Unsupervised large-cone-angle artifact suppression for CBCT reconstruction via neural radiance fields

Cone-beam computed tomography (CBCT) is vital for clinical imaging, but reconstructing data acquired from circular trajectories inherently suffers from severe large-cone-angle artifacts (LCAA) due to null space deficiency. While existing methods can reduce artifacts, they typically require extensive prior information or paired data, limiting clinical applicability. To address these issues, we propose an unsupervised Neural Radiance Fields framework (LCAA-NeRF) to suppress LCAA directly from projection data. First, we design an axial-aware anisotropic scaling hash encoding ( A 3 Hash) mechanism to enhance the representational capacity for large cone angles. Second, we introduce a path-length-adaptive ray sampling (PLARS) strategy to dynamically capture features across varying ray paths. Finally, we incorporate a stochastic structural similarity (S3IM) loss to further enforce geometric consistency. The superiority and robustness of our method are validated across both simulated and real datasets under large cone angles.