OCT

Optical Coherence Tomography

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
🥇	ScoreOCT ScoreOCT Wei et al., ECCV 2025 37.95 dB SSIM 0.973 Checkpoint unavailable	0.869	37.95	0.973	✓ Certified	Wei et al., ECCV 2025
🥈	DiffusionOCT DiffusionOCT Zhang et al., NeurIPS 2024 37.52 dB SSIM 0.970 Checkpoint unavailable	0.860	37.52	0.970	✓ Certified	Zhang et al., NeurIPS 2024
🥉	SpeckleFormer SpeckleFormer Devalla et al., ECCV 2024 36.85 dB SSIM 0.964 Checkpoint unavailable	0.846	36.85	0.964	✓ Certified	Devalla et al., ECCV 2024
4	RetinalFormer RetinalFormer Chen et al., ICCV 2024 36.35 dB SSIM 0.960 Checkpoint unavailable	0.836	36.35	0.960	✓ Certified	Chen et al., ICCV 2024
5	OCT-ViT OCT-ViT Tian et al., ICCV 2024 36.12 dB SSIM 0.958 Checkpoint unavailable	0.831	36.12	0.958	✓ Certified	Tian et al., ICCV 2024
6	OCTA-Net OCTA-Net Hybrid U-Net+Transformer, 2023 34.6 dB SSIM 0.942 Checkpoint unavailable	0.798	34.6	0.942	✓ Certified	Hybrid U-Net+Transformer, 2023
7	U-Net-OCT U-Net-OCT U-Net variant 33.85 dB SSIM 0.935 Checkpoint unavailable	0.782	33.85	0.935	✓ Certified	U-Net variant
8	Speckle-DenoiseNet Speckle-DenoiseNet Devalla et al., BOE 2019 33.1 dB SSIM 0.925 Checkpoint unavailable	0.764	33.1	0.925	✓ Certified	Devalla et al., BOE 2019
9	NLM-OCT NLM-OCT Non-local means variant 30.2 dB SSIM 0.870 Try in SpecLab →	0.688	30.2	0.870	✓ Certified	Non-local means variant
10	BM4D BM4D Maggioni et al., IEEE TIP 2013 29.3 dB SSIM 0.850 Try in SpecLab →	0.663	29.3	0.850	✓ Certified	Maggioni et al., IEEE TIP 2013
11	TV-Denoising TV-Denoising TV regularization 28.5 dB SSIM 0.815 Try in SpecLab →	0.632	28.5	0.815	✓ Certified	TV regularization
12	Speckle-Lee Speckle-Lee Lee, IEEE TGRS 1980 27.85 dB SSIM 0.790 Try in SpecLab →	0.609	27.85	0.790	✓ Certified	Lee, IEEE TGRS 1980
13	FFT-OCT FFT-OCT Analytical baseline 25.6 dB SSIM 0.720 Try in SpecLab →	0.537	25.6	0.720	✓ Certified	Analytical baseline

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
🥇	SpeckleFormer + gradient SpeckleFormer + gradient Devalla et al., ECCV 2024 Score 0.773 Correct & Reconstruct →	0.773	0.806 34.31 dB / 0.963	0.783 33.26 dB / 0.955	0.731 29.87 dB / 0.915	✓ Certified	Devalla et al., ECCV 2024
🥈	RetinalFormer + gradient RetinalFormer + gradient Chen et al., ICCV 2024 Score 0.771 Correct & Reconstruct →	0.771	0.824 34.92 dB / 0.967	0.765 31.56 dB / 0.938	0.723 30.02 dB / 0.917	✓ Certified	Chen et al., ICCV 2024
🥉	OCT-ViT + gradient OCT-ViT + gradient Tian et al., ICCV 2024 Score 0.770 Correct & Reconstruct →	0.770	0.819 34.39 dB / 0.964	0.768 31.66 dB / 0.939	0.723 30.29 dB / 0.921	✓ Certified	Tian et al., ICCV 2024
4	ScoreOCT + gradient ScoreOCT + gradient Wei et al., ECCV 2025 Score 0.769 Correct & Reconstruct →	0.769	0.843 36.51 dB / 0.976	0.759 30.92 dB / 0.930	0.705 29.4 dB / 0.907	✓ Certified	Wei et al., ECCV 2025
5	DiffusionOCT + gradient DiffusionOCT + gradient Zhang et al., NeurIPS 2024 Score 0.746 Correct & Reconstruct →	0.746	0.837 36.33 dB / 0.975	0.719 28.61 dB / 0.893	0.682 27.32 dB / 0.866	✓ Certified	Zhang et al., NeurIPS 2024
6	OCTA-Net + gradient OCTA-Net + gradient Hybrid U-Net+Transformer, 2023 Score 0.672 Correct & Reconstruct →	0.672	0.777 32.09 dB / 0.944	0.645 25.08 dB / 0.805	0.593 22.87 dB / 0.726	✓ Certified	Hybrid U-Net+Transformer, 2023
7	Speckle-DenoiseNet + gradient Speckle-DenoiseNet + gradient Devalla et al., BOE 2019 Score 0.662 Correct & Reconstruct →	0.662	0.752 30.11 dB / 0.918	0.637 25.37 dB / 0.814	0.596 23.87 dB / 0.764	✓ Certified	Devalla et al., BOE 2019
8	Speckle-Lee + gradient Speckle-Lee + gradient Lee, IEEE TGRS 1980 Score 0.653 Correct & Reconstruct →	0.653	0.655 25.19 dB / 0.808	0.654 25.24 dB / 0.810	0.651 25.71 dB / 0.824	✓ Certified	Lee, IEEE TGRS 1980
9	TV-Denoising + gradient TV-Denoising + gradient Rudin et al., Phys. A 1992 Score 0.645 Correct & Reconstruct →	0.645	0.706 27.48 dB / 0.869	0.633 25.0 dB / 0.802	0.595 22.86 dB / 0.726	✓ Certified	Rudin et al., Phys. A 1992
10	BM4D + gradient BM4D + gradient Maggioni et al., IEEE TIP 2013 Score 0.643 Correct & Reconstruct →	0.643	0.687 26.8 dB / 0.853	0.629 24.39 dB / 0.782	0.612 23.51 dB / 0.751	✓ Certified	Maggioni et al., IEEE TIP 2013
11	U-Net-OCT + gradient U-Net-OCT + gradient Ronneberger et al., MICCAI 2015 (OCT variant) Score 0.639 Correct & Reconstruct →	0.639	0.792 32.8 dB / 0.951	0.619 24.32 dB / 0.780	0.507 20.6 dB / 0.627	✓ Certified	Ronneberger et al., MICCAI 2015 (OCT variant)
12	NLM-OCT + gradient NLM-OCT + gradient Buades et al., Multiscale Model. Simul. 2005 Score 0.628 Correct & Reconstruct →	0.628	0.730 28.44 dB / 0.890	0.632 24.68 dB / 0.792	0.521 20.32 dB / 0.614	✓ Certified	Buades et al., Multiscale Model. Simul. 2005
13	FFT-OCT + gradient FFT-OCT + gradient Analytical baseline Score 0.588 Correct & Reconstruct →	0.588	0.612 23.61 dB / 0.754	0.612 23.76 dB / 0.760	0.539 21.62 dB / 0.673	✓ Certified	Analytical baseline

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 3 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	ScoreOCT + gradient	0.843	36.51	0.976
2	DiffusionOCT + gradient	0.837	36.33	0.975
3	RetinalFormer + gradient	0.824	34.92	0.967
4	OCT-ViT + gradient	0.819	34.39	0.964
5	SpeckleFormer + gradient	0.806	34.31	0.963
6	U-Net-OCT + gradient	0.792	32.8	0.951
7	OCTA-Net + gradient	0.777	32.09	0.944
8	Speckle-DenoiseNet + gradient	0.752	30.11	0.918
9	NLM-OCT + gradient	0.730	28.44	0.89
10	TV-Denoising + gradient	0.706	27.48	0.869
11	BM4D + gradient	0.687	26.8	0.853
12	Speckle-Lee + gradient	0.655	25.19	0.808
13	FFT-OCT + gradient	0.612	23.61	0.754

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
dispersion	-200.0	400.0	fs²
reference_delay	-5.0	10.0	μm
spectral_roll_off	-1.0	2.0	dB/mm

View Details →

Dev 3 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	SpeckleFormer + gradient	0.783	33.26	0.955
2	OCT-ViT + gradient	0.768	31.66	0.939
3	RetinalFormer + gradient	0.765	31.56	0.938
4	ScoreOCT + gradient	0.759	30.92	0.93
5	DiffusionOCT + gradient	0.719	28.61	0.893
6	Speckle-Lee + gradient	0.654	25.24	0.81
7	OCTA-Net + gradient	0.645	25.08	0.805
8	Speckle-DenoiseNet + gradient	0.637	25.37	0.814
9	TV-Denoising + gradient	0.633	25.0	0.802
10	NLM-OCT + gradient	0.632	24.68	0.792
11	BM4D + gradient	0.629	24.39	0.782
12	U-Net-OCT + gradient	0.619	24.32	0.78
13	FFT-OCT + gradient	0.612	23.76	0.76

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
dispersion	-240.0	360.0	fs²
reference_delay	-6.0	9.0	μm
spectral_roll_off	-1.2	1.8	dB/mm

Submit Reconstruction View Details →

Hidden 3 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	SpeckleFormer + gradient	0.731	29.87	0.915
2	RetinalFormer + gradient	0.723	30.02	0.917
3	OCT-ViT + gradient	0.723	30.29	0.921
4	ScoreOCT + gradient	0.705	29.4	0.907
5	DiffusionOCT + gradient	0.682	27.32	0.866
6	Speckle-Lee + gradient	0.651	25.71	0.824
7	BM4D + gradient	0.612	23.51	0.751
8	Speckle-DenoiseNet + gradient	0.596	23.87	0.764
9	TV-Denoising + gradient	0.595	22.86	0.726
10	OCTA-Net + gradient	0.593	22.87	0.726
11	FFT-OCT + gradient	0.539	21.62	0.673
12	NLM-OCT + gradient	0.521	20.32	0.614
13	U-Net-OCT + gradient	0.507	20.6	0.627

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
dispersion	-140.0	460.0	fs²
reference_delay	-3.5	11.5	μm
spectral_roll_off	-0.7	2.3	dB/mm

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

OCT is a low-coherence interferometric imaging technique that measures depth-resolved backscattering profiles (A-scans) by interfering sample-arm reflections with a reference mirror. In spectral-domain OCT, the interference spectrum is recorded by a spectrometer and the axial profile is obtained via Fourier transform. Axial resolution is determined by the source bandwidth (typically 3-7 um in tissue) and imaging depth by spectrometer resolution. Dominant artifacts include speckle noise, motion artifacts, and sensitivity roll-off with depth.

Principle

Optical Coherence Tomography uses low-coherence interferometry to produce cross-sectional images of tissue microstructure. A broadband light source (superluminescent diode, ~840 nm or ~1310 nm) is split between sample and reference arms. Interference occurs only when the path lengths match within the coherence length (~5-10 μm), providing axial resolution. Spectral-domain OCT records the spectral interferogram and uses FFT for fast depth-resolved imaging.

How to Build the System

Build or acquire a spectral-domain OCT system: broadband SLD source (center 840 nm, 50 nm bandwidth for retinal; 1310 nm for dermal/cardiac), fiber-based Michelson interferometer, galvo scanner for lateral scanning, and a spectrometer with line camera (2048-4096 pixels) for spectral detection. Calibrate wavelength-to-wavenumber mapping, dispersion compensation, and reference arm delay. For swept-source OCT, use a frequency-swept laser (100-400 kHz sweep rate) and balanced detector.

Common Reconstruction Algorithms

FFT-based spectral-domain OCT reconstruction (spectral interferogram → A-scan)
Dispersion compensation (numerical or hardware)
Speckle reduction (spatial/angular compounding, or deep-learning)
Segmentation of retinal layers (graph-based, U-Net, or transformer models)
OCT Angiography (OCTA) via decorrelation or phase-variance of repeated B-scans

Common Mistakes

Dispersion mismatch between sample and reference arms degrading axial resolution
Mirror image artifact from complex conjugate ambiguity in SD-OCT
Sensitivity roll-off at deeper imaging depths not compensated
Motion artifacts in 3-D OCT volumes (eye motion for ophthalmic OCT)
Incorrect refractive index assumption for depth scale calibration

How to Avoid Mistakes

Match fiber lengths and add numerical dispersion compensation in reconstruction
Place the zero-delay near the sample surface; use full-range OCT if needed
Use swept-source OCT for reduced roll-off; optimize spectrometer for uniform sensitivity
Apply eye-tracking or motion-correction algorithms; average repeated B-scans
Calibrate depth scale with a known-thickness reference standard

Forward-Model Mismatch Cases

The widefield fallback applies spatial blur, but OCT acquires spectral interferograms that encode depth via low-coherence interferometry — the interference fringe pattern bears no resemblance to a blurred image
OCT depth resolution comes from the broadband source coherence length (~5-10 um), not from spatial PSF — the widefield operator cannot model the axial sectioning, dispersion, or spectral-to-depth FFT relationship

How to Correct the Mismatch

Use the OCT operator that models spectral-domain interferometry: y(k) = |E_ref + E_sample(k)|^2, where depth information is encoded in the spectral fringe frequency
Reconstruct A-scans via FFT of the spectral interferogram after dispersion compensation and k-linearization; B-scans are formed by lateral scanning

Experimental Setup — Signal Chain

Experimental setup diagram for Optical Coherence Tomography

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

25.174717813548444 dB SSIM 0.8355680560955927

II — Mismatch

Ground Truth

Measurement

Reconstruction

20.543864852005157 dB SSIM 0.6765135454458621

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

19.600940235091095 dB SSIM 0.6342307678718445

I — Ideal

Ground Truth

Measurement

Reconstruction

24.158380010327747 dB SSIM 0.8259076877222974

II — Mismatch

Ground Truth

Measurement

Reconstruction

20.05261602729396 dB SSIM 0.678279400385818

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

18.858046846209472 dB SSIM 0.6294072205236141

I — Ideal

Ground Truth

Measurement

Reconstruction

22.7018034077921 dB SSIM 0.7591982367538689

II — Mismatch

Ground Truth

Measurement

Reconstruction

18.353446916031444 dB SSIM 0.6360522397070137

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

17.48278963227588 dB SSIM 0.5424081139820738

I — Ideal

Ground Truth

Measurement

Reconstruction

25.711598825518927 dB SSIM 0.8596903392549893

II — Mismatch

Ground Truth

Measurement

Reconstruction

20.686385980179104 dB SSIM 0.7166397560073531

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

19.789852575390306 dB SSIM 0.677009872001462

Mean PSNR Across All Scenes

24.4366250142968 dB

I — Ideal

19.90907844387742 dB

II — Mismatch

18.93290732224169 dB

III — Oracle

Degradation: -4.5 dB Recovery: +-1.0 dB

Per-scene PSNR breakdown (4 scenes)

Scene	I (PSNR)	I (SSIM)	II (PSNR)	II (SSIM)	III (PSNR)	III (SSIM)
scene_00	25.174717813548444	0.8355680560955927	20.543864852005157	0.6765135454458621	19.600940235091095	0.6342307678718445
scene_01	24.158380010327747	0.8259076877222974	20.05261602729396	0.678279400385818	18.858046846209472	0.6294072205236141
scene_02	22.7018034077921	0.7591982367538689	18.353446916031444	0.6360522397070137	17.48278963227588	0.5424081139820738
scene_03	25.711598825518927	0.8596903392549893	20.686385980179104	0.7166397560073531	19.789852575390306	0.677009872001462
Mean	24.4366250142968	0.8200910799566872	19.90907844387742	0.6768712353865117	18.93290732224169	0.6207639935947485

Experimental Setup

Instrument: Heidelberg Spectralis HRA+OCT / Zeiss Cirrus HD-OCT 5000

Wavelength Nm: 840

Bandwidth Nm: 45

Axial Resolution Um: 5

Lateral Resolution Um: 15

A Scan Rate Khz: 40

Scan Width Mm: 6.0

B Scan Lines: 512

A Scans Per B: 512

Snr Db: 98

Key References

Huang et al., 'Optical coherence tomography', Science 254, 1178 (1991)
de Boer et al., 'Twenty-five years of OCT', Biomed. Opt. Express 8, 3248 (2017)

Canonical Datasets

Duke SD-OCT DME dataset (Chiu et al.)
RETOUCH Challenge (retinal OCT)
OCTA-500 (Li et al., Scientific Data 2024)

Spec DAG — Forward Model Pipeline

P(low-coherence) → Σ(interference) → D(g, η₁)

P Low-Coherence Source (low-coherence)

Σ Interferometric Sum (interference)

D Spectrometer / Balanced Detector (g, η₁)

Mismatch Parameters

Symbol	Parameter	Description	Perturbed
ΔGVD	dispersion	Dispersion mismatch (fs²)	200
Δz_r	reference_delay	Reference delay error (μm)	5.0
ΔR	spectral_roll_off	Spectral roll-off error (dB/mm)	1.0

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.