Abstract
3D point cloud registration in laparoscopic surgery estimates the transformation between an intraoperative organ reconstructed from video and its preoperative mesh. Since ground-truth transformations are unavailable for real data, supervised networks are trained on synthetic organ pairs. At test time, however, real reconstructions differ from synthetic ones and are noisy, sparse and occluded, degrading correspondence estimation. Test-time adaptation (TTA) methods can address this domain shift at inference and have been applied successfully on tasks such as classification and segmentation. However, these methods typically rely on logits, entropy, class prototypes, or cache memory mechanisms unavailable in registration. Moreover, TTA methods assume a single shifted input, whereas registration involves a pair with an asymmetric shift, which mainly affects the intraoperative cloud, making TTA for registration much more challenging. This paper provides state-of-the-art TTA methods for 3D registration tasks across three families: model, normalization, and input adaptation. In particular, we analyze and modify four representative methods from those families that perform: auxiliary-task
model update, backpropagation-free token purging, feature alignment, and layer-normalization calibration. TTA methods are modified to account for the asymmetric domain shifts between the preoperative and intraoperative point clouds, and to replace the entropy-based mechanism from classification. Given a correspondence-based model trained on clean source synthetic data, two scenarios were considered for the target: synthetic data with corruption and real data. We experiment on the P2P and P2ILReg datasets. Eight corruptions were applied separately to those datasets on the synthetic target data only, such as uniform noise and global density decrease, with five increasing levels of sever ity. Our experiments show that all methods improve registration on the P2P dataset, whereas on P2ILReg only input adaptation reduces the error, while normalization adaptation degrades it. Given the computational overhead of the backpropagation-based method, input adaptation is a more promising family for laparoscopic surgery, with low inference latency and a consistent reduction in error across all datasets.
Our code: https://github.com/ninaa-git/survey_pc_registration_tta
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COLAS_004.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=J56meTbzGz
BibTex
@InProceedings{BodNin_Test_MICCAISAT2026,
author = { Bodelot, Nina AND Belharbi, Soufiane AND Granger, Eric},
title = { { Test Time Adaptation Methods for Point Cloud Registration in Laparoscopic Surgery } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS pending},
month = {pending},
page = {pending}
}