Fundus

Fundus Camera

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
🥇 Swin-Fundus 0.790 34.2 0.940 ✓ Certified Chen et al., MICCAI 2023
🥈 cofe-Net 0.747 32.5 0.910 ✓ Certified Shen et al., IEEE TMI 2020
🥉 PnP-BM3D 0.645 28.8 0.830 ✓ Certified Danielyan et al., 2012
4 Richardson-Lucy 0.498 24.5 0.680 ✓ Certified Richardson 1972 / Lucy 1974

Dataset: PWM Benchmark (4 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
🥇 Swin-Fundus + gradient 0.684
0.795
32.67 dB / 0.950
0.673
26.39 dB / 0.843
0.585
23.49 dB / 0.750
✓ Certified Li et al., IEEE TMI 2023
🥈 cofe-Net + gradient 0.656
0.747
30.14 dB / 0.919
0.636
24.51 dB / 0.786
0.584
23.32 dB / 0.743
✓ Certified Shen et al., IEEE TMI 2020
🥉 PnP-BM3D + gradient 0.630
0.678
26.35 dB / 0.842
0.623
24.56 dB / 0.788
0.590
23.55 dB / 0.752
✓ Certified Danielyan et al., 2012
4 Richardson-Lucy + gradient 0.571
0.611
23.1 dB / 0.735
0.574
21.99 dB / 0.690
0.528
20.72 dB / 0.633
✓ Certified Richardson 1972 / Lucy 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 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 Swin-Fundus + gradient 0.795 32.67 0.95
2 cofe-Net + gradient 0.747 30.14 0.919
3 PnP-BM3D + gradient 0.678 26.35 0.842
4 Richardson-Lucy + gradient 0.611 23.1 0.735
Spec Ranges (3 parameters)
Parameter Min Max Unit
pupil_dilation -0.5 1.0 mm
focus -0.25 0.5 diopters
vignetting -5.0 10.0 %
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 Swin-Fundus + gradient 0.673 26.39 0.843
2 cofe-Net + gradient 0.636 24.51 0.786
3 PnP-BM3D + gradient 0.623 24.56 0.788
4 Richardson-Lucy + gradient 0.574 21.99 0.69
Spec Ranges (3 parameters)
Parameter Min Max Unit
pupil_dilation -0.6 0.9 mm
focus -0.3 0.45 diopters
vignetting -6.0 9.0 %
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 PnP-BM3D + gradient 0.590 23.55 0.752
2 Swin-Fundus + gradient 0.585 23.49 0.75
3 cofe-Net + gradient 0.584 23.32 0.743
4 Richardson-Lucy + gradient 0.528 20.72 0.633
Spec Ranges (3 parameters)
Parameter Min Max Unit
pupil_dilation -0.35 1.15 mm
focus -0.175 0.575 diopters
vignetting -3.5 11.5 %
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

A fundus camera captures a 2D color photograph of the retinal surface by illuminating the fundus through the pupil with a ring-shaped flash and imaging the reflected light through the central pupillary zone. The optical system images the curved retina onto a flat detector with 30-50 degree field of view. Image quality is degraded by media opacities (cataract), small pupil, and uneven illumination. Fundus images are widely used for automated screening of diabetic retinopathy, glaucoma, and AMD via deep learning.

Principle

A fundus camera images the posterior segment of the eye (retina, optic disc, macula, vasculature) by illuminating the retina through the pupil and capturing the reflected/backscattered light. The optical path is designed to separate illumination and observation through different portions of the pupil to avoid corneal reflections. Standard fundus imaging provides 30-50° field-of-view color photographs of the retina.

How to Build the System

Use a dedicated fundus camera (e.g., Topcon TRC-NW400, Canon CR-2 AF) or a scanning laser ophthalmoscope (Optos for widefield). Dilate the patient's pupil (tropicamide 1%) for standard fundus photography. Align the camera to center on the macula or optic disc. Set appropriate flash intensity and focus. Capture color and red-free (green channel) images. For fluorescein angiography, inject sodium fluorescein IV and capture timed image series with excitation/barrier filters.

Common Reconstruction Algorithms

  • Image quality assessment and auto-focus/auto-exposure
  • Vessel segmentation (U-Net, DeepVessel)
  • Optic disc and cup segmentation for glaucoma screening
  • Diabetic retinopathy grading (deep-learning classifiers)
  • Multi-frame averaging and super-resolution for fundus images

Common Mistakes

  • Insufficient pupil dilation causing vignetting at the field edges
  • Corneal reflections (flare) obscuring the central retinal image
  • Image out of focus due to refractive error not compensated
  • Eyelash or eyelid obstruction in the image
  • Uneven illumination across the retinal image

How to Avoid Mistakes

  • Ensure adequate mydriasis (>5 mm pupil diameter) before imaging
  • Align the camera carefully to separate illumination and observation through different pupil zones
  • Use auto-focus and compensate for patient refractive error in the camera optics
  • Ask patients to open eyes wide; use a fixation target for gaze direction
  • Verify uniform illumination before capture; adjust camera alignment if uneven

Forward-Model Mismatch Cases

  • The widefield fallback applies a generic Gaussian PSF, but fundus imaging has a unique optical path through the eye's optics (cornea and lens) with specific aberrations and the pupil-splitting illumination/observation geometry
  • The retinal image is formed after double-pass through the ocular media, with wavelength-dependent absorption (hemoglobin, melanin, macular pigment) — the widefield achromatic Gaussian blur cannot model spectral absorption or ocular aberrations

How to Correct the Mismatch

  • Use the fundus operator that models the eye's optical path: illumination through one pupil zone, retinal reflection/fluorescence, and collection through a separate pupil zone, with ocular aberration and media absorption
  • Include wavelength-dependent retinal reflectance for color fundus imaging, or fluorescein excitation/emission model for fluorescein angiography

Experimental Setup — Signal Chain

Experimental setup diagram for Fundus Camera

Experimental Setup

Instrument: Topcon TRC-NW400 / Canon CR-2 AF
Image Size: 2124x2056
Field Of View Deg: 45
Wavelength Range Nm: 500-700
Flash Exposure Ms: 0.04
Dataset: EyePACS, DRIVE, MESSIDOR, APTOS

Key References

  • Gulshan et al., 'Development and validation of a deep learning algorithm for detection of diabetic retinopathy', JAMA 316, 2402 (2016)
  • Staal et al., 'Ridge-based vessel segmentation (DRIVE)', IEEE TMI 23, 501 (2004)

Canonical Datasets

  • EyePACS (diabetic retinopathy screening)
  • DRIVE (Digital Retinal Images for Vessel Extraction)
  • MESSIDOR-2
  • APTOS 2019 Blindness Detection

Spec DAG — Forward Model Pipeline

C(PSF_optic) → D(g, η₁)

C Ophthalmic PSF (PSF_optic)
D CCD Camera (g, η₁)

Mismatch Parameters

Symbol Parameter Description Nominal Perturbed
Δd_p pupil_dilation Pupil dilation error (mm) 0 0.5
Δf focus Focus error (diopters) 0 0.25
Δv vignetting Vignetting uniformity error (%) 0 5.0

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.