Semantic-Aware 3D Polyp Segmentation in Dynamic Endoscopy Scenes
Authors: Liu, J., Tang, W., Shi, Y., Huang, D.
Journal: Proceedings International Symposium on Biomedical Imaging
Publication Date: 01/01/2026
Volume: 2026-April
eISSN: 1945-8452
ISSN: 1945-7928
DOI: 10.1109/ISBI61048.2026.11515363
Abstract:This paper proposes the first 3D polyp segmentation method based on Neural Radiance Fields (NeRFs) for dynamic endoscopy scenes. A two-staged training approach, SAPolyp-NeRF (Segment Anything Polyp in NeRFs), is introduced. The first stage is to reconstruct the dynamic implicit field of soft tissue. Then, a multi-view self-prompt mechanism is employed to automatically predict 2D polyp masks for the rendered views generated by the dynamic implicit field. The second stage is a 2D-to-3D segmentation module, in which a view-alternating weak supervision strategy is introduced to train SAPolypNeRF for the second time using 2D polyp masks supervision aggregated from multiple training views to produce a dynamic 3D mask implicit field for 3D polyp segmentation. The experimental results demonstrate that our method achieves higher-quality comprehensive reconstruction and accurate segmentation of the lesion region than the state-of-the-art 3D segmentation methods, resulting in the best overall performance.
Source: Scopus