Abstract
Feed-forward 3D foundation models jointly predict dense geometry and camera parameters, but their direct transfer to colonoscopy is hindered by domain-specific appearance, non-rigid tissue motion, and close-range imaging. We introduce ColScale3R, a parameter-efficient adaptation of MapAnything for native-scale reconstruction from calibrated three-view colonoscopy windows. Weight-Decomposed Low-Rank Adaptation (DoRA) with rank-stabilized LoRA (rsLoRA) scaling is applied to 72 image-encoder projections, while all remaining components are frozen. This leaves only 5.886M parameters (0.477% of the network) trainable. A metric-preserving objective combines reference-view depth and pointmap supervision, metric-factor re-anchoring, and an adapter-disabled structural anchor that preserves pretrained ray and camera-pose predictions. On C3VD, ColScale3R achieves a mean three-view depth RMSE of 3.263 mm and a native fused Chamfer distance of 2.172 mm, compared with 8.580 mm and 6.069 mm, respectively, for a C3VD-trained Endo3R checkpoint using a single scale factor estimated only from the training split. It also outperforms zero-shot MapAnything and DA3 under the same evaluation protocol. Without target-domain training or test-time median scaling, ColScale3R reduces one-view SimCol depth RMSE from 155.94 cm to 1.12 cm. On the physical subset of PolypSense3D, the frozen model achieves a supplied-mask sizing MAE of 4.02 mm. Virtual-only Geometry-Preserving Scale Calibration and Aggregation (GSCA), applied solely to downstream diameter estimates, reduces the MAE to 2.90 mm. These results demonstrate parameter-efficient native-scale colonoscopic reconstruction with cross-dataset transfer and quantitative sizing capability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EndoLINA_015.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=NHL0txUxxW
BibTex
@InProceedings{KhaSaa_ColScale3R_MICCAISAT2026,
author = { Khalil, Saad AND Hwang, Youngbae},
title = { { ColScale3R: Parameter-Efficient Native-Scale Adaptation of Feed-Forward 3D Reconstruction for Colonoscopy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17275},
month = {pending},
page = {pending}
}
