Public

DEXA — Public Tier

(3 scenes)

Full-access development tier with all data visible.

Measurements (y), ideal forward operator (H), spec ranges, ground truth (x_true), and true mismatch spec.

Load HDF5 → compare reconstruction vs x_true → check consistency → iterate.

Reconstructed signals (x_hat) and corrected spec as HDF5.

✓

True spec visible — use these exact values for Scenario III oracle reconstruction.

Parameter	Spec Range	True Value	Unit
energy_offset	-1.0 – 2.0	0.5	keV
soft_tissue	-3.0 – 6.0	1.5	%
beam_overlap	-0.02 – 0.04	0.01

Method: CPU_baseline — Mismatch parameter: nominal

9.33 dB

SSIM 0.0730

8.45 dB

SSIM 0.0249

10.85 dB

SSIM 0.1611

Per-scene breakdown (4 scenes)

Scene	PSNR I	SSIM I	PSNR II	SSIM II	PSNR III	SSIM III
scene_00	8.85	0.0678	8.09	0.0247	10.36	0.1590
scene_01	9.91	0.0750	8.97	0.0252	11.44	0.1618
scene_02	9.58	0.0745	8.61	0.0252	11.27	0.1707
scene_03	8.98	0.0744	8.13	0.0246	10.34	0.1529

#	Method	Score	PSNR	SSIM	Consistency	Trust	Source
1	DiffusionDXA + gradient	0.851	38.57	0.984	0.86	✓ Certified	Blattmann et al., arXiv 2023 (DEXA adapt.)
2	PhysDXA + gradient	0.832	36.52	0.976	0.88	✓ Certified	Raissi et al., J. Comput. Phys. 2019 (DEXA)
3	SwinDXA + gradient	0.820	35.23	0.969	0.9	✓ Certified	Liu et al., ICCV 2021 (DEXA adapt.)
4	DXA-U-Net + gradient	0.794	33.8	0.959	0.86	✓ Certified	Huo et al., IEEE TMED 2021
5	PnP-DXA + gradient	0.769	31.32	0.935	0.91	✓ Certified	Venkatakrishnan et al., 2013 (DEXA adapt.)
6	DXA-CNN + gradient	0.764	31.0	0.931	0.91	✓ Certified	Lee et al., Bone 2020
7	TV-DEXA + gradient	0.711	28.33	0.888	0.86	✓ Certified	Sidky & Pan, Phys. Med. Biol. 2008 (DEXA)
8	BML-Sep + gradient	0.673	26.02	0.833	0.9	✓ Certified	Lehmann et al., Med. Phys. 1981
9	FBP-DEXA + gradient	0.632	24.45	0.784	0.87	✓ Certified	Mazess et al., Am. J. Clin. Nutr. 1990

y H_ideal spec_ranges x_true true_spec

Format: HDF5

Scenes: 3

0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖)

PSNR: 40% SSIM: 40% Consistency: 20%