Training-Free 3D CT From Sparse X-Rays Using Diffusion Priors

A team of researchers (Zhenkai Zhang, Markus Hiller, Krista A. Ehinger, Tom Drummond) has proposed TF-PRDiT, a training-free conditional sampling framework that repurposes a frozen voxel-level 3D Diffusion Transformer as a universal inverse problem solver for medical imaging. Unlike traditional supervised methods that require retraining for each task and measurement setting, TF-PRDiT leverages a predictor-corrector sampler combined with likelihood-based guidance on denoised predictions. This enforces measurement consistency via a task-specific forward operator—such as a differentiable Digitally Reconstructed Radiograph (DRR) projector for X-ray-to-CT reconstruction—without updating model weights. The method is validated on the LIDC-IDRI dataset, where it reconstructs 3D CT volumes from as few as 1 to 12 input X-rays, with quality consistently improving as more views are added.

Beyond X-ray-to-CT, TF-PRDiT demonstrates remarkable versatility: by simply swapping the forward operator, the same frozen model can perform 3D super-resolution, volumetric infilling, and deblurring with no retraining. This showcases how a single pretrained 3D diffusion prior can serve as a universal solver for diverse volumetric medical inverse problems. The approach breaks the traditional dependency on task-specific supervised models, offering a scalable, efficient path for clinical applications where training data or computing resources are limited. The paper (arXiv:2606.20763) provides full experimental details and comparisons.