Confocal 3D

Confocal 3D Z-Stack

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
🥇 DiffusionMicro 0.896 39.9 0.963 ✓ Certified Gao 2024
🥈 Restormer-3D 0.869 38.6 0.951 ✓ Certified Zamir 2022
🥉 SwinIR-3D 0.846 37.5 0.942 ✓ Certified Liang 2021
4 U-Net-3D 0.810 35.9 0.924 ✓ Certified Çiçek 2016
5 CARE 0.785 34.8 0.910 ✓ Certified Weigert 2018
6 Noise2Void 0.756 33.5 0.895 ✓ Certified Krull 2019
7 IRCNN-Confocal 0.724 32.1 0.878 ✓ Certified Zhang 2017
8 Wiener-3D 0.639 28.5 0.828 ✓ Certified Wiener 1942
9 Richardson-Lucy 0.597 26.8 0.801 ✓ Certified Richardson 1972

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
🥇 SwinIR-3D + gradient 0.779
0.817
35.47 dB / 0.971
0.787
33.45 dB / 0.956
0.733
30.02 dB / 0.917
✓ Certified Liang et al., ICCV 2021 (3D adapted)
🥈 DiffusionMicro + gradient 0.771
0.844
37.28 dB / 0.979
0.765
31.85 dB / 0.941
0.705
27.93 dB / 0.879
✓ Certified Gao et al., Nat. Methods 2024
🥉 Restormer-3D + gradient 0.767
0.851
37.5 dB / 0.980
0.759
31.6 dB / 0.938
0.690
27.88 dB / 0.878
✓ Certified Zamir et al., CVPR 2022 (3D adapted)
4 IRCNN-Confocal + gradient 0.711
0.740
29.69 dB / 0.912
0.729
29.53 dB / 0.909
0.665
27.21 dB / 0.863
✓ Certified Zhang et al., CVPR 2017
5 Noise2Void + gradient 0.681
0.764
31.52 dB / 0.937
0.667
25.99 dB / 0.832
0.612
23.87 dB / 0.764
✓ Certified Krull et al., CVPR 2019
6 CARE + gradient 0.680
0.781
32.73 dB / 0.950
0.679
26.78 dB / 0.853
0.580
23.37 dB / 0.745
✓ Certified Weigert et al., Nat. Methods 2018
7 U-Net-3D + gradient 0.672
0.820
34.89 dB / 0.967
0.640
24.69 dB / 0.792
0.555
21.67 dB / 0.676
✓ Certified Çiçek et al., MICCAI 2016
8 Wiener-3D + gradient 0.645
0.669
25.83 dB / 0.827
0.660
26.03 dB / 0.833
0.605
24.22 dB / 0.776
✓ Certified Wiener, 1942
9 Richardson-Lucy + gradient 0.638
0.662
25.04 dB / 0.803
0.647
25.34 dB / 0.813
0.606
24.09 dB / 0.772
✓ Certified Richardson, J. Opt. Soc. Am. 1972

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 Restormer-3D + gradient 0.851 37.5 0.98
2 DiffusionMicro + gradient 0.844 37.28 0.979
3 U-Net-3D + gradient 0.820 34.89 0.967
4 SwinIR-3D + gradient 0.817 35.47 0.971
5 CARE + gradient 0.781 32.73 0.95
6 Noise2Void + gradient 0.764 31.52 0.937
7 IRCNN-Confocal + gradient 0.740 29.69 0.912
8 Wiener-3D + gradient 0.669 25.83 0.827
9 Richardson-Lucy + gradient 0.662 25.04 0.803
Spec Ranges (3 parameters)
Parameter Min Max Unit
z_step -50.0 100.0 nm
spherical_aberr -0.1 0.2 waves
refractive_index 1.505 1.535
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 SwinIR-3D + gradient 0.787 33.45 0.956
2 DiffusionMicro + gradient 0.765 31.85 0.941
3 Restormer-3D + gradient 0.759 31.6 0.938
4 IRCNN-Confocal + gradient 0.729 29.53 0.909
5 CARE + gradient 0.679 26.78 0.853
6 Noise2Void + gradient 0.667 25.99 0.832
7 Wiener-3D + gradient 0.660 26.03 0.833
8 Richardson-Lucy + gradient 0.647 25.34 0.813
9 U-Net-3D + gradient 0.640 24.69 0.792
Spec Ranges (3 parameters)
Parameter Min Max Unit
z_step -60.0 90.0 nm
spherical_aberr -0.12 0.18 waves
refractive_index 1.503 1.533
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 SwinIR-3D + gradient 0.733 30.02 0.917
2 DiffusionMicro + gradient 0.705 27.93 0.879
3 Restormer-3D + gradient 0.690 27.88 0.878
4 IRCNN-Confocal + gradient 0.665 27.21 0.863
5 Noise2Void + gradient 0.612 23.87 0.764
6 Richardson-Lucy + gradient 0.606 24.09 0.772
7 Wiener-3D + gradient 0.605 24.22 0.776
8 CARE + gradient 0.580 23.37 0.745
9 U-Net-3D + gradient 0.555 21.67 0.676
Spec Ranges (3 parameters)
Parameter Min Max Unit
z_step -35.0 115.0 nm
spherical_aberr -0.07 0.23 waves
refractive_index 1.508 1.538
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

Three-dimensional confocal imaging by acquiring a z-stack of optical sections. Each slice is convolved with the 3D confocal PSF. The anisotropic PSF (axial resolution ~3x worse than lateral) is a key challenge. 3D Richardson-Lucy or CARE-3D are used for volumetric deconvolution. The forward model is y(x,y,z) = PSF_3d *** x(x,y,z) + n where *** denotes 3D convolution.

Principle

Same confocal principle as live-cell mode but acquiring a full z-stack by stepping the objective or sample through the focal plane. Each optical section is convolved with the 3-D confocal PSF, and the full volume is reconstructed by 3-D deconvolution to recover isotropic resolution.

How to Build the System

Use a high-NA objective (60-100x, 1.4 NA oil or 1.2 NA water) with a piezo z-stage for precise, repeatable z-steps (typ. 200-300 nm). Acquire z-stacks covering the specimen thickness with Nyquist z-sampling. For fixed samples, oil immersion is preferred; for thick tissue, use silicone oil or glycerol objectives to minimize RI mismatch deep in the sample.

Common Reconstruction Algorithms

  • 3-D Richardson-Lucy deconvolution
  • 3-D Wiener / Tikhonov deconvolution
  • Huygens Professional iterative deconvolution
  • DeconvolutionLab2 (GPU-accelerated 3-D)
  • Deep-learning volumetric restoration (3-D U-Net, RCAN3D)

Common Mistakes

  • Using z-step larger than Nyquist, causing axial aliasing
  • Depth-dependent spherical aberration from RI mismatch not corrected
  • Not accounting for signal attenuation deeper in the sample
  • Applying 2-D deconvolution slice-by-slice instead of full 3-D
  • Incorrect PSF model (2-D Gaussian instead of 3-D Born & Wolf model)

How to Avoid Mistakes

  • Calculate Nyquist z-step (λ / (4·n·(1-cos α))) and sample accordingly
  • Use depth-dependent PSF models or adaptive optics for thick specimens
  • Apply intensity normalization per z-slice before deconvolution
  • Always perform true 3-D deconvolution to preserve axial information
  • Use measured 3-D PSF from sub-diffraction beads embedded at the correct depth

Forward-Model Mismatch Cases

  • The widefield fallback processes only 2D (64,64) images, but confocal 3D requires volumetric input (32,64,64) — the entire z-stack is discarded, losing all axial information
  • Applying 2D deconvolution slice-by-slice instead of true 3D deconvolution produces incorrect axial resolution and misses inter-slice correlations from the 3D PSF

How to Correct the Mismatch

  • Use the 3D confocal operator that processes full z-stack volumes with the anisotropic 3D PSF (worse axial than lateral resolution)
  • Perform true 3D deconvolution using the measured or modeled 3D confocal PSF; never decompose a z-stack into independent 2D slices

Experimental Setup — Signal Chain

Experimental setup diagram for Confocal 3D Z-Stack

Experimental Setup

Instrument: Zeiss LSM 880 / Leica TCS SP8
Objective: Plan Apo 63x / 1.40 NA oil
Pixel Size Nm: 80
Excitation Source: 561 nm DPSS laser
Pinhole Au: 1.0
Dwell Time Us: 8
Z Step Nm: 300
Z Slices: 64
Lateral Resolution Nm: 180
Image Size: 512x512
Reconstruction: Richardson-Lucy 3D deconvolution

Key References

  • McNally et al., 'Three-dimensional imaging by deconvolution microscopy', Methods 23, 210-217 (1999)
  • Weigert et al., 'Isotropic reconstruction of 3D fluorescence microscopy images using convolutional neural networks', MICCAI 2017

Canonical Datasets

  • Planaria 3D confocal dataset (Weigert et al.)
  • BioSR confocal 3D subset

Spec DAG — Forward Model Pipeline

C(PSF_3D) → D(g, η₃)

C 3D Confocal PSF (PSF_3D)
D PMT / HyD (g, η₃)

Mismatch Parameters

Symbol Parameter Description Nominal Perturbed
Δz z_step Z-step size error (nm) 0 50
C_s spherical_aberr Spherical aberration (waves) 0 0.1
Δn refractive_index Refractive index mismatch 1.515 1.525

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.