Endoscopy

Fiber Bundle Endoscopy

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
🥇 DiffEndo 0.890 39.7 0.957 ✓ Certified Gao et al. 2024
🥈 PhysEndo 0.863 38.4 0.947 ✓ Certified Chen et al. 2024
🥉 SwinEndo 0.840 37.3 0.937 ✓ Certified Li et al. 2023
4 TransEndo 0.809 35.9 0.921 ✓ Certified Wang et al. 2022
5 EndoSLAM-Net 0.758 33.8 0.889 ✓ Certified Ozyoruk et al. 2021
6 DnCNN-Endo 0.701 31.4 0.855 ✓ Certified Zhang et al. 2017
7 BM3D-Endo 0.638 28.9 0.812 ✓ Certified Dabov et al. 2007
8 CLAHE-Endo 0.578 26.5 0.772 ✓ Certified Zuiderveld 1994
9 Histogram-Eq 0.521 24.1 0.738 ✓ Certified Gonzalez & Woods 2002

Dataset: PWM Benchmark (9 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
🥇 SwinEndo + gradient 0.779
0.813
34.66 dB / 0.966
0.786
32.78 dB / 0.951
0.739
30.27 dB / 0.921
✓ Certified Li et al., IEEE TMI 2023
🥈 TransEndo + gradient 0.774
0.820
34.89 dB / 0.967
0.774
32.61 dB / 0.949
0.727
29.29 dB / 0.905
✓ Certified Wang et al., Med. Image Anal. 2022
🥉 PhysEndo + gradient 0.769
0.826
35.58 dB / 0.971
0.755
31.73 dB / 0.940
0.726
29.32 dB / 0.906
✓ Certified Chen et al., Med. Image Anal. 2024
4 DiffEndo + gradient 0.757
0.841
36.77 dB / 0.977
0.749
30.13 dB / 0.919
0.681
26.69 dB / 0.850
✓ Certified Gao et al., MICCAI 2024
5 EndoSLAM-Net + gradient 0.680
0.791
32.53 dB / 0.948
0.662
26.08 dB / 0.834
0.586
23.15 dB / 0.737
✓ Certified Ozyoruk et al., Med. Image Anal. 2021
6 DnCNN-Endo + gradient 0.660
0.729
29.15 dB / 0.903
0.642
25.4 dB / 0.815
0.608
24.17 dB / 0.775
✓ Certified Zhang et al., IEEE TIP 2017
7 BM3D-Endo + gradient 0.648
0.683
26.69 dB / 0.850
0.649
24.99 dB / 0.802
0.611
24.25 dB / 0.777
✓ Certified Dabov et al., IEEE TIP 2007
8 CLAHE-Endo + gradient 0.609
0.663
25.34 dB / 0.813
0.597
23.42 dB / 0.747
0.566
22.35 dB / 0.705
✓ Certified Zuiderveld, Graphics Gems IV 1994
9 Histogram-Eq + gradient 0.563
0.597
22.41 dB / 0.707
0.568
22.16 dB / 0.697
0.523
20.96 dB / 0.644
✓ Certified Gonzalez & Woods, Digital Image Processing 2002

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 DiffEndo + gradient 0.841 36.77 0.977
2 PhysEndo + gradient 0.826 35.58 0.971
3 TransEndo + gradient 0.820 34.89 0.967
4 SwinEndo + gradient 0.813 34.66 0.966
5 EndoSLAM-Net + gradient 0.791 32.53 0.948
6 DnCNN-Endo + gradient 0.729 29.15 0.903
7 BM3D-Endo + gradient 0.683 26.69 0.85
8 CLAHE-Endo + gradient 0.663 25.34 0.813
9 Histogram-Eq + gradient 0.597 22.41 0.707
Spec Ranges (3 parameters)
Parameter Min Max Unit
fiber_coupling -5.0 10.0 %
core_spacing -0.5 1.0 μm
bending_loss -0.3 0.6 dB
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 SwinEndo + gradient 0.786 32.78 0.951
2 TransEndo + gradient 0.774 32.61 0.949
3 PhysEndo + gradient 0.755 31.73 0.94
4 DiffEndo + gradient 0.749 30.13 0.919
5 EndoSLAM-Net + gradient 0.662 26.08 0.834
6 BM3D-Endo + gradient 0.649 24.99 0.802
7 DnCNN-Endo + gradient 0.642 25.4 0.815
8 CLAHE-Endo + gradient 0.597 23.42 0.747
9 Histogram-Eq + gradient 0.568 22.16 0.697
Spec Ranges (3 parameters)
Parameter Min Max Unit
fiber_coupling -6.0 9.0 %
core_spacing -0.6 0.9 μm
bending_loss -0.36 0.54 dB
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 SwinEndo + gradient 0.739 30.27 0.921
2 TransEndo + gradient 0.727 29.29 0.905
3 PhysEndo + gradient 0.726 29.32 0.906
4 DiffEndo + gradient 0.681 26.69 0.85
5 BM3D-Endo + gradient 0.611 24.25 0.777
6 DnCNN-Endo + gradient 0.608 24.17 0.775
7 EndoSLAM-Net + gradient 0.586 23.15 0.737
8 CLAHE-Endo + gradient 0.566 22.35 0.705
9 Histogram-Eq + gradient 0.523 20.96 0.644
Spec Ranges (3 parameters)
Parameter Min Max Unit
fiber_coupling -3.5 11.5 %
core_spacing -0.35 1.15 μm
bending_loss -0.21 0.69 dB
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

Fiber bundle endoscopy transmits images through a coherent fiber bundle of 10,000-50,000 individual optical fibers. Each fiber core acts as a spatial sample, producing a honeycomb pattern. Image quality is limited by inter-core spacing (pixelation), inter-core coupling (crosstalk), and core-to-core transmission variation. White-light or narrow-band illumination is delivered through the bundle or alongside it. Reconstruction involves core localization, transmission calibration, interpolation to a regular grid, and denoising.

Principle

Fiber-bundle endoscopy transmits an image through a flexible coherent fiber bundle (10,000-100,000 individual fiber cores) to visualize internal body cavities. Each fiber core acts as a single pixel, transmitting light from the distal end to the proximal end where a camera captures the image. The hexagonal fiber packing imposes a fixed pixelation pattern (comb/honeycomb structure) on the image.

How to Build the System

A medical endoscope has a flexible insertion tube containing the coherent fiber bundle (or a distal CMOS chip for video endoscopes), illumination fibers, working channels, and air/water channels. Light source: LED or Xenon lamp transmitted through illumination fibers. For fiber-bundle type: attach a high-resolution camera and relay lens at the proximal end. Calibrate fiber core positions and individual fiber transmission for computational image improvement.

Common Reconstruction Algorithms

  • Fiber core mapping and interpolation (honeycomb artifact removal)
  • Deep-learning super-resolution for fiber-bundle images
  • Structure-from-motion for endoscopic 3-D reconstruction
  • Defogging / dehazing for underwater or smoke-obscured endoscopy
  • Real-time mosaicking for extended field-of-view endoscopy

Common Mistakes

  • Honeycomb pattern artifact from fiber core spacing not removed
  • Broken fibers (dark spots) accumulating over time and degrading image quality
  • Specular reflections (glare) from wet tissue surfaces saturating the image
  • Insufficient illumination causing noisy images in deep body cavities
  • Image distortion from fiber bundle bending not corrected

How to Avoid Mistakes

  • Apply fiber core interpolation or deep-learning super-resolution in post-processing
  • Replace fiber bundles when broken fiber percentage exceeds acceptable threshold
  • Use polarization filtering or computational specular removal algorithms
  • Use bright LED sources and adjust exposure/gain for adequate signal
  • Calibrate and correct for bending-dependent distortion using test patterns

Forward-Model Mismatch Cases

  • The widefield fallback produces a (64,64) image, but fiber-bundle endoscopy transmits images through discrete fiber cores creating a hexagonal pixelation pattern — output shape (n_fibers,) is a 1D vector of per-core intensities
  • The fiber bundle imposes a fixed sampling grid (honeycomb structure) with inter-core crosstalk and dead fibers — the widefield continuous Gaussian blur has no relationship to the discrete fiber sampling and transmission physics

How to Correct the Mismatch

  • Use the endoscopy operator that models per-fiber-core sampling: each of the ~10,000-100,000 cores transmits a point sample from the distal end to the proximal camera, with known core positions and transmission coefficients
  • Reconstruct using fiber-core interpolation, honeycomb artifact removal, or deep-learning super-resolution that account for the known fiber bundle geometry and per-core response

Experimental Setup — Signal Chain

Experimental setup diagram for Fiber Bundle Endoscopy

Experimental Setup

Instrument: Olympus GIF-H290Z / Karl Storz IMAGE1 S
Fiber Cores: 30000
Resolution: 1920x1080 (HD output)
Frame Rate Fps: 60
Field Of View Deg: 140
Working Channel Mm: 3.7
Wavelength Range Nm: 400-700 (white light)
Dataset: Kvasir, CVC-ClinicDB, HyperKvasir

Key References

  • Lee & Bhatt, 'Fiber bundle endoscopy advances', J. Biophotonics 12, e201900004 (2019)

Canonical Datasets

  • Kvasir-SEG (polyp segmentation)
  • CVC-ClinicDB (colonoscopy)
  • HyperKvasir (multi-class GI dataset)

Spec DAG — Forward Model Pipeline

C(PSF_fiber) → D(g, η₁)

C Fiber Bundle PSF (PSF_fiber)
D Camera (g, η₁)

Mismatch Parameters

Symbol Parameter Description Nominal Perturbed
Δη_f fiber_coupling Fiber coupling efficiency error (%) 0 5.0
Δd core_spacing Core spacing error (μm) 0 0.5
Δα_b bending_loss Bending loss error (dB) 0 0.3

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.