List of Papers Browse by Subject Areas Author List
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. Our data and code will be publicly available at https://nubagcilab.github.io/BronchoAccess/.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2633_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://nubagcilab.github.io/BronchoAccess/
Link to the Dataset(s)
https://nubagcilab.github.io/BronchoAccess/
BibTex
@InProceedings{PenLin_AnatomyAware_MICCAI2026,
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},
title = { { Anatomy-Aware Prediction of Bronchoscopic Accessibility from 3D CT } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16893},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The contributions are three-fold: 1.To address this severely under-studied clinical challenge, the authors formalize bronchoscopy accessibility prediction as an end-to-end supervised deep learning task. This work presents the first deep learning application dedicated to predicting functional procedural outcomes directly from pre-operative CT scans. 2.The authors introduce an anatomy-aware MoE architecture that integrates volumetric imaging, anatomical priors, and airway geometry within a unified model. 3.The authors establish the first comprehensive benchmark for bronchoscopy accessibility prediction using 438 clinically verified cases and evaluate our framework against diverse computational baselines and human expert assessments.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The authors propose an AnatomyAware mixture-of-experts (MoE) model that integrates several specialized modules, including 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.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1、The authors use mask labels in the methodology section. They should discuss whether this may introduce a risk of label leakage. In addition, the paper does not explain how the mask labels are obtained during testing. 2、The method is mainly described in words, with a lack of rigorous mathematical formulation, which makes the overall pipeline somewhat unclear. 3、There are no references from the past three years in the manuscript, which raises concerns regarding the novelty of both the research direction and the proposed method. 4、The figure uses the term “Path Expert,” whereas the main text refers to it as “Path Geometry Expert. ” This inconsistency should be corrected. 5、The manuscript does not clearly describe how the outputs of the expert modules are incorporated into the gating network.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The score is assigned based on the authors’ descriptions of the methodology and experimental results.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have provided concise and targeted responses to my questions, and I have no further questions.
Review #2
- Please describe the contribution of the paper
The paper targets an underexplored clinical question: given a pre-op CT, will a bronchoscopy successfully reach a target lesion? So far the literature has been on tasks like airway segmentation, lesion detection and (the closest) path planning. The most related prediction tasks from CT has only been done with handcrafted geometric features.
The main contributions of the paper are:
- framing the prediction of the procedure’s success as an end-to-end supervised task (classification)
- a gated fusion model that combines a 3D CNN patch encoder, lobe identity embedding and BiGRU (over geometric features)
- a multi-center dataset (438 cases) with documented procedural outcomes (which authors promise to release).
The paper calls the fusion module a Mixture-of-Experts. In practice, what they have used is a soft-gated multi-modal fusion (not a sparse MoE). The core idea is: local morphology, anatomical region and the geometric profile of the approach path.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Most CAD work in bronchoscopy is either segmentation or planning. Predicting whether a procedure will succed from static CT affects patient triage. The paper makes a reasonable case that experts struggle with this decision.
2.The dataset contribution is the strongest. 438 cases from 3 centers with procedure outcome labels is very much valuable.
3.The path expert is the most defensible modelling choice. Converting the approach trajectory into a sequence of geometric features (and feeding it into a BiGRU) is a neat idea. The ablation shows that this module alone adds a +0.04 AUROC on top of CT alone.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The MoE framing is not defended properly. The key comparison is missing (simple concatenation of the three expert features followed by an MLP with no gating). Without this, the adaptive gating contribution cannot be separated from the decomposition itself.
2.The lobe expert adds almost noting alone (0.55 AUROC) and a marginal +0.03 in combination. A simple sanity check (concatenate a one-hot lobe vector to the CT features) is missing. If that matches the lobe expert, embedding + MLP framing is not justified.
3.Multi-center validation is not done despite having 3 centers. Pooled 5-fold CV with patient-level splits is okay, but leave-one-center-out would test whether the model generalizes across sites.
4.The dataset description misses what “success” means operationally. Is it based on the diagnostic yeild from the sample or is it a subjective call or something else? Details like per-center counts, peripheral vs central lesions breakdown, lesion size distributions, etc. can be reported.
5.The human performance comparison is hard to interpret because the aggregation protocol is not described. The metric of AUROC is computed from three probability values (mapped to the 3-level difficulty) and with atmost 3 threshold points vs. a continuous 0-1 model output, AUROC may not reflect the relevant evaluation properly. A reasonable fix is to report sensitivity and specificity at each of the 3 difficulty levels.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Rebuttal-fixable 1.Clarify the clinician protocol. Who rated what, with what inputs, and how was aggregation done. 2.Report expert performance using SN and SP along with AUROC at the 3-difficulty levels. 3.Add DeLong tests for the Table 2 comparison. 4.Add the concatenation-without-gating ablation. 5.Report per-center counts and other dataset statistics. 6.Clarify what procedure “success” means.
One observation on the MoE naming. What you have is dense gated multi-modal fusion. Consider renaming. This is not a scientific issue but it will confuse readers coming from the LLM literature.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper targets an underexplored problem and the dataset is a valuable contribution. The paper overclaims in two places. The MoE framing is not what the architecture actually does and the claim “beats experts” relies on an apples-to-oranges AUROC comparison between a 3-difficulty level human rating and a 0-1 model output. The ablations isolate the experts but not the fusion mechanism (so the MoE-specific claim is not supported). If the authors clarify the expert study protocol/details, add the fusion ablation, and run statistical significance tests on the same-input comparison, this paper can get into weak accept territory. Overall, the dataset alone has value for the community.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
1.The two ablations I asked for were provided. Without gating, the model drops from 0.8052 to 0.7070.With one-hot lobe instead of the learned embedding, it drops to 0.7405.So the gating and the embedding are doing real work. 2.The rebuttal also says the gating network takes anatomical context as input and not the expert features. This was not clear in the paper. It needs to be stated clearly in the camera-ready. 3.One thing the authors did not address: I pointed out that what they call Mixture-of-Experts is actually gated multi-modal fusion. Real MoE (Shazeer-style) is sparse while theirs is dense. The authors chose not to respond. The camera-ready should rename this. 4.EfficientNet-b0 in Table 2 is not statistically significantly different from their model. This is the fairest comparison and the paper should say so explicitly in the camera-ready. 5.Overall, the clinical problem is underexplored, the dataset is useful and the method works. The framing and reporting need fixes in the camera-ready.
Review #3
- Please describe the contribution of the paper
1.Integration of local morphology, lobe information and airway path sequence to the targeted lung lesion for end-to-end bronchoscopy accessibility prediction(binary class) using 3D CT. For encoding each of these component into latent sapce, separate networks 3DCNN, Embedding+MLP layers, bidirectional GRU are used. Finally the embeddings are weighted and fused using a gated network for binary classification. 2.To make the end-to-end bronchoscopy accessibility prediction as supervised learning, the authors created a dataset using documented clinical outcomes of 438 patients who underwent bronchoscopy for lung lesion.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.To the reviewer’s knowledge, this work is the first to formulate the supervised end-to-end bronchoscopy accessibility prediction problem. 2.The fusion of the features from three different modality(Raw CT, Lobe location, airway segmentation) 3.The quantitative performance of the proposed work over the compared state-of-the-arts method has big margin. 4.The curation of the largest dataset for bronchoscopy accessibility prediction from three clinical centers.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Even though, the application of fusion network is new to the formulated problem, the fusion framework is conceptually not novel. Similar fusion techniques are reported in [1],[2] 2.The bronchoscopy accessibility prediction(Success/Failure) can be due to any of the three components CT data, lobe location or airpathway. The proposed work is not addressing the reasons behind the prediction. 3.To ensure the high performance of the proposed work is not due to richer input representation, authors compared the work with CT based CNN networks using CT intensity, airway distance maps, and lesion masks. However, the comparison is not made using airway path sequence data with graph-based networks in Table2.This appears to give unequal advantage to the proposed network. 4.The lack of comparison of the proposed path geometry expert with the sequence networks such as Long Short-Term Memory (LSTM) or Temporal Convolutional Networks.
[1] Gao, Siheng, et al. “A parallel feature fusion network combining GRU and CNN for motor imagery EEG decoding.” Brain Sciences 12.9 (2022): 1233.[2] Y. Zhuoning, Z. Xun, and Y. Tianbao, “Hetero-convlstm: A deeplearning approach to traffic accident prediction on heterogeneous spatio-temporal data,” in Proc. 24th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2018, pp. 984–992.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
1.There is a mismatch between the visual representation (Fig.1) and textual explanation of the proposed Lobe expert. Visual representation mentions “it contains bi-directional Gated recurrent unit” whereas textual explanation mentions “embedding layer followed by MLP layers”.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The formulation of the bronchoscopy accessibility prediction problem, high marginal performance of the proposed work over the other methods and curation of the dataset for the problem made the reviewer to opt acceptance. However, the lack of novelty in fusion method, non-ideal comparison in table 2, and the lack of reasons for the prediction made the reviewer to opt weak acceptance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal from author’s clarified the interpretability issues raised earlier. Despite the usage of existing fusion network, the novel formulation of bronchoscopy accessibility prediction problem and curation of multicenter dataset opened way for research direction.
Author Feedback
We thank the reviewers for their constructive feedback. Our ablations and main results validate our architectural design choices. Please see our responses below:
1.Anatomy-Conditioned Gating vs. Generic Fusion (R1, R2, R3). R2 and R3 rightly questioned whether gains stem from MoE generically or from anatomy-conditioned design specifically. This is an important question, and our ablation provides a clear answer. Replacing our gating with Concat+MLP under identical inputs collapses AUROC from 0.8052 to 0.7070, a substantial 9.8% absolute drop. This single experiment isolates the entire contribution: it is not the mixture that matters, it is the anatomical conditioning of the mixture. Even a 2-expert gated variant without the Lobe Expert (AUROC 0.7756) vastly outperforms 3-expert concatenation (0.7070), proving that gating superiority is independent of problem decomposition. To clarify R1: the gating network never consumes expert outputs. It routes purely from anatomical context via w=Softmax(MLP([pos;lobe_emb;stats])); we will make this explicit in §2.2. 2.The Lobe Module Paradox (R2). R2 noted the standalone Lobe Expert is weak (AUROC 0.5513) and questioned its value. This skepticism is well-founded on its face, but ablations reveal a more nuanced picture. Lobe context enters via two synergistic pathways: as an expert and as a gating condition. Removing the Lobe Expert while keeping its embedding in the gate drops AUROC to 0.7756.Removing lobe from the gate entirely drops it to 0.7494.Replacing the learned embedding with naive one-hot drops it to 0.7405.The pattern is clear: lobe information is indispensable not as a standalone predictor but as an anatomical context anchor that teaches the gating network where it is operating. We will reframe accordingly.
3.Path Expert: Why Sequences, Not Graphs (R3). R3 correctly intuited that airway topology might favor graphs. However, our input is a linearized trachea-to-target path ordered by generation, inherently a 1D sequence. Evaluating alternative backbones on identical features: BiGRU achieves AUROC 0.6040 vs. TCN at 0.5446 and LSTM at 0.5234, confirming BiGRU as optimal. Crucially, all standalone path AUROCs remain modest, yet the full model reaches 0.8052.The leap is entirely attributable to multimodal fusion, not backbone choice, which is precisely our thesis.
4.Interpretability and Clinical Necessity (R2, R3). R3 asked for case-level attribution: gating weights w provide this natively. w_ct rises for central lesions relying on morphology; w_path spikes for distal/lower-lobe lesions governed by airway geometry. On clinical difficulty (R2): 22 clinicians blindly rated anonymized CTs without procedural records. Results are revealing: at strict thresholds, SN=0.50±0.13, SP=0.63±0.19; relaxing raised SN to 0.80 but collapsed SP to 0.27.Inter-rater agreement was poor (6.2% unanimous, 64% spanned all difficulty levels). This high human variance underscores both the task’s difficulty and the value of a consistent objective predictor.
5.Clarifications (R1, R2, R3). No label leakage (R1): masks are from frozen models trained on disjoint datasets. Dataset (R2): “success” means pathology-confirmed tissue, an objective ground truth; the cohort (73% peripheral) reflects clinical reality. To our knowledge, this is the first dataset mapping CT to procedural yield, which is a proof-of-concept milestone; larger multi-center studies with site-stratified validation are future work. DeLong (R2): all baselines except EfficientNet-b0 differ significantly (p<0.05); Table 2 will be annotated. Fig.1 (R1, R3): swapped labels are a figure-only typo; code matches text exactly; figure will be redrawn. References (R1): we will add 2023-2025 citations including Naito (IJCARS 2023, already as [19]) and recent airway/multimodal works.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper receives mixed reviews and rebuttal is invited.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper initially received 2 negative and 1 positive reviews. After rebuttal, authors address most of the concerns and all reviewers agree to accept this work.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All three reviewers recommend acceptance after rebuttal. Reviewers agreed that bronchoscopy accessibility prediction is an important and underexplored problem, and that the dataset is a valuable contribution to the community. The rebuttal successfully clarified several key concerns regarding the gating mechanism, the role of the lobe expert, the clinician comparison protocol, and the contribution of anatomy-conditioned fusion beyond simple feature concatenation. While the architectural novelty is moderate and some presentation issues remain, the task formulation, dataset, and empirical results together make a meaningful contribution. The presentation could also be improved, including clearer methodological descriptions, figures, and terminology. I recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes a clinically useful method, and its strong performance was demonstrated through the experiments. Therefore, I believe this paper is worthy of acceptance.
