CBCT

Cone-Beam Computed 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
🥇 DiffusionCBCT 0.900 40.1 0.964 ✓ Certified Gao 2024
🥈 CTFormer 0.877 39.0 0.953 ✓ Certified Wang 2023
🥉 DuDoTrans 0.859 38.2 0.944 ✓ Certified Wang 2022
4 DuDoNet 0.834 37.1 0.932 ✓ Certified Lin 2019
5 Learned Primal-Dual 0.817 36.4 0.921 ✓ Certified Adler 2018
6 Metal-AR-Net 0.803 35.8 0.912 ✓ Certified Zhang 2018
7 FBPConvNet 0.770 34.5 0.891 ✓ Certified Jin 2017
8 TV-ADMM 0.696 31.2 0.851 ✓ Certified Boyd 2011
9 FDK 0.614 27.8 0.801 ✓ Certified Feldkamp 1984

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
🥇 CTFormer + gradient 0.787
0.835
36.97 dB / 0.978
0.782
33.99 dB / 0.961
0.744
30.74 dB / 0.927
✓ Certified Wang et al., MICCAI 2023
🥈 DuDoTrans + gradient 0.782
0.823
35.31 dB / 0.970
0.772
33.21 dB / 0.954
0.750
31.51 dB / 0.937
✓ Certified Wang et al., IEEE TMI 2022
🥉 DiffusionCBCT + gradient 0.745
0.848
37.92 dB / 0.982
0.728
29.32 dB / 0.906
0.658
26.72 dB / 0.851
✓ Certified Gao et al., Med. Phys. 2024
4 Learned Primal-Dual + gradient 0.724
0.804
34.48 dB / 0.964
0.719
28.48 dB / 0.891
0.649
25.54 dB / 0.819
✓ Certified Adler & Oktem, IEEE TMI 2018
5 DuDoNet + gradient 0.722
0.811
34.53 dB / 0.965
0.700
28.55 dB / 0.892
0.655
26.68 dB / 0.850
✓ Certified Lin et al., CVPR 2019
6 Metal-AR-Net + gradient 0.703
0.818
34.72 dB / 0.966
0.669
26.5 dB / 0.846
0.622
24.54 dB / 0.787
✓ Certified Zhang & Yu, IEEE TMI 2018
7 FBPConvNet + gradient 0.666
0.774
31.78 dB / 0.940
0.647
25.36 dB / 0.813
0.576
22.68 dB / 0.718
✓ Certified Jin et al., IEEE TIP 2017
8 FDK + gradient 0.658
0.685
26.2 dB / 0.838
0.665
26.38 dB / 0.842
0.625
24.41 dB / 0.783
✓ Certified Feldkamp et al., J. Opt. Soc. Am. A 1984
9 TV-ADMM + gradient 0.653
0.749
29.63 dB / 0.911
0.628
24.26 dB / 0.778
0.583
22.63 dB / 0.716
✓ Certified Boyd et al., Found. Trends 2011

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 10 scenes

Full-access tier: 10 real CBCT/CT volumes from AAPM, LIDC-IDRI, CBCTLiTS, MMDental, CTooth+, 2DeteCT, HTC, Walnut CT, CQ500, DM4CT.

What you get & how to use

What you get: Cone-beam projections (y), ideal geometry (H), spec ranges, ground truth volume (x_true), and true mismatch spec.

How to use: Load cbct_challenge_public.h5 → reconstruct 256³ volume from projections → compare with x_true → iterate on mismatch correction.

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

Public Leaderboard
# Method Score PSNR SSIM
1 DiffusionCBCT + gradient 0.848 37.92 0.982
2 CTFormer + gradient 0.835 36.97 0.978
3 DuDoTrans + gradient 0.823 35.31 0.97
4 Metal-AR-Net + gradient 0.818 34.72 0.966
5 DuDoNet + gradient 0.811 34.53 0.965
6 Learned Primal-Dual + gradient 0.804 34.48 0.964
7 FBPConvNet + gradient 0.774 31.78 0.94
8 TV-ADMM + gradient 0.749 29.63 0.911
9 FDK + gradient 0.685 26.2 0.838
Spec Ranges (6 parameters)
Parameter Min Max Unit
source_offset_x -1.2 2.8 mm
source_offset_z -1.0 2.0 mm
detector_tilt -0.35 0.65 deg
detector_shift_u -1.8 4.2 px
beam_hardening -0.015 0.135
scatter_fraction -0.01 0.09
View Details →
Dev 20 scenes

Blind evaluation: 20 procedural 256³ phantoms (anatomy-inspired, based on CQ500/AAPM/CBCTLiTS/MMDental characteristics).

What you get & how to use

What you get: Cone-beam projections (y), ideal geometry (H), and spec ranges. No ground truth.

How to use: Apply your pipeline from Public tier. Self-check via consistency metric. Ground truth scored server-side.

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

Dev Leaderboard
# Method Score PSNR SSIM
1 CTFormer + gradient 0.782 33.99 0.961
2 DuDoTrans + gradient 0.772 33.21 0.954
3 DiffusionCBCT + gradient 0.728 29.32 0.906
4 Learned Primal-Dual + gradient 0.719 28.48 0.891
5 DuDoNet + gradient 0.700 28.55 0.892
6 Metal-AR-Net + gradient 0.669 26.5 0.846
7 FDK + gradient 0.665 26.38 0.842
8 FBPConvNet + gradient 0.647 25.36 0.813
9 TV-ADMM + gradient 0.628 24.26 0.778
Spec Ranges (6 parameters)
Parameter Min Max Unit
source_offset_x -1.5 2.5 mm
source_offset_z -1.2 1.8 mm
detector_tilt -0.4 0.6 deg
detector_shift_u -2.2 3.8 px
beam_hardening -0.035 0.115
scatter_fraction -0.02 0.08
Submit Reconstruction View Details →
Hidden 20 scenes

Fully blind: 20 adversarial 256³ phantoms (extreme complexity: TPMS, reaction-diffusion, multi-metal, depth-6 vascular trees). Strongest mismatch.

What you get & how to use

What you get: No data download. Algorithm runs server-side on hidden projections.

How to use: Package algorithm as Docker container / Python script accepting projections + geometry, outputting reconstructed volume + corrected spec.

What to submit: Containerized algorithm. Scored server-side against adversarial phantoms.

Hidden Leaderboard
# Method Score PSNR SSIM
1 DuDoTrans + gradient 0.750 31.51 0.937
2 CTFormer + gradient 0.744 30.74 0.927
3 DiffusionCBCT + gradient 0.658 26.72 0.851
4 DuDoNet + gradient 0.655 26.68 0.85
5 Learned Primal-Dual + gradient 0.649 25.54 0.819
6 FDK + gradient 0.625 24.41 0.783
7 Metal-AR-Net + gradient 0.622 24.54 0.787
8 TV-ADMM + gradient 0.583 22.63 0.716
9 FBPConvNet + gradient 0.576 22.68 0.718
Spec Ranges (6 parameters)
Parameter Min Max Unit
source_offset_x -0.5 3.5 mm
source_offset_z -0.5 2.5 mm
detector_tilt -0.15 0.85 deg
detector_shift_u -0.8 5.2 px
beam_hardening 0.045 0.195
scatter_fraction 0.03 0.13
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

Cone-beam CT (CBCT) uses a divergent cone-shaped X-ray beam and a flat-panel 2D detector to acquire volumetric data in a single rotation, unlike fan-beam CT which acquires slice-by-slice. The 3D Feldkamp-Davis-Kress (FDK) algorithm performs approximate filtered back-projection for cone geometry. CBCT is widely used in dental, ENT, and image-guided radiation therapy. Primary artifacts include cone-beam artifacts at large cone angles, scatter, and truncation. Sparse-view CBCT reduces scan time and dose but introduces streak artifacts.

Principle

Cone-Beam CT uses a divergent cone-shaped X-ray beam and a 2-D flat-panel detector to acquire a volumetric CT dataset in a single rotation. Unlike multi-slice CT with a narrow fan beam, CBCT covers the full volume simultaneously, enabling faster acquisition but with increased scatter and cone-beam artifacts compared to conventional CT.

How to Build the System

Mount a flat-panel detector (typically 30×40 cm, CsI scintillator) opposite an X-ray tube on a rotating gantry or C-arm. Common implementations: dental CBCT (small FOV, 90 kVp), image-guided radiation therapy CBCT (kV source on linac gantry), and C-arm CBCT (interventional). Calibrate: geometric parameters (source-detector distances, isocenter), detector offset corrections, and scatter correction LUTs.

Common Reconstruction Algorithms

  • FDK (Feldkamp-Davis-Kress) cone-beam filtered back-projection
  • Iterative CBCT (SART, SIRT with cone-beam projector)
  • Scatter correction (measurement-based or Monte Carlo simulation)
  • Motion-compensated CBCT (4D-CBCT for respiratory motion)
  • Deep-learning CBCT-to-CT synthesis for radiation therapy planning

Common Mistakes

  • Severe scatter artifacts (cupping, shading) in large FOV acquisitions
  • Cone-beam artifacts near the edges of the FOV (Feldkamp approximation breaks down)
  • Truncation artifacts when anatomy extends outside the FOV
  • Motion artifacts in thorax/abdomen from respiratory and cardiac motion
  • Insufficient angular sampling causing streak artifacts

How to Avoid Mistakes

  • Apply scatter correction (anti-scatter grid, software correction, or beam-blocker method)
  • Limit cone angle or use exact reconstruction algorithms for large cone angles
  • Use extended FOV techniques (shifted detector, multiple scans) for large anatomy
  • Apply 4D-CBCT or gated acquisition for moving anatomy
  • Acquire sufficient projections (≥600 for a full rotation) with uniform angular spacing

Forward-Model Mismatch Cases

  • The widefield fallback produces a blurred (64,64) image, but cone-beam CT acquires a sinogram of shape (n_angles, n_detector_rows * n_detector_cols) from a 2D detector rotating around the patient — the data is a set of cone-beam projections, not a blurred image
  • CBCT cone-beam geometry introduces axial cone-angle artifacts (Feldkamp approximation errors) that are absent from the widefield model — any reconstruction expecting cone-beam projection data will fail with the blurred image

How to Correct the Mismatch

  • Use the CBCT operator implementing cone-beam projection (Radon transform in 3D divergent geometry) for each source-detector angle, producing the correct sinogram/projection data shape
  • Reconstruct using FDK (Feldkamp-Davis-Kress) algorithm or iterative cone-beam methods (SART, ADMM) with the correct cone-beam system matrix

Experimental Setup — Signal Chain

Experimental setup diagram for Cone-Beam Computed Tomography

Experimental Setup

Instrument: Varian TrueBeam / Elekta XVI / iCAT dental CBCT
Image Size: 512x512
Projection Views: 360
Sparse Views: 20
Detector Size: 150x150 px
Pixel Pitch Mm: 0.4
Kvp: 90
Tube Current Ma: 8
Voxel Size Mm: 0.3
Reconstruction: FDK / iterative

Key References

  • Feldkamp et al., 'Practical cone-beam algorithm', JOSA A 1, 612-619 (1984)

Canonical Datasets

  • ICASSP 2024 CBCT Challenge

Spec DAG — Forward Model Pipeline

R(θ) → Π(cone) → D(g, η₁)

R Gantry Rotation (θ)
Π Cone-Beam Projection (cone)
D Detector (g, η₁)

Mismatch Parameters

Symbol Parameter Description Nominal Perturbed
Δc center_offset Center-of-rotation offset (pixels) 0 2.0
Δd_s source_dist Source-to-isocenter distance error (mm) 0 1.0
Δα cone_angle Cone half-angle error (deg) 0 0.3
φ detector_tilt Detector tilt (deg) 0 0.5

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.