Seeing Down the Line: Endoscopic Reconstruction with Centerline Constraints

Andrea Dunn Beltran, Romain Hardy, Pranav Rajpurkar
Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, PMLR 315:360-378, 2026.

Abstract

Colonoscopy remains the gold standard for colorectal cancer screening, but there is still no real-time, geometry-aware way to quantify which parts of the colon have been inspected during a procedure. We revisit 3D Gaussian endoscopic reconstruction as a representation and geometry problem rather than a new network design. Assuming known camera poses and off-the-shelf depth or photometric supervision, we add a simple centerline-based coordinate system and priors on top of an existing Gaussian mapping backbone. From the noisy pose stream we maintain an online centerline and Bishop frame, assign each Gaussian tubular coordinates $(s,r,\theta)$, and use these coordinates both to regularize the map toward a hollow tube and to accumulate coverage statistics in colon-intrinsic space. On long C3VD phantom colonoscopy sequences, this lightweight modification achieves Chamfer distance comparable to or better than an endoscopy-specific 3D Gaussian SLAM baseline while running at frame rates close to MonoGS and yielding improved rendering quality, with negligible additional computation. At the same time, the same representation produces unrolled colon views and segment-wise coverage summaries essentially "for free", making centerline-aware Gaussian mapping a practical drop-in component for future real-time quality monitoring tools in colonoscopy.

Cite this Paper

BibTeX
@InProceedings{pmlr-v315-dunn-beltran26a, title = {Seeing Down the Line: Endoscopic Reconstruction with Centerline Constraints}, author = {Dunn Beltran, Andrea and Hardy, Romain and Rajpurkar, Pranav}, booktitle = {Proceedings of The 9th International Conference on Medical Imaging with Deep Learning}, pages = {360--378}, year = {2026}, editor = {Huo, Yuankai and Gao, Mingchen and Kuo, Chang-Fu and Jin, Yueming and Deng, Ruining}, volume = {315}, series = {Proceedings of Machine Learning Research}, month = {08--10 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v315/main/assets/dunn-beltran26a/dunn-beltran26a.pdf}, url = {https://proceedings.mlr.press/v315/dunn-beltran26a.html}, abstract = {Colonoscopy remains the gold standard for colorectal cancer screening, but there is still no real-time, geometry-aware way to quantify which parts of the colon have been inspected during a procedure. We revisit 3D Gaussian endoscopic reconstruction as a representation and geometry problem rather than a new network design. Assuming known camera poses and off-the-shelf depth or photometric supervision, we add a simple centerline-based coordinate system and priors on top of an existing Gaussian mapping backbone. From the noisy pose stream we maintain an online centerline and Bishop frame, assign each Gaussian tubular coordinates $(s,r,\theta)$, and use these coordinates both to regularize the map toward a hollow tube and to accumulate coverage statistics in colon-intrinsic space. On long C3VD phantom colonoscopy sequences, this lightweight modification achieves Chamfer distance comparable to or better than an endoscopy-specific 3D Gaussian SLAM baseline while running at frame rates close to MonoGS and yielding improved rendering quality, with negligible additional computation. At the same time, the same representation produces unrolled colon views and segment-wise coverage summaries essentially "for free", making centerline-aware Gaussian mapping a practical drop-in component for future real-time quality monitoring tools in colonoscopy.} }
Endnote
%0 Conference Paper %T Seeing Down the Line: Endoscopic Reconstruction with Centerline Constraints %A Andrea Dunn Beltran %A Romain Hardy %A Pranav Rajpurkar %B Proceedings of The 9th International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2026 %E Yuankai Huo %E Mingchen Gao %E Chang-Fu Kuo %E Yueming Jin %E Ruining Deng %F pmlr-v315-dunn-beltran26a %I PMLR %P 360--378 %U https://proceedings.mlr.press/v315/dunn-beltran26a.html %V 315 %X Colonoscopy remains the gold standard for colorectal cancer screening, but there is still no real-time, geometry-aware way to quantify which parts of the colon have been inspected during a procedure. We revisit 3D Gaussian endoscopic reconstruction as a representation and geometry problem rather than a new network design. Assuming known camera poses and off-the-shelf depth or photometric supervision, we add a simple centerline-based coordinate system and priors on top of an existing Gaussian mapping backbone. From the noisy pose stream we maintain an online centerline and Bishop frame, assign each Gaussian tubular coordinates $(s,r,\theta)$, and use these coordinates both to regularize the map toward a hollow tube and to accumulate coverage statistics in colon-intrinsic space. On long C3VD phantom colonoscopy sequences, this lightweight modification achieves Chamfer distance comparable to or better than an endoscopy-specific 3D Gaussian SLAM baseline while running at frame rates close to MonoGS and yielding improved rendering quality, with negligible additional computation. At the same time, the same representation produces unrolled colon views and segment-wise coverage summaries essentially "for free", making centerline-aware Gaussian mapping a practical drop-in component for future real-time quality monitoring tools in colonoscopy.
APA
Dunn Beltran, A., Hardy, R. & Rajpurkar, P.. (2026). Seeing Down the Line: Endoscopic Reconstruction with Centerline Constraints. Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 315:360-378 Available from https://proceedings.mlr.press/v315/dunn-beltran26a.html.

Related Material