Abstract
Automated tooth and root-canal delineation in cone-beam computed tomography (CBCT) requires both accurate foreground seg- mentation and correct anatomical identification. STSR 2026 Task 1 makes this explicit with a 65-class target and only 30 annotated volumes along- side 299 unannotated ones. We train a 3D nnU-Net on 24 annotated cases and a 40-case unannotated subset with supervised segmentation and a detached weak-to-strong consistency objective. A training-free correction then estimates a patient-specific affine-coordinate midline from predicted teeth and relabels only tooth components whose FDI laterality conflicts with their physical side, leaving root canals untouched. On the 20-case official validation set the correction raised IA from 0.2518 to 0.5493 and Challenge Score from 0.5158 to 0.7207, while image-level Dice, mIoU and normalized surface Dice stayed at 0.9563, 0.9166 and 0.9934. Memory- aware inference stayed within 8 GB and 60 s per case on an NVIDIA A10. Team djw placed sixth in the final Task 1 evaluation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ODIN_challenges_002.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=UjlpuQe4ag
BibTex
@InProceedings{DuYuk_GeometryGuided_MICCAISAT2026,
author = { Du, Yukun AND Ji, Yude AND Xu, Wei AND Zhang, Rui},
title = { { Geometry-Guided Identity Correction for Semi-Supervised CBCT Tooth and Root-Canal Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17270},
month = {pending},
page = {pending}
}