NeRF 기반 CT 재구성에서 광선 샘플링과 레이어 정규화의 기여도 분석

초록

CT enables precise examination of three-dimensional anatomical structures inside the human body, but it also exposes patients to high radiation because it requires many X-ray projections. This has increased the demand for sparse-view CT, which reconstructs volumes from fewer images. NeRF has opened a promising route for efficient sparse-view CT reconstruction, and recent work seeks improvements not only via network design but also through training choices such as ray sampling and normalization. Here, we compared several importance-based ray sampling strategies under identical settings and found that preserving global coverage in ray selection is critical for reconstruction quality. We further show that applying layer normalization across sampling conditions stabilizes internal feature scales, improving training stability and final reconstruction performance.

키워드