Abstract
Endoluminal localization is essential for endoluminal navigation, yet many navigation systems depend on preoperative computed tomography, electromagnetic tracking, or explicit registration. In many clinical scenarios, however, coarse anatomical localization is more practical than metric pose estimation: the physician needs to know which airway region the endoscope is traversing and whether the trajectory is anatomically plausible. We present AirwayTopoLoc, a topology-aware vision-language framework for CT-free bronchoscopic localization. We formulate the task as video-based airway sequence labeling, where every frame is assigned an anatomical airway label while maintaining temporal and bronchial-tree consistency. The proposed method adapts Qwen3-VL to bronchoscopy videos using LoRA-based supervised fine-tuning, extracts temporal video hidden states, and attaches a lightweight emission head for dense airway-label scoring. A linear-chain CRF is used during training to learn data-driven transition preferences without a hard topology mask, while a hard-topology Viterbi decoder is applied during inference to enforce anatomically valid trajectories. On a bronchoscopy video dataset with 27 airway labels, AirwayTopoLoc achieves 85.26% accuracy and 60.17% Macro-F1 with zero invalid transitions, outperforming visual encoder baselines and structured ablations. These results demonstrate that combining VLM representations with topology-constrained sequence decoding is effective for bronchoscopic anatomical localization.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EndoLINA_005.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=Vhii9g4dVJ
BibTex
@InProceedings{LuoMin_TopologyConstrained_MICCAISAT2026,
author = { Luo, Mingyi AND Wu, Junyang AND Zhang, Minghui AND Liu, Yaoyu AND Chen, Yirong AND Zhang, Hanxiao AND Li, Qingbiao AND Gu, Yun},
title = { { Topology-Constrained Vision-Language Sequence Modeling for Bronchoscopic Localization } },
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}
}