per pixel
Denoising algorithms trained on Gaussian noise statistics will underperform on Poisson-dominated low-dose data, producing biased estimates and residual artifacts

How to Correct the Mismatch

Use the low-dose widefield operator that applies a Poisson-Gaussian noise model: y = Poisson(alpha * PSF ** x) / alpha + N(0, sigma^2)
Train or select denoising algorithms that explicitly model Poisson statistics (Anscombe transform + BM3D, or Poisson-aware deep networks like Noise2Void)

Experimental Setup — Signal Chain

Experimental setup diagram for Low-Dose Widefield Microscopy

Reconstruction Gallery — 4 Scenes × 3 Scenarios

Method: CPU_baseline | Mismatch: nominal (nominal=True, perturbed=False)

I Ideal II Mismatch (uncorrected) III Oracle correction

I — Ideal

Ground Truth

Measurement

Reconstruction

17.022206064497794 dB SSIM 0.38380508849556166

II — Mismatch

Ground Truth

Measurement

Reconstruction

14.93801985037615 dB SSIM 0.24399594480247574

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

20.039527108294998 dB SSIM 0.5012899560940253

I — Ideal

Ground Truth

Measurement

Reconstruction

14.231375949985203 dB SSIM 0.29689664293183293

II — Mismatch

Ground Truth

Measurement

Reconstruction

12.92413737908083 dB SSIM 0.20980780569697918

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

19.225830712320835 dB SSIM 0.547962131323724

I — Ideal

Ground Truth

Measurement

Reconstruction

8.520061941863302 dB SSIM 0.3871062840596513

II — Mismatch

Ground Truth

Measurement

Reconstruction

8.013928482144143 dB SSIM 0.23077294300010404

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

20.113884181775745 dB SSIM 0.3416888134023018

I — Ideal

Ground Truth

Measurement

Reconstruction

13.297073801541249 dB SSIM 0.5284335563352818

II — Mismatch

Ground Truth

Measurement

Reconstruction

11.66205075678512 dB SSIM 0.26941399804629346

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

19.61468818224941 dB SSIM 0.44080752632935544

Mean PSNR Across All Scenes

13.267679439471888 dB

I — Ideal

11.884534117096562 dB

II — Mismatch

19.748482546160247 dB

III — Oracle

Degradation: -1.4 dB Recovery: +7.9 dB

Per-scene PSNR breakdown (4 scenes)

Scene	I (PSNR)	I (SSIM)	II (PSNR)	II (SSIM)	III (PSNR)	III (SSIM)
scene_00	17.022206064497794	0.38380508849556166	14.93801985037615	0.24399594480247574	20.039527108294998	0.5012899560940253
scene_01	14.231375949985203	0.29689664293183293	12.92413737908083	0.20980780569697918	19.225830712320835	0.547962131323724
scene_02	8.520061941863302	0.3871062840596513	8.013928482144143	0.23077294300010404	20.113884181775745	0.3416888134023018
scene_03	13.297073801541249	0.5284335563352818	11.66205075678512	0.26941399804629346	19.61468818224941	0.44080752632935544
Mean	13.267679439471888	0.39906039295558193	11.884534117096562	0.23849767288646312	19.748482546160247	0.45793710678735156

Experimental Setup

Instrument: Nikon Eclipse Ti2-E / Zeiss Axio Observer 7

Objective: Plan Apo 60x / 1.40 NA oil immersion

Pixel Size Nm: 65

Excitation Source: LED (attenuated to 2 mW, 4% power)

Excitation Nm: 488

Emission Nm: 520

Exposure Ms: 5

Photon Budget: 50-200 photons/pixel

Detector: Hamamatsu ORCA-Flash4.0 V3 sCMOS

Reconstruction: PnP-HQS / Noise2Void / CARE

Key References

Krull et al., 'Noise2Void - Learning Denoising from Single Noisy Images', CVPR 2019
Weigert et al., 'Content-aware image restoration (CARE)', Nature Methods 15, 1090-1097 (2018)

Canonical Datasets

BioSR low-SNR subset
Planaria / Tribolium datasets (Weigert et al.)

Spec DAG — Forward Model Pipeline

C(PSF) → D(g, η₃)

C PSF Convolution (PSF)

D sCMOS Camera (g, η₃)

Mismatch Parameters

Symbol	Parameter	Description	Nominal	Perturbed
Δσ	psf_sigma	PSF width error (%)	0	10.0
ΔN	photon_budget	Photon budget error (%)	0	20.0
Δσ_r	read_noise	Read noise error (e-)	1.5	2.5

Credits System

40%

Platform Profit Pool

Revenue allocated to benchmark rewards

30%

Winner Share

Top algorithm receives from pool

$100

Min Withdrawal

Minimum payout threshold

Spec Primitives Reference (11 primitives)

P Propagation

Free-space or medium propagation kernel (Fresnel, Rayleigh-Sommerfeld).

M Mask / Modulation

Spatial or spatio-temporal amplitude modulation (coded aperture, SLM pattern).

Π Projection

Geometric projection operator (Radon transform, fan-beam, cone-beam).

F Fourier Sampling

Sampling in the Fourier / k-space domain (MRI, ptychography).

C Convolution

Shift-invariant convolution with a point-spread function (PSF).

Σ Summation / Integration

Summation along a physical dimension (spectral, temporal, angular).

D Detector

Sensor readout with gain g and noise model η (Gaussian, Poisson, mixed).

S Structured Illumination

Patterned illumination (block, Hadamard, random) applied to the scene.

W Wavelength Dispersion

Spectral dispersion element (prism, grating) with shift α and aperture a.

R Rotation / Motion

Sample or gantry rotation (CT, electron tomography).

Λ Wavelength Selection

Spectral filter or monochromator selecting a wavelength band.