Anatomy-Aware Prediction of Bronchoscopic Accessibility from 3D CT
Authors: Linkai Peng, Cuiling Sun, Bin Wang, Jamie Rowell, Catherine Gao, Oyku Ikizgul, Eminenur Sentasci, Andrea Bejar, +5 more
Organizations: Department of Electrical and Computer Engineering, Northwestern University · Department of Computer Science, Northwestern University · Department of Medicine, Northwestern University · Department of Radiology, Istanbul Faculty of Medicine · Department of Radiology, Northwestern University
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. Our data and code will be publicly available at https://nubagcilab.github.io/BronchoAccess/.
Figures & tables
Figure 1: Overview of the proposed anatomy-aware mixture-of-experts framework for bronchoscopy accessibility prediction.
Input
Method
Acc.
F1
AUROC
PRAUC
Params
Geometric Features
Logistic Regression [ 9 ]
0.6136
0.7536
0.5128
0.6897
N/A
Random Forest [ 2 ]
0.5682
0.7031
0.4590
0.5785
N/A
CT Volumes
Densenet121 [ 11 ] †
0.5909
0.7313
0.5301
0.6543
11.24M
SwinTransformer [ 15 ] †
0.6136
0.7499
0.5418
0.8069
38.50M
Efficientnetb0 [ 21 ] †
0.6022
0.7517
0.5725
0.7459
4.69M
ConvNeXt [ 16 ] †
0.6023
0.7482
0.5982
0.7750
31.30M
Table 1: Performance comparison of the proposed method against various baseline models. Human assessments were provided by 22 clinicians based on CT volumes and target lesion locations. † indicates statistically significant difference compared with the proposed method (DeLong test, p < 0.05).
Method
Acc.
F1
AUROC
PRAUC
SwinTransformer † [ 15 ]
0.6250
0.7626
0.5243
0.6698
ConvNeXt † [ 16 ]
0.6477
0.7438
0.6629
0.7057
Densenet121 † [ 11 ]
0.6590
0.7058
0.6640
0.7611
Resnet18 † [ 8 ]
0.6818
0.7586
0.7109
0.7834
Efficientnetb0 [ 21 ]
0.6704
0.7819
0.7310
0.8269
Ours
0.8068
0.8595
0.8052
0.8496
Table 2: Comparison with CT-based deep networks trained using multi-channel volumetric inputs (CT intensity, airway distance transform, and lesion mask). † indicates statistically significant difference compared with the proposed method (DeLong test, p < 0.05).
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
Table 3: Ablation study evaluating the contribution of each expert module. The full model integrates all experts through adaptive gating.
Figure 2: Representative qualitative examples illustrating true positive (TP), false positive (FP), false negative (FN), and true negative (TN) predictions. Airway trees are visualized together with target lesion locations (red markers).