Proton Therapy Imaging

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
🥇	Score-CT Score-CT Song et al., NeurIPS 2024 39.92 dB SSIM 0.984 Checkpoint unavailable	0.907	39.92	0.984	✓ Certified	Song et al., NeurIPS 2024
🥈	DiffusionCT DiffusionCT Kazemi et al., ECCV 2024 39.68 dB SSIM 0.982 Checkpoint unavailable	0.902	39.68	0.982	✓ Certified	Kazemi et al., ECCV 2024
🥉	CTFormer CTFormer Li et al., ICCV 2024 39.45 dB SSIM 0.980 Checkpoint unavailable	0.897	39.45	0.980	✓ Certified	Li et al., ICCV 2024
4	CT-ViT CT-ViT Guo et al., NeurIPS 2024 39.15 dB SSIM 0.978 Checkpoint unavailable	0.891	39.15	0.978	✓ Certified	Guo et al., NeurIPS 2024
5	DOLCE DOLCE Liu et al., ICCV 2023 38.32 dB SSIM 0.971 Checkpoint unavailable	0.874	38.32	0.971	✓ Certified	Liu et al., ICCV 2023
6	DuDoTrans DuDoTrans Wang et al., MLMIR 2022 37.68 dB SSIM 0.962 Checkpoint unavailable	0.859	37.68	0.962	✓ Certified	Wang et al., MLMIR 2022
7	Learned Primal-Dual Learned Primal-Dual Adler & Oktem, IEEE TMI 2018 36.42 dB SSIM 0.947 Checkpoint unavailable	0.831	36.42	0.947	✓ Certified	Adler & Oktem, IEEE TMI 2018
8	FBPConvNet FBPConvNet Jin et al., IEEE TIP 2017 35.81 dB SSIM 0.939 Checkpoint unavailable	0.816	35.81	0.939	✓ Certified	Jin et al., IEEE TIP 2017
9	RED-CNN RED-CNN Chen et al., IEEE TMI 2017 33.56 dB SSIM 0.908 Checkpoint unavailable	0.763	33.56	0.908	✓ Certified	Chen et al., IEEE TMI 2017
10	PnP-DnCNN PnP-DnCNN Zhang et al., 2017 33.45 dB SSIM 0.905 Try in SpecLab →	0.760	33.45	0.905	✓ Certified	Zhang et al., 2017
11	PnP-ADMM PnP-ADMM Venkatakrishnan et al., 2013 32.64 dB SSIM 0.891 Try in SpecLab →	0.740	32.64	0.891	✓ Certified	Venkatakrishnan et al., 2013
12	TV-ADMM TV-ADMM Sidky et al., 2008 30.15 dB SSIM 0.862 Try in SpecLab →	0.683	30.15	0.862	✓ Certified	Sidky et al., 2008
13	FBP FBP Kak & Slaney, 1988 27.38 dB SSIM 0.790 Try in SpecLab →	0.601	27.38	0.790	✓ Certified	Kak & Slaney, 1988

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
🥇	CT-ViT + gradient CT-ViT + gradient Guo et al., NeurIPS 2024 Score 0.803 Correct & Reconstruct →	0.803	0.835 36.5 dB / 0.976	0.803 34.18 dB / 0.962	0.772 32.03 dB / 0.943	✓ Certified	Guo et al., NeurIPS 2024
🥈	CTFormer + gradient CTFormer + gradient Li et al., ICCV 2024 Score 0.797 Correct & Reconstruct →	0.797	0.839 36.8 dB / 0.977	0.797 34.86 dB / 0.967	0.756 32.35 dB / 0.946	✓ Certified	Li et al., ICCV 2024
🥉	Score-CT + gradient Score-CT + gradient Song et al., NeurIPS 2024 Score 0.755 Correct & Reconstruct →	0.755	0.845 37.5 dB / 0.980	0.745 30.81 dB / 0.928	0.676 26.74 dB / 0.852	✓ Certified	Song et al., NeurIPS 2024
4	DiffusionCT + gradient DiffusionCT + gradient Kazemi et al., ECCV 2024 Score 0.751 Correct & Reconstruct →	0.751	0.863 38.36 dB / 0.983	0.718 29.48 dB / 0.909	0.672 26.32 dB / 0.841	✓ Certified	Kazemi et al., ECCV 2024
5	Learned Primal-Dual + gradient Learned Primal-Dual + gradient Adler & Oktem, IEEE TMI 2018 Score 0.750 Correct & Reconstruct →	0.750	0.805 34.54 dB / 0.965	0.740 29.71 dB / 0.912	0.706 28.85 dB / 0.898	✓ Certified	Adler & Oktem, IEEE TMI 2018
6	DuDoTrans + gradient DuDoTrans + gradient Wang et al., MLMIR 2022 Score 0.744 Correct & Reconstruct →	0.744	0.839 36.25 dB / 0.975	0.713 29.29 dB / 0.905	0.680 26.94 dB / 0.857	✓ Certified	Wang et al., MLMIR 2022
7	DOLCE + gradient DOLCE + gradient Liu et al., ICCV 2023 Score 0.736 Correct & Reconstruct →	0.736	0.826 35.87 dB / 0.973	0.726 29.57 dB / 0.910	0.656 26.33 dB / 0.841	✓ Certified	Liu et al., ICCV 2023
8	PnP-ADMM + gradient PnP-ADMM + gradient Venkatakrishnan et al., IEEE GlobalSIP 2013 Score 0.726 Correct & Reconstruct →	0.726	0.751 30.67 dB / 0.926	0.721 28.57 dB / 0.892	0.706 28.77 dB / 0.896	✓ Certified	Venkatakrishnan et al., IEEE GlobalSIP 2013
9	FBPConvNet + gradient FBPConvNet + gradient Jin et al., IEEE TIP 2017 Score 0.726 Correct & Reconstruct →	0.726	0.815 34.05 dB / 0.961	0.700 28.21 dB / 0.885	0.664 26.36 dB / 0.842	✓ Certified	Jin et al., IEEE TIP 2017
10	PnP-DnCNN + gradient PnP-DnCNN + gradient Zhang et al., IEEE TIP 2017 Score 0.712 Correct & Reconstruct →	0.712	0.760 30.96 dB / 0.930	0.702 27.54 dB / 0.871	0.675 27.1 dB / 0.861	✓ Certified	Zhang et al., IEEE TIP 2017
11	TV-ADMM + gradient TV-ADMM + gradient Sidky et al., Phys. Med. Biol. 2008 Score 0.698 Correct & Reconstruct →	0.698	0.707 27.83 dB / 0.877	0.698 27.43 dB / 0.868	0.688 27.86 dB / 0.878	✓ Certified	Sidky et al., Phys. Med. Biol. 2008
12	RED-CNN + gradient RED-CNN + gradient Chen et al., IEEE TMI 2017 Score 0.695 Correct & Reconstruct →	0.695	0.758 30.6 dB / 0.926	0.679 27.39 dB / 0.867	0.648 26.08 dB / 0.834	✓ Certified	Chen et al., IEEE TMI 2017
13	FBP + gradient FBP + gradient Kak & Slaney, IEEE Press 1988 Score 0.655 Correct & Reconstruct →	0.655	0.676 25.79 dB / 0.826	0.658 25.95 dB / 0.831	0.631 24.42 dB / 0.783	✓ Certified	Kak & Slaney, IEEE Press 1988

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	DiffusionCT + gradient	0.863	38.36	0.983
2	Score-CT + gradient	0.845	37.5	0.98
3	CTFormer + gradient	0.839	36.8	0.977
4	DuDoTrans + gradient	0.839	36.25	0.975
5	CT-ViT + gradient	0.835	36.5	0.976
6	DOLCE + gradient	0.826	35.87	0.973
7	FBPConvNet + gradient	0.815	34.05	0.961
8	Learned Primal-Dual + gradient	0.805	34.54	0.965
9	PnP-DnCNN + gradient	0.760	30.96	0.93
10	RED-CNN + gradient	0.758	30.6	0.926
11	PnP-ADMM + gradient	0.751	30.67	0.926
12	TV-ADMM + gradient	0.707	27.83	0.877
13	FBP + gradient	0.676	25.79	0.826

Spec Ranges (4 parameters)

Parameter	Min	Max	Unit
range_uncertainty	-0.6	1.2	mm
scattering_power_error	0.99	1.02	-
detector_efficiency_drift	0.84	0.87	-
setup_error	-0.4	0.8	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	CT-ViT + gradient	0.803	34.18	0.962
2	CTFormer + gradient	0.797	34.86	0.967
3	Score-CT + gradient	0.745	30.81	0.928
4	Learned Primal-Dual + gradient	0.740	29.71	0.912
5	DOLCE + gradient	0.726	29.57	0.91
6	PnP-ADMM + gradient	0.721	28.57	0.892
7	DiffusionCT + gradient	0.718	29.48	0.909
8	DuDoTrans + gradient	0.713	29.29	0.905
9	PnP-DnCNN + gradient	0.702	27.54	0.871
10	FBPConvNet + gradient	0.700	28.21	0.885
11	TV-ADMM + gradient	0.698	27.43	0.868
12	RED-CNN + gradient	0.679	27.39	0.867
13	FBP + gradient	0.658	25.95	0.831

Spec Ranges (4 parameters)

Parameter	Min	Max	Unit
range_uncertainty	-0.72	1.08	mm
scattering_power_error	0.988	1.018	-
detector_efficiency_drift	0.838	0.868	-
setup_error	-0.48	0.72	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	CT-ViT + gradient	0.772	32.03	0.943
2	CTFormer + gradient	0.756	32.35	0.946
3	Learned Primal-Dual + gradient	0.706	28.85	0.898
4	PnP-ADMM + gradient	0.706	28.77	0.896
5	TV-ADMM + gradient	0.688	27.86	0.878
6	DuDoTrans + gradient	0.680	26.94	0.857
7	Score-CT + gradient	0.676	26.74	0.852
8	PnP-DnCNN + gradient	0.675	27.1	0.861
9	DiffusionCT + gradient	0.672	26.32	0.841
10	FBPConvNet + gradient	0.664	26.36	0.842
11	DOLCE + gradient	0.656	26.33	0.841
12	RED-CNN + gradient	0.648	26.08	0.834
13	FBP + gradient	0.631	24.42	0.783

Spec Ranges (4 parameters)

Parameter	Min	Max	Unit
range_uncertainty	-0.42	1.38	mm
scattering_power_error	0.993	1.023	-
detector_efficiency_drift	0.843	0.873	-
setup_error	-0.28	0.92	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

Spec DAG — Forward Model Pipeline

Π → D

Π Projection

D Detector

Mismatch Parameters

Symbol	Parameter	Description	Nominal	Perturbed
r_u	range_uncertainty	Range uncertainty (mm)	0.0	0.6
s_p	scattering_power_error	Scattering power error (-)	1.0	1.01
d_e	detector_efficiency_drift	Detector efficiency drift (-)	0.85	0.86
s_e	setup_error	Setup error (mm)	0.0	0.4

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.