Summary
The paper removes the dependency on paired low-dose and normal-dose data by training directly from low-dose measurements. It adapts split-measurement style self-supervision to the CT setting and shows that clean targets are not always required to learn a useful reconstruction model.
The key move is to apply self-supervision in the projection domain, where the noise assumptions are better aligned with the requirements of the training scheme. Rather than treating CT as a generic image denoising problem, the paper uses the structure of the acquisition process itself.
Method
The approach uses low-dose sinograms as their own training targets by splitting measurements and optimizing the reconstruction pipeline without access to normal-dose references. In practice, this means the filtering stage of FBP and the parameters of a denoising network are learned together under self-supervision.
Why it matters
This conference paper establishes one of the lab’s central research directions: reconstruction methods that remain usable when paired clean data is unavailable. It is also the direct conceptual precursor to the later Proj2Proj journal paper, which develops the idea into a more complete and validated framework.
Highlights
- Tailors self-supervised learning to CT projection data rather than generic images.
- Establishes a practical route for labs and clinics without paired training datasets.
- Serves as a direct precursor to the later Proj2Proj journal work.