Anatomy-Aware Prediction of Bronchoscopic Accessibility from 3D CT

Linkai Peng¹, Cuiling Sun², Bin Wang¹, Jamie Rowell³, Catherine Gao³, Oyku Ikizgul⁴, Eminenur Sentasci⁵, Andrea Bejar⁵, Halil Ertugrul Aktas⁵, Gorkem Durak⁵, Momen Wahidi³, Christopher Kapp³, Ulas Bagci⁵†

¹Department ofElectrical and Computer Engineering, Northwestern University ²Department ofComputer Science, Northwestern University
³Department of Medicine, Northwestern University ⁴Department of Radiology, Istanbul Faculty of Medicine
⁵Department of Radiology, Northwestern University

MICCAI 2026
†Corresponding author: ulas.bagci@northwestern.edu

arXiv Code 🤗 HuggingFace

Abstract

Pre-operative planning for bronchoscopy is critical for the diagnosis of lung lesions. Current accessibility assessment relies on subjective manual inspection of CT scans, which is time-consuming and prone to inter-observer variability. In this paper, we formalize bronchoscopy accessibility prediction as a novel supervised learning task and present the first end-to-end framework to address it. We propose an Anatomy-Aware Mixture-of-Experts (MoE) model that integrates specialized modules: a CT Expert for local morphological features, a Lobe Expert for anatomical priors, and a Path Geometry Expert that encodes the sequential constraints of the bronchial tree. To support this task, we curated the first clinical dataset of 438 cases with pre-operative CT scans and documented procedural outcomes. Experimental results demonstrate that our method achieves an AUROC of 0.8052, significantly outperforming both state-of-the-art baselines and experienced human experts. This work establishes a new benchmark for computer-aided interventional planning in pulmonary medicine.

Framework overview: anatomy-aware adaptive expert fusion.

Quantitative Results

Comparison with Diverse Baselines

Conventional ML on handcrafted features shows limited discriminative ability; CT-based deep networks reach at best 0.6333 AUROC, and graph neural networks up to 0.6483. Human experts achieve only 0.5661 AUROC, indicating that visual CT inspection alone is insufficient. Our hybrid model reaches 0.8052 AUROC with just 3.83M parameters.

Input	Method	Acc.	F1	AUROC	PRAUC	Params
Geometric Features	Logistic Regression	0.6136	0.7536	0.5128	0.6897	N/A
Geometric Features	Random Forest	0.5682	0.7031	0.4590	0.5785	N/A
CT Volumes	DenseNet121†	0.5909	0.7313	0.5301	0.6543	11.24M
	Swin Transformer†	0.6136	0.7499	0.5418	0.8069	38.50M
	EfficientNet-b0†	0.6022	0.7517	0.5725	0.7459	4.69M
	ConvNeXt†	0.6023	0.7482	0.5982	0.7750	31.30M
	ResNet18†	0.5795	0.6185	0.6333	0.7700	33.16M
Airway Graph	GAT†	0.6386	0.7794	0.5767	0.7115	52.35K
	GIN†	0.6265	0.7257	0.5774	0.6782	101.12K
	GraphSAGE†	0.6203	0.7541	0.5897	0.7129	86.14K
	GCN†	0.6329	0.7752	0.6483	0.7666	51.59K
—	Human Experts	0.5272	0.5423	0.5661	0.7434	N/A
Hybrid	Ours	0.8068	0.8595	0.8052	0.8496	3.83M

† indicates a statistically significant difference compared with the proposed method (DeLong test, p < 0.05). Human assessments provided by 22 clinicians.

Comparison Under Identical CT Inputs

To verify the gain does not come from richer inputs alone, we compare against CT-based deep networks trained on the same multi-channel inputs (CT intensity, airway distance map, lesion mask). Our model still leads clearly, confirming the benefit comes from architectural decomposition and adaptive fusion.

Method	Acc.	F1	AUROC	PRAUC
Swin Transformer†	0.6250	0.7626	0.5243	0.6698
ConvNeXt†	0.6477	0.7438	0.6629	0.7057
DenseNet121†	0.6590	0.7058	0.6640	0.7611
ResNet18†	0.6818	0.7586	0.7109	0.7834
EfficientNet-b0	0.6704	0.7819	0.7310	0.8269
Ours	0.8068	0.8595	0.8052	0.8496

† indicates a statistically significant difference (DeLong test, p < 0.05).

Ablation Study

We evaluate the contribution of each expert. The CT Expert is the strongest single branch (0.7382 AUROC); the Lobe Expert performs near chance alone, consistent with its role as a global prior. Combining all three through adaptive gating gives the best result.

Lobe Expert	Path Expert	CT Expert	Acc.	F1	AUROC	PRAUC
✓	—	—	0.6190	0.7647	0.5513	0.7030
—	✓	—	0.6310	0.7704	0.6040	0.6665
—	—	✓	0.7500	0.8035	0.7382	0.8028
✓	✓	—	0.7045	0.7968	0.5960	0.6857
✓	—	✓	0.7841	0.8504	0.7477	0.8141
—	✓	✓	0.7500	0.8358	0.7756	0.8030
✓	✓	✓	0.8068	0.8595	0.8052	0.8496

Qualitative Analysis

Representative TP, FP, FN, TN predictions with airway trees and lesion locations

Representative true positive (TP), false positive (FP), false negative (FN), and true negative (TN) cases. Airway trees are rendered together with target lesion locations (red markers). The TP case shows a lesion near a well-connected branch with sufficient diameter; the FP case sits close to the distal bronchi but with abrupt branching changes, where the CT Expert may overweigh local proximity; the FN involves a peripheral lesion that may still be sampled in practice via operator expertise or ultra-thin bronchoscopes; the TN is a deeply peripheral lesion with sparse connectivity. Accessibility depends on both local lesion–airway adjacency and cumulative geometric constraints.

BibTeX

@inproceedings{peng2026bronchoacc,
  title={Anatomy-Aware Prediction of Bronchoscopic Accessibility from 3D CT},
  author={Peng, Linkai and Sun, Cuiling and Wang, Bin and Rowell, Jamie and Gao, Catherine and Ikizgul, Oyku and Sentasci, Eminenur and Bejar, Andrea and Aktas, Halil Ertugrul and Durak, Gorkem and Wahidi, Momen and Kapp, Christopher and Bagci, Ulas},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)},
  year={2026}
}