Mammography

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 0.801 Correct & Reconstruct →	0.801	0.857 37.86 dB / 0.982	0.805 35.0 dB / 0.968	0.741 31.2 dB / 0.933	✓ Certified	Guo et al., NeurIPS 2024
🥈	CTFormer + gradient CTFormer + gradient Li et al., ICCV 2024 Score 0.795 Correct & Reconstruct →	0.795	0.842 37.57 dB / 0.980	0.799 34.5 dB / 0.964	0.745 30.5 dB / 0.924	✓ Certified	Li et al., ICCV 2024
🥉	DiffusionCT + gradient DiffusionCT + gradient Kazemi et al., ECCV 2024 Score 0.791 Correct & Reconstruct →	0.791	0.843 37.53 dB / 0.980	0.779 33.33 dB / 0.955	0.750 31.65 dB / 0.939	✓ Certified	Kazemi et al., ECCV 2024
4	DOLCE + gradient DOLCE + gradient Liu et al., ICCV 2023 Score 0.784 Correct & Reconstruct →	0.784	0.847 36.83 dB / 0.977	0.764 31.91 dB / 0.942	0.742 30.68 dB / 0.927	✓ Certified	Liu et al., ICCV 2023
5	Score-CT + gradient Score-CT + gradient Song et al., NeurIPS 2024 Score 0.782 Correct & Reconstruct →	0.782	0.866 38.45 dB / 0.984	0.772 32.51 dB / 0.948	0.709 29.18 dB / 0.903	✓ Certified	Song et al., NeurIPS 2024
6	DuDoTrans + gradient DuDoTrans + gradient Wang et al., MLMIR 2022 Score 0.723 Correct & Reconstruct →	0.723	0.820 35.84 dB / 0.973	0.700 27.35 dB / 0.866	0.650 25.86 dB / 0.828	✓ Certified	Wang et al., MLMIR 2022
7	Learned Primal-Dual + gradient Learned Primal-Dual + gradient Adler & Oktem, IEEE TMI 2018 Score 0.716 Correct & Reconstruct →	0.716	0.826 35.4 dB / 0.970	0.683 27.52 dB / 0.870	0.640 25.69 dB / 0.823	✓ Certified	Adler & Oktem, IEEE TMI 2018
8	FBPConvNet + gradient FBPConvNet + gradient Jin et al., IEEE TIP 2017 Score 0.712 Correct & Reconstruct →	0.712	0.816 34.44 dB / 0.964	0.678 27.35 dB / 0.866	0.643 25.73 dB / 0.824	✓ Certified	Jin et al., IEEE TIP 2017
9	PnP-ADMM + gradient PnP-ADMM + gradient Venkatakrishnan et al., IEEE GlobalSIP 2013 Score 0.698 Correct & Reconstruct →	0.698	0.771 30.91 dB / 0.930	0.690 27.47 dB / 0.869	0.634 24.7 dB / 0.792	✓ Certified	Venkatakrishnan et al., IEEE GlobalSIP 2013
10	RED-CNN + gradient RED-CNN + gradient Chen et al., IEEE TMI 2017 Score 0.695 Correct & Reconstruct →	0.695	0.762 30.97 dB / 0.930	0.687 27.6 dB / 0.872	0.635 24.66 dB / 0.791	✓ Certified	Chen et al., IEEE TMI 2017
11	PnP-DnCNN + gradient PnP-DnCNN + gradient Zhang et al., IEEE TIP 2017 Score 0.676 Correct & Reconstruct →	0.676	0.759 30.94 dB / 0.930	0.670 26.76 dB / 0.852	0.600 23.29 dB / 0.742	✓ Certified	Zhang et al., IEEE TIP 2017
12	TV-ADMM + gradient TV-ADMM + gradient Sidky et al., Phys. Med. Biol. 2008 Score 0.665 Correct & Reconstruct →	0.665	0.731 28.72 dB / 0.895	0.646 25.85 dB / 0.828	0.618 24.01 dB / 0.769	✓ Certified	Sidky et al., Phys. Med. Biol. 2008
13	FBP + gradient FBP + gradient Kak & Slaney, IEEE Press 1988 Score 0.615 Correct & Reconstruct →	0.615	0.652 25.33 dB / 0.812	0.622 24.45 dB / 0.784	0.571 22.88 dB / 0.726	✓ 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	Score-CT + gradient	0.866	38.45	0.984
2	CT-ViT + gradient	0.857	37.86	0.982
3	DOLCE + gradient	0.847	36.83	0.977
4	DiffusionCT + gradient	0.843	37.53	0.98
5	CTFormer + gradient	0.842	37.57	0.98
6	Learned Primal-Dual + gradient	0.826	35.4	0.97
7	DuDoTrans + gradient	0.820	35.84	0.973
8	FBPConvNet + gradient	0.816	34.44	0.964
9	PnP-ADMM + gradient	0.771	30.91	0.93
10	RED-CNN + gradient	0.762	30.97	0.93
11	PnP-DnCNN + gradient	0.759	30.94	0.93
12	TV-ADMM + gradient	0.731	28.72	0.895
13	FBP + gradient	0.652	25.33	0.812

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
compression	-2.0	4.0	mm
anode_angle	-0.5	1.0	deg
scatter	0.25	0.4

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	0.805	35.0	0.968
2	CTFormer + gradient	0.799	34.5	0.964
3	DiffusionCT + gradient	0.779	33.33	0.955
4	Score-CT + gradient	0.772	32.51	0.948
5	DOLCE + gradient	0.764	31.91	0.942
6	DuDoTrans + gradient	0.700	27.35	0.866
7	PnP-ADMM + gradient	0.690	27.47	0.869
8	RED-CNN + gradient	0.687	27.6	0.872
9	Learned Primal-Dual + gradient	0.683	27.52	0.87
10	FBPConvNet + gradient	0.678	27.35	0.866
11	PnP-DnCNN + gradient	0.670	26.76	0.852
12	TV-ADMM + gradient	0.646	25.85	0.828
13	FBP + gradient	0.622	24.45	0.784

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
compression	-2.4	3.6	mm
anode_angle	-0.6	0.9	deg
scatter	0.24	0.39

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	DiffusionCT + gradient	0.750	31.65	0.939
2	CTFormer + gradient	0.745	30.5	0.924
3	DOLCE + gradient	0.742	30.68	0.927
4	CT-ViT + gradient	0.741	31.2	0.933
5	Score-CT + gradient	0.709	29.18	0.903
6	DuDoTrans + gradient	0.650	25.86	0.828
7	FBPConvNet + gradient	0.643	25.73	0.824
8	Learned Primal-Dual + gradient	0.640	25.69	0.823
9	RED-CNN + gradient	0.635	24.66	0.791
10	PnP-ADMM + gradient	0.634	24.7	0.792
11	TV-ADMM + gradient	0.618	24.01	0.769
12	PnP-DnCNN + gradient	0.600	23.29	0.742
13	FBP + gradient	0.571	22.88	0.726

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
compression	-1.4	4.6	mm
anode_angle	-0.35	1.15	deg
scatter	0.265	0.415

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

Full-field digital mammography (FFDM) produces high-resolution X-ray projection images of compressed breast tissue for cancer screening. The low-energy X-ray beam (25-32 kVp with W/Rh or Mo/Mo target-filter) maximizes soft tissue contrast. Amorphous selenium flat-panel detectors provide direct conversion with ~50 um pixel pitch. The forward model follows Beer-Lambert with energy-dependent attenuation. Primary challenges include overlapping tissue structures, microcalcification detection, and dense breast tissue masking lesions.

Principle

Mammography uses low-energy X-rays (25-35 kVp) with specialized anode/filter combinations (Mo/Mo, Mo/Rh, W/Rh) to optimize contrast between breast tissue types (adipose, glandular, calcifications). Breast compression reduces thickness and scatter, improving contrast and reducing dose. Digital mammography uses flat-panel detectors for direct or indirect X-ray detection.

How to Build the System

A dedicated mammography unit with a compression paddle, specialized X-ray tube (Mo, Rh, or W anode), and high-resolution flat-panel detector (50-100 μm pixel size, amorphous selenium for direct conversion). Automatic optimization of target/filter and kVp based on compressed breast thickness. Regular quality assurance per ACR/MQSA requirements: phantom images, SNR measurements, artifact checks, and AEC calibration.

Common Reconstruction Algorithms

Contrast-limited adaptive histogram equalization (CLAHE) for display
Computer-aided detection (CAD) for microcalcification and mass detection
Digital breast tomosynthesis (DBT) reconstruction (FBP or iterative)
Deep-learning breast density classification (BI-RADS categories)
Synthetic 2D mammography from DBT volumes

Common Mistakes

Insufficient breast compression, increasing dose and reducing contrast
Positioning errors cutting off breast tissue (especially axillary tail)
Grid artifacts or grid cutoff from misaligned Bucky grid
Exposure errors from AEC sensor placed over dense tissue vs. adipose
Motion blur from long exposure times in thick or dense breasts

How to Avoid Mistakes

Apply firm, consistent compression; verify thickness readout is reasonable
Follow standardized positioning protocols (CC, MLO) with technologist training
Verify grid alignment and use reciprocating grid to eliminate grid lines
Position AEC sensor appropriately for breast density; adjust manually if needed
Use shortest possible exposure with adequate mAs; consider large-angle tomosynthesis

Forward-Model Mismatch Cases

The widefield fallback applies Gaussian blur, but mammography uses low-energy X-ray transmission (25-35 kVp) with tissue-specific attenuation coefficients optimized for fat/glandular tissue contrast — the physics model is fundamentally different
Mammographic image formation involves compression geometry, scatter grid rejection, anti-scatter grid, and detector-specific MTF — none of these are captured by a simple spatial Gaussian blur

How to Correct the Mismatch

Use the mammography operator implementing Beer-Lambert transmission at mammographic energies with tissue-specific attenuation: y = I_0 * exp(-mu_tissue * t) for fat, glandular, and calcification components
Include scatter rejection model, detector quantum efficiency (DQE), and geometric magnification for accurate forward modeling and quantitative breast density estimation

Experimental Setup — Signal Chain

Experimental setup diagram for Mammography

Reconstruction Gallery — 4 Scenes × 3 Scenarios

Method: CPU_baseline | Mismatch: nominal (nominal=True, perturbed=False)

I Ideal II Mismatch (uncorrected) III Oracle correction

I — Ideal

Ground Truth

Measurement

Reconstruction

18.75706809390867 dB SSIM 0.8193796826772168

II — Mismatch

Ground Truth

Measurement

Reconstruction

16.51708673843107 dB SSIM 0.8010193209571606

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

2.026051331025295 dB SSIM -0.029079755804036336

I — Ideal

Ground Truth

Measurement

Reconstruction

17.573653750956822 dB SSIM 0.7843662566135395

II — Mismatch

Ground Truth

Measurement

Reconstruction

12.948442065810703 dB SSIM 0.7571764078246858

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

2.1587216587240707 dB SSIM -0.07899526490180157

I — Ideal

Ground Truth

Measurement

Reconstruction

23.79125671172779 dB SSIM 0.9272437152554989

II — Mismatch

Ground Truth

Measurement

Reconstruction

25.60238639864057 dB SSIM 0.9300905234761239

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

1.6125778550107865 dB SSIM 0.010872066143016603

I — Ideal

Ground Truth

Measurement

Reconstruction

25.021414573805604 dB SSIM 0.9012129326692745

II — Mismatch

Ground Truth

Measurement

Reconstruction

18.825287032760002 dB SSIM 0.8967124120893478

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

1.908795384138298 dB SSIM -0.03821022143108241

Mean PSNR Across All Scenes

21.285848282599723 dB

I — Ideal

18.473300558910587 dB

II — Mismatch

1.9265365572246127 dB

III — Oracle

Degradation: -2.8 dB Recovery: +-16.5 dB

Per-scene PSNR breakdown (4 scenes)

Scene	I (PSNR)	I (SSIM)	II (PSNR)	II (SSIM)	III (PSNR)	III (SSIM)
scene_00	18.75706809390867	0.8193796826772168	16.51708673843107	0.8010193209571606	2.026051331025295	-0.029079755804036336
scene_01	17.573653750956822	0.7843662566135395	12.948442065810703	0.7571764078246858	2.1587216587240707	-0.07899526490180157
scene_02	23.79125671172779	0.9272437152554989	25.60238639864057	0.9300905234761239	1.6125778550107865	0.010872066143016603
scene_03	25.021414573805604	0.9012129326692745	18.825287032760002	0.8967124120893478	1.908795384138298	-0.03821022143108241
Mean	21.285848282599723	0.8580506468038824	18.473300558910587	0.8462496660868296	1.9265365572246127	-0.03385329399847593

Experimental Setup

Instrument: Hologic Selenia Dimensions / Siemens MAMMOMAT Revelation

Image Size: 2294x1914

Kvp: 28

Target Filter: W/Rh

Mas: 60

Detector: flat-panel amorphous selenium (direct conversion)

Pixel Pitch Um: 70

Dataset: VinDr-Mammo, CBIS-DDSM, INbreast

Key References

VinDr-Mammo, Scientific Data 2023
Lee et al., 'A curated mammography dataset (CBIS-DDSM)', Scientific Data 4, 170177 (2017)

Canonical Datasets

VinDr-Mammo (5000 4-view exams)
CBIS-DDSM (curated DDSM subset)
INbreast (410 images, Moreira et al.)

Spec DAG — Forward Model Pipeline

Π(contact) → D(g, η₁)

Π Contact Projection (contact)

D Detector (g, η₁)

Mismatch Parameters

Symbol	Parameter	Description	Nominal	Perturbed
Δh	compression	Compression thickness error (mm)	0	2.0
Δα	anode_angle	Anode angle error (deg)	0	0.5
f_s	scatter	Scatter fraction error	0.3	0.35

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.