Summary
Proj2Proj trains directly in the projection domain by partitioning low-dose CT measurements into complementary subsets and learning a mapping between them. That lets the model improve reconstruction quality without relying on paired standard-dose reference scans.
Compared with the earlier conference version of the idea, this paper develops a fuller methodological treatment and a more mature experimental pipeline. It positions projection-domain self-supervision as a practical answer to the data availability problem in low-dose CT.
Method
The model learns from complementary projection subsets rather than from low-dose/normal-dose image pairs. This keeps the supervision signal tied to the measurements actually produced by the scanner and allows the reconstruction network to be trained in a way that respects the acquisition process.
Why it matters
Proj2Proj is important because it turns a promising self-supervised idea into a reproducible journal-level result. It provides a public reference point for later lab work on task-aware reconstruction, data-efficient training, and projection-domain learning.
Highlights
- Extends the lab’s earlier self-supervised CT work into a complete journal-level pipeline.
- Keeps supervision aligned with the measurements that are actually available in practice.
- Provides a reproducible foundation for follow-up work on data-efficient LDCT reconstruction.