Summary
3D U-NetR applies volumetric convolutions so the model can use relationships across adjacent slices instead of treating every slice independently. That improves continuity along the volume axis and establishes a baseline for 3D-aware reconstruction in the lab’s public work.
This matters because CT data is naturally volumetric: anatomical structures extend across slices, and slice-wise processing can miss that continuity. The paper explores what changes when reconstruction is allowed to use neighboring slice information directly.
Method
The architecture replaces standard 2D processing with 3D convolutions so the network can operate on volumetric patches. That gives the model access to inter-slice context and helps it produce reconstructions that are more coherent through the stack rather than optimized one slice at a time.
Why it matters
Even as a preprint, this paper is useful as a public baseline for the lab’s volumetric reconstruction direction. It shows an early transition from single-slice enhancement toward models that better reflect the structure of CT volumes.
Highlights
- Moves from 2D slice processing to volumetric reconstruction.
- Targets inter-slice consistency, not just per-slice denoising quality.
- Represents an early public step toward richer CT context modeling.