Loading…
Skeletal CT reconstruction from two orthogonal X-ray projections, using conditional latent diffusion that classifies scoliosis as the volume forms.
No providers recorded yet. Browse all providers
An orthopedic clinic images a curved spine with two X-rays rather than a CT, because the CT dose is substantially higher and the patient is often an adolescent who will be re-imaged for years. The cost is that both projections flatten every vertebra onto a plane: axial rotation, the depth of a curve, and the contour of a joint surface all have to be read out of overlapping shadows. SkeDiff treats the missing third dimension as something to generate rather than to measure. It is a conditional latent diffusion model that takes a pair of orthogonal radiographs and denoises its way to the 3D skeletal CT volume those views are consistent with.
The model comes from the group of Rongjun Ge and Yang Chen at Southeast University, working with the Shandong Computer Science Center at Qilu University of Technology in Jinan, and was published in the IEEE Journal of Biomedical and Health Informatics. Three trained components make up the system: a cross-dimensional conditional encoder, a 3D diffusion model that runs in a compressed latent space, and a diffusion-based scoliosis classifier that reads a diagnosis out of the same reverse process. All three are fixed after training: a new patient's radiographs are inputs, not something the network is re-fit to.
X-ray-to-CT reconstruction has moved from regression CNNs through GAN generators to diffusion, and the catalog holds neighbours on either side of this one: DeepSparse reconstructs cone-beam CT from a handful of sparse projections, while X2CT-CLIP aligns chest radiographs with CT embeddings instead of producing a volume at all. What distinguishes SkeDiff is the combination of a two-view budget, denoising in latent rather than pixel space, and a clinical readout produced by the generative process itself.
SkeDiff reconstructs volumes at 128×128×128 and takes roughly 900 ms per case at about 12 GB of GPU memory — moderate for a 3D generative model, and a direct consequence of denoising in a latent space rather than over voxels. Training uses SGD rather than an adaptive optimizer, which the authors found generalized better in the low-data regime typical of paired CT and radiograph collections, at a global batch size of 8 across four GPUs for 200 epochs. On the spine dataset, supplementary sensitivity runs put PSNR near 33.3 dB and SSIM near 0.81; repeating training under three random seeds moved those metrics by about 1%, and varying the global batch size from 8 to 16 left them essentially unchanged. Evaluation covers spine, hip, and knee datasets against recent baselines, including multi-view methods that consume comparable or greater compute.
The clear use is orthopedic assessment where CT is unavailable, unjustified by dose, or slower than the clinical question allows: emergency triage on the radiographs already taken, pre-operative appreciation of hip and knee geometry, and follow-up of adolescent spinal curvature, where repeated CT is what clinicians try to avoid. Because the scoliosis classifier runs inside reconstruction, a screening workflow can get a curvature category and an inspectable 3D volume from one pass over a biplanar study.
SkeDiff is a worked argument that conditional generative modelling, not better regression, is the productive way to recover 3D skeletal anatomy from two projections — and that a diagnostic head can be taken from the diffusion trajectory instead of bolted on downstream. The honest boundaries are narrow ones. Evaluation is confined to skeletal anatomy, with no claim of transfer beyond bone; results are computational comparisons against CT ground truth rather than a clinical study; and a generated volume is a plausible reconstruction rather than a measurement, which matters near a surgical decision. The authors' repository carries the training and evaluation code alongside the paper's supplementary material, but publishes no trained checkpoints and no license.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.