DEXA

Dual-Energy X-ray Absorptiometry

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
🥇 DiffusionDXA 0.901 40.4 0.956 ✓ Certified Blattmann 2023
🥈 PhysDXA 0.865 38.7 0.940 ✓ Certified Raissi 2019
🥉 SwinDXA 0.847 37.9 0.931 ✓ Certified Liu 2021
4 DXA-U-Net 0.797 35.6 0.907 ✓ Certified Huo 2021
5 PnP-DXA 0.767 34.2 0.893 ✓ Certified Venkatakrishnan 2013
6 DXA-CNN 0.754 33.8 0.881 ✓ Certified Lee 2020
7 TV-DEXA 0.672 30.1 0.841 ✓ Certified Sidky 2008
8 BML-Sep 0.635 28.7 0.813 ✓ Certified Lehmann 1981
9 FBP-DEXA 0.581 26.4 0.782 ✓ Certified Mazess 1990

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
🥇 SwinDXA + gradient 0.789
0.820
35.23 dB / 0.969
0.804
34.8 dB / 0.966
0.742
30.46 dB / 0.924
✓ Certified Liu et al., ICCV 2021 (DEXA adapt.)
🥈 PhysDXA + gradient 0.776
0.832
36.52 dB / 0.976
0.768
31.38 dB / 0.936
0.728
29.36 dB / 0.907
✓ Certified Raissi et al., J. Comput. Phys. 2019 (DEXA)
🥉 DiffusionDXA + gradient 0.764
0.851
38.57 dB / 0.984
0.747
31.2 dB / 0.933
0.693
28.62 dB / 0.893
✓ Certified Blattmann et al., arXiv 2023 (DEXA adapt.)
4 PnP-DXA + gradient 0.733
0.769
31.32 dB / 0.935
0.730
29.11 dB / 0.902
0.699
28.35 dB / 0.888
✓ Certified Venkatakrishnan et al., 2013 (DEXA adapt.)
5 DXA-CNN + gradient 0.696
0.764
31.0 dB / 0.931
0.686
27.48 dB / 0.869
0.637
24.64 dB / 0.791
✓ Certified Lee et al., Bone 2020
6 DXA-U-Net + gradient 0.672
0.794
33.8 dB / 0.959
0.651
25.18 dB / 0.808
0.570
22.84 dB / 0.725
✓ Certified Huo et al., IEEE TMED 2021
7 BML-Sep + gradient 0.650
0.673
26.02 dB / 0.833
0.658
26.14 dB / 0.836
0.618
24.04 dB / 0.770
✓ Certified Lehmann et al., Med. Phys. 1981
8 TV-DEXA + gradient 0.627
0.711
28.33 dB / 0.888
0.604
23.36 dB / 0.745
0.565
21.81 dB / 0.682
✓ Certified Sidky & Pan, Phys. Med. Biol. 2008 (DEXA)
9 FBP-DEXA + gradient 0.600
0.632
24.45 dB / 0.784
0.600
23.77 dB / 0.760
0.568
22.06 dB / 0.693
✓ Certified Mazess et al., Am. J. Clin. Nutr. 1990

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 DiffusionDXA + gradient 0.851 38.57 0.984
2 PhysDXA + gradient 0.832 36.52 0.976
3 SwinDXA + gradient 0.820 35.23 0.969
4 DXA-U-Net + gradient 0.794 33.8 0.959
5 PnP-DXA + gradient 0.769 31.32 0.935
6 DXA-CNN + gradient 0.764 31.0 0.931
7 TV-DEXA + gradient 0.711 28.33 0.888
8 BML-Sep + gradient 0.673 26.02 0.833
9 FBP-DEXA + gradient 0.632 24.45 0.784
Spec Ranges (3 parameters)
Parameter Min Max Unit
energy_offset -1.0 2.0 keV
soft_tissue -3.0 6.0 %
beam_overlap -0.02 0.04
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 SwinDXA + gradient 0.804 34.8 0.966
2 PhysDXA + gradient 0.768 31.38 0.936
3 DiffusionDXA + gradient 0.747 31.2 0.933
4 PnP-DXA + gradient 0.730 29.11 0.902
5 DXA-CNN + gradient 0.686 27.48 0.869
6 BML-Sep + gradient 0.658 26.14 0.836
7 DXA-U-Net + gradient 0.651 25.18 0.808
8 TV-DEXA + gradient 0.604 23.36 0.745
9 FBP-DEXA + gradient 0.600 23.77 0.76
Spec Ranges (3 parameters)
Parameter Min Max Unit
energy_offset -1.2 1.8 keV
soft_tissue -3.6 5.4 %
beam_overlap -0.024 0.036
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 SwinDXA + gradient 0.742 30.46 0.924
2 PhysDXA + gradient 0.728 29.36 0.907
3 PnP-DXA + gradient 0.699 28.35 0.888
4 DiffusionDXA + gradient 0.693 28.62 0.893
5 DXA-CNN + gradient 0.637 24.64 0.791
6 BML-Sep + gradient 0.618 24.04 0.77
7 DXA-U-Net + gradient 0.570 22.84 0.725
8 FBP-DEXA + gradient 0.568 22.06 0.693
9 TV-DEXA + gradient 0.565 21.81 0.682
Spec Ranges (3 parameters)
Parameter Min Max Unit
energy_offset -0.7 2.3 keV
soft_tissue -2.1 6.9 %
beam_overlap -0.014 0.046
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

DEXA measures bone mineral density (BMD) by acquiring two X-ray projections at different energies (typically 70 and 140 kVp) and decomposing the attenuation into bone and soft-tissue components using their known energy-dependent mass attenuation coefficients. The dual-energy forward model is y_E = I_0(E) * exp(-(mu_b(E)*t_b + mu_s(E)*t_s)) + n for each energy E. Output is areal BMD (g/cm2) and T-score for osteoporosis diagnosis. Precision errors of ~1% are achievable.

Principle

Dual-Energy X-ray Absorptiometry uses two X-ray beam energies to decompose the body into bone mineral and soft tissue compartments. The differential attenuation of the two energies allows separation of bone from soft tissue. Bone mineral density (BMD, g/cm²) is computed by comparing attenuation to calibration phantoms.

How to Build the System

A DEXA scanner (Hologic Discovery/Horizon or GE Lunar) uses a fan-beam or pencil-beam X-ray source with two energies (typically 70 and 140 kVp, or k-edge filtration). The detector is directly opposite the source below the patient table. Daily quality assurance with a calibration phantom (anthropomorphic spine) is mandatory. Cross-calibration is needed when changing scanners. Scan modes include AP spine, dual femur, whole body, and lateral vertebral assessment.

Common Reconstruction Algorithms

  • Dual-energy decomposition (two-material model: bone + soft tissue)
  • Edge detection for region-of-interest (ROI) identification
  • BMD calculation relative to calibration phantom
  • T-score / Z-score computation against normative databases
  • Body composition analysis (lean mass, fat mass from whole-body scans)

Common Mistakes

  • Patient positioning errors (rotation, wrong vertebral level) affecting BMD
  • Not removing metal objects (belts, jewelry) that artifactually increase BMD
  • Comparing BMD values from different scanner manufacturers without cross-calibration
  • Degenerative changes (osteophytes) falsely elevating spine BMD
  • Analyzing the wrong vertebral levels or including fractured vertebrae

How to Avoid Mistakes

  • Standardize patient positioning with positioning aids; verify on scout image
  • Remove all metal from scan field; use lateral spine view to avoid artifacts
  • Use same scanner for serial monitoring; cross-calibrate if changing equipment
  • Evaluate AP spine image for degenerative changes; consider lateral spine or femur
  • Follow ISCD guidelines for vertebral inclusion/exclusion criteria in analysis

Forward-Model Mismatch Cases

  • The widefield fallback produces a single 2D (64,64) image, but DEXA acquires dual-energy X-ray measurements — output shape (2,64,64) has two channels (high and low energy) for material decomposition
  • DEXA uses the energy-dependent difference in attenuation between bone and soft tissue to measure bone mineral density — the single-energy widefield blur cannot distinguish materials and produces no BMD information

How to Correct the Mismatch

  • Use the DEXA operator that models dual-energy Beer-Lambert transmission: y_E = I_0(E) * exp(-(mu_bone(E)*t_bone + mu_tissue(E)*t_tissue)) for E = low and high energy
  • Decompose the dual-energy measurements into bone and soft tissue components using the known energy-dependent attenuation coefficients to compute areal bone mineral density (g/cm^2)

Experimental Setup — Signal Chain

Experimental setup diagram for Dual-Energy X-ray Absorptiometry

Experimental Setup

Instrument: Hologic Discovery A / GE Lunar iDXA
Energies Kvp: [70, 140]
Pixel Size Mm: 0.5
Scan Time S: 30
Dose Usv: 1
Output: BMD (g/cm2), T-score
Sites: lumbar spine, proximal femur

Key References

  • Blake & Fogelman, 'The role of DXA bone density scans in the diagnosis and treatment of osteoporosis', Postgrad. Med. J. 83, 509-517 (2007)

Canonical Datasets

  • NHANES DXA reference data (CDC)

Spec DAG — Forward Model Pipeline

Λ(E₁,E₂) → Π(proj) → D(g, η₁)

Λ Dual-Energy Selection (E₁,E₂)
Π X-ray Projection (proj)
D Detector (g, η₁)

Mismatch Parameters

Symbol Parameter Description Nominal Perturbed
ΔE energy_offset Energy calibration offset (keV) 0 1.0
Δμ_s soft_tissue Soft-tissue attenuation error (%) 0 3.0
f_o beam_overlap Spectral overlap fraction 0 0.02

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.