AXON

A chest radiograph is fast, inexpensive, and available almost everywhere; a chest CT is none of those things and delivers far more radiation. Recovering a 3D volume from the 2D projection is badly ill-posed — one projection collapses depth and superimposes structures that must be pulled apart again. The harder obstacle is methodological: nearly all X-ray-to-CT work is trained and scored on digitally reconstructed radiographs, synthetic projections computed from the very CT the model must recover, perfectly registered and free of scatter, beam hardening, detector crosstalk, and vendor preprocessing. None of that holds for a clinical radiograph.

AXON (Advanced X-ray to CT-volume Network) is a multi-stage conditional diffusion framework built for the real-radiograph setting, from Christian Wachinger's Lab for Artificial Intelligence in Medical Imaging at the Technical University of Munich with the Munich Center for Machine Learning and the University of Western Australia, and accepted at BMVC 2026. Rather than regress a volume directly, it decomposes the mapping into intermediate stages: a diffusion model lifts the radiograph into a coarse CT estimate, a second diffusion model conditioned on that estimate restores local intensity detail, and a super-resolution network scales the result up. Synthetic projections enter only as training augmentation: every validation and test case is a real X-ray paired with its own CT, and the trained model is also evaluated unchanged on an external clinical cohort.

#Key Features

• Brownian bridge lift out of the projection domain: CoarseDiff builds a stochastic bridge between an encoded X-ray feature volume and the CT distribution, so sampling starts from the conditioning input rather than from Gaussian noise.
• ControlNet-guided 3D refinement: FineDiff trains a rectified-flow 3D U-Net as an unconditional CT prior, then freezes it while a ControlNet branch conditioned on the coarse volume steers generation toward the observed anatomy.
• Bi-planar conditioning through a learned fusion block: Frontal and lateral views share one image encoder and merge in a residual block of 3×3×3 volumetric convolutions, supplying the depth cue a single projection cannot.
• Super-resolution decoupled from the diffusion cost: Diffusion runs at 128³ and a separate 3D dense network upsamples to 256³, sidestepping the cubic memory growth of volumetric generation.
• Real radiographs at evaluation time: DeepDRR projections expand training only, so reported fidelity reflects clinical input.

#Technical Details

CoarseDiff is a Brownian Bridge Diffusion Model with 1,000 timesteps over a 3D U-Net of 64 base channels with attention at the 16³ and 8³ resolutions. FineDiff's unconditional prior trains for 320 epochs before the ControlNet branch adds 150 more; the upscaler stacks six residual-in-residual dense blocks. Training data comes from LIDC-IDRI, where 232 of 1,018 chest CT scans carry paired real radiographs, split 197/10/25 for single-planar input and 149/8/16 for bi-planar; 786 further scans contribute DeepDRR projections. Volumes are resampled to 1 mm³, rescaled to 128³, and clipped to [-100, 900] HU, with X-rays standardized to 1024². Inference runs ten CoarseDiff and 100 FineDiff sampling steps, roughly 32 seconds per case on one A100. On LIDC-IDRI with single-planar input the pipeline reaches 21.21 dB PSNR and 0.540 SSIM against 19.68 dB and 0.524 for XctDiff, the strongest baseline there; bi-planar input raises this to 21.71 dB and 0.576, and at 256³ the bi-planar configuration reaches 22.03 dB and 0.628 against X2CT-GAN's 19.69 dB and 0.566. On the 220-case external cohort, acquired on Philips and Siemens scanners and unseen in training, AXON records 20.41 dB PSNR and 0.484 SSIM, gains of 8.0% and 16.9% over the strongest baseline.

#Applications

The target setting is any point of care where radiography is routine and CT is constrained by cost, dose, or availability: screening programs, resource-limited hospitals, follow-up imaging where a repeat scan is hard to justify. Reconstructed volumes show rib cage, lung parenchyma, and vascular structure in three dimensions from an exam the patient has already had. The stages are assembled from public implementations — CoarseDiff from Cor2Vox and BBDM, FineDiff from the diffusion U-Net and ControlNet of MAISI, the upscaler from a 3D ESRGAN — which makes the cascade a template for other cross-dimensional translation problems.

#Impact

AXON sets a reference point for X-ray-to-CT translation measured on clinical rather than simulated input, and the external-cohort result matters most: a model trained only on LIDC-IDRI transfers to other scanners and protocols without retraining. The authors are candid about the limits. The 256³ ceiling exceeds the 128³ common in prior work but stays below native clinical CT, and scaling further meets the memory cost of 3D diffusion. Evaluation rests on image-quality metrics rather than a downstream clinical task, so "high fidelity" means improvement over baselines, not agreement with ground truth. And because the translation is one-to-many, structures in ambiguous regions can be plausible without being veridical. The repository releases code for all three stages under GPL-3.0 but no pretrained checkpoints.