This paper reconstructs 3D X-ray computed-tomography volumes from incomplete measurements - sparse views and limited angles - acquired by a proposed multi-source static CT (MSCT) scanner that fixes several X-ray sources around the ring to cut both radiation dose and scan time. Reconstruction is an ill-posed linear inverse problem, and the authors solve it with a conditional denoising-diffusion process: a pretrained diffusion image prior generates each slice by reverse-time sampling, and at every step the sample is projected onto the affine set of volumes consistent with the measured projections to enforce data fidelity. To get a continuous, resolution-flexible high-resolution volume, the 3D phantom is parameterized implicitly by a coordinate neural network (implicit neural representation), and a self-supervised objective using the measured data refines it and suppresses accumulated sampling error. The combined method, DIP-ASPINS (diffusion image prior + affine-set projection + INR + self-supervised learning), is evaluated on simulated sparse-view and limited-angle MSCT settings and shown to improve reconstruction quality.
This paper reconstructs 3D X-ray computed-tomography volumes from incomplete measurements - sparse views and limited angles - acquired by a proposed multi-source static CT (MSCT) scanner that fixes s…