Source-linked AI summary
BronchoTop: Bronchoscopy Navigation via RGB-Only Topological Localization
Clara Tomasini, Ana Cristina Murillo, Luis Riazuelo
TL;DR
Bronchoscopy localization must help clinicians track the scope through a complex airway tree, while existing navigation methods often rely on patient-specific CT or additional sensors. BronchoTop estimates location from RGB video within a generic airway model using a four-module pipeline. Its full pipeline improves accuracy by 21% over the version without switch verification, with comparable results across sequences.
Problem
Clinicians must mentally track the scope through a complex airway tree, and misidentifying segments can lengthen procedures and lead to missed lesions.
Source: §I · p. 1Method
BronchoTop estimates scope location online from RGB bronchoscopy video in a generic airway model using lumen tracking, branch association, probability estimation, and switch verification.
Source: §I · pp. 1–2Results
21% increase in accuracy for full-pipeline BronchoTop (A+B+C+D) over its ablated version without Switch Verification (A+B+C), with comparable results across sequences.
Source: D. Topological Localization performance · p. 6Takeaways & Limitations
Topological localization in a generic, clinically used airway model provides sufficient spatial context for navigational guidance.
Source: §I · p. 1
Abstract
from arXiv · showhide
Accurate localization of the bronchoscope within the bronchial tree is essential for clinicians to be able to reach target lesions, perform biopsies and avoid misidentification of airway segments during diagnostic and therapeutic procedures. However, existing navigation systems typically rely on patient-specific CT scans or additional external sensors, increasing cost, setup time and patient radiation exposure. This work presents BronchoTop, a real-time, RGB-only framework for topological bronchoscopy localization that eliminates the need for patient-specific data. BronchoTop estimates scope location relative to a generic airway model through four modules: lumen detection and tracking, lumen-branch label association, probabilistic scope location estimation, and switch verification. By using only standard bronchoscopy video input, BronchoTop provides practical, real-time navigational assistance to physicians. Evaluation on phantom, simulated and real data demonstrates state-of-the-art accuracy, improving existing approaches performance by over 20% on real bronchoscopy sequences. BronchoTop is the first published framework including both the localization algorithms as well as all the real data used, together with code to generate additional simulations, encouraging and facilitating further developments and benchmarking. The results highlight BronchoTop's potential to enhance procedural safety, efficiency and accessibility in clinical and robotic bronchoscopy.
I. INTRODUCTION
Bronchoscopy requires clinicians to track the scope through a complex airway tree, while misidentifying segments can lengthen procedures and cause missed lesions. BronchoTop offers RGB-only online localization in a generic airway model, alongside released data and tools for further research.
I. INTRODUCTION
- Misidentifying airway segments can lengthen procedures and lead to missed lesions, motivating navigational assistance during bronchoscopy.Clinicians otherwise mentally track the bronchoscope through the airway’s complex tree-like structure.Source: §I · p. 1
- Existing navigation methods commonly require patient-specific CT scans, external sensors, or image registration, adding radiation exposure and setup time.Topological localization in a generic, clinically used airway model is presented as sufficient spatial context for navigational guidance.Source: §I · p. 1
- BronchoTop provides online RGB-only localization in a generic bronchial tree without patient-specific CT scans or additional hardware.Its four modules are lumen detection and tracking, lumen-branch label association, scope-location probability estimation, and switch verification.Source: §I · pp. 1–2
- The work releases annotated real procedures and synthetic sequences, plus the framework and trained models for benchmarking and generating simulations.The dataset contribution is presented as addressing a data-scarce domain.Source: §I · p. 2
II. RELATED WORK
Prior bronchoscopy localization approaches use frame classification, tracked exploration trees, or anatomical models, but face limits in temporal consistency, clinical correspondence, online use, or data availability. BronchoTop proposes online RGB-image localization within a generic bronchial model.
II. RELATED WORK
- Earlier frame-wise classifiers do not model temporal procedure dynamics, and their trained models and large labeled real-image sets were not publicly available.These methods classify individual frames as locations in the bronchial tree.Source: §II · p. 2
- Tracked exploration trees capture temporal structure but lack direct correspondence to clinical anatomy, while CT registration adds patient-specific scans and setup overhead.Such relative trees can indicate branch depth but cannot distinguish specific nodes such as the Right and Left Main Bronchi.Source: §II · p. 2
- Prior generic-model localization uses CNN likelihoods and an HMM, but is offline, requires sequence endpoints, and had only phantom evaluation.The model relies on airway similarity across patients up to the fourth branch level.Source: §II · p. 2
- BronchoTop localizes online from bronchoscope RGB images in the same generic bronchial-tree model, without patient-specific CT or challenging registration.The authors describe simplified, homogeneous identification of previously detected regions of interest for follow-up procedures and report promising results on released real-procedure data.Source: §II · p. 2
III. METHOD
BronchoTop estimates the scope’s probability of occupying each node in a 15-class, six-level generic airway model using image data alone. Its pipeline combines lumen tracking and branch association with probability estimation and switch verification.
III. METHOD
- BronchoTop uses only bronchoscope image frames, without external sensors or patient data, and its clinical airway topology contains 15 classes over six levels.For each frame, it computes the probability of scope location at every tree node and determines the most likely position.Source: §III · p. 2
- The pipeline has four modules: lumen detection and tracking, lumen-branch label association, scope-location probability estimation, and switch verification.The probability module uses tracked labeled lumens and camera moving direction estimated with RAFT optical flow.Source: Figure 2 · p. 3
B. Lumen-Branch Label Association;
- Lumen-branch label association assigns tracked lumens labels with respect to the bronchial tree model.The method’s four-module description places this association between lumen tracking and scope-location probability estimation.Source: B. Lumen-Branch Label Association · p. 2, Figure 2 · p. 3
A. Lumen detection and tracking.
BronchoTop detects and tracks lumens, associates them with airway branches, and updates a probability distribution over possible scope locations. Its localization combines lumen observations with anatomical transitions and camera motion direction.
A. Lumen detection and tracking.
- Grayscale K-means clustering identifies dark lumen regions, and an IoU-based tracker maintains lumen tracks across frames.Detections match when box IoU exceeds 30%; tracks are removed after five consecutive missed frames, helping address lighting variation and liquids.Source: A. Lumen detection and tracking · p. 3
B. Lumen-Branch Label Association.
- Tracked lumens are labeled using the previous scope location and their left-to-right order, with candidate branches drawn from its children and grandchildren.The left-to-right assignment assumes the scope enters upright; the tracker captures subsequent orientation changes.Source: B. Lumen-Branch Label Association · p. 3, Figure 3 · p. 3
C. Scope Location Probability Estimation.
- A recursive Bayesian update estimates scope-location probabilities from labeled lumens, the airway tree, and camera movement direction.The transition model permits only anatomically valid child, parent, or same-node moves, depending on direction.Source: C. Scope Location Probability Estimation · pp. 3–4
- Maintaining probabilities across the generic airway tree lets the estimator handle sporadic false positives and missed detections.The paper contrasts this with deterministic bronchial-orifice tracking.Source: C. Scope Location Probability Estimation · p. 3
- The observation likelihood uses active-lumen counts to indicate branch points and whether the scope has moved into a child or parent location.A return to one active lumen after forward movement indicates passage through a branch point; an increase to two or more during backward movement indicates retraction to the parent branch point.Source: C. Scope Location Probability Estimation · p. 4
D. Switch Verification.
Switch Verification checks suspected location changes against frames associated with different active-lumen counts. It accepts, delays, or rejects a change based on frame similarity and subsequent lumen observations.
D. Switch Verification.
- For a suspected switch, the module compares the current frame with a reference frame having a different active-lumen count.For forward motion, the reference is the last frame with at least two active tracked lumens; for backward motion, it is the last frame with one tracked lumen.Source: D. Switch Verification · p. 4
- A lightweight three-layer siamese network measures frame similarity, and a validation-set threshold determines whether the images are different enough to confirm the switch.If the comparison does not confirm a change, the system waits for 10 frames before reevaluating; persistent lumen counts validate the change, while a changed count rejects it and retains the previous location.Source: D. Switch Verification · p. 4
IV. BRONCHOTOP DATASET
The BronchoTop dataset provides real and simulated bronchoscopy sequences with expert frame-level topological labels, alongside a smaller lumen-detection set. Its subsets cover distinct data conditions and tasks.
IV. BRONCHOTOP DATASET
- BronchoTop was released to address the lack of public annotated real data for bronchoscopy topological localization.It includes expert frame-level labels for real and simulated sequences.Source: §IV · p. 4
BronchoTop-Sim
- BronchoTop-Sim contains nine CT-based simulated sequences from two patients, with 700–1000 frames each and paths reaching branching level 4.The sequences are generated in Blender, and the framework for generating additional sequences is available.Source: §IV · p. 4
BronchoTop-Real
- BronchoTop-Real contains six procedure sequences from three patients, with 505–1285 frames each and paths reaching branching level 3.Motion includes forward-only and forward–backward trajectories; Real1–Real4 have doctor-assigned frame labels, while Real5–Real6 are unlabeled.Source: §IV · p. 4
BronchoTop-Det
- BronchoTop-Det provides 90 manually boxed images from Real1 and Real2, selected to vary illumination and lumen visibility.The dataset addresses lumen detection, while BronchoTop-Sim and BronchoTop-Real provide simulated and real sequences for topological localization.Source: §IV · p. 4
V. EXPERIMENTS AND RESULTS
The experiments evaluate BronchoTop on phantom, simulated, and real bronchoscopy sequences, using public benchmarks and the new BronchoTop dataset. The evaluation also compares the complete pipeline with related work and examines selected module alternatives.
V. EXPERIMENTS AND RESULTS
- Related-work baselines were implemented fully or partially, with adaptations for online topological localization on the public datasets.The authors report that original implementations could not be found.Source: §V · p. 5
- The main experiment compares complete BronchoTop with the closest related topological-localization work, while separate experiments evaluate lumen detection and scope-location probability modules.The pipeline-specific branch-association and switch-verification steps are validated in the main experiment.Source: §V · p. 5
A. Data
- The evaluation combines public and newly collected datasets, and the paper summarizes which datasets serve each task.Its reported evaluation spans phantom, simulated, and real bronchoscopy sequences.Source: §V · p. 4, A. Data · p. 5, Table II · p. 5
- For lumen detection, the evaluation uses BronchoTop-Det and CVC-DBLumen, a public set of 125 real frames with segmentation masks.Detection boxes in CVC-DBLumen are derived by grouping connected components in the masks.Source: A. Data · p. 5
- Topological localization is evaluated on BronchoTop-Sim, BronchoTop-Real, and phantom sequences Phantom0 and Phantom7.The other 14 phantom sequences are used for training; the evaluated phantom sequences and labels match those used in.Source: A. Data · p. 5
B. Lumen detection evaluation (module A).
The evaluation compares lumen-detection strategies on CVC-DBLumen and BronchoTop-Det. BronchoTop’s detector offers the best precision–recall balance on both datasets while supporting real-time inference.
B. Lumen detection evaluation (module A).
- M1 detects lumens from image gradients and pixel intensity, while M2 thresholds CycleGAN-generated depth maps to retain the darkest pixels.M1 is evaluated on CVC-DBLumen; M2 is based on.Source: B. Lumen detection evaluation (module A) · p. 5
- M3 trains a Lite R-ASPP segmentation model on simulated data using pseudo-labels derived by K-means clustering of depth maps.M2 and M3 are re-implemented; M1 results are reported from because its code and training set are unavailable.Source: B. Lumen detection evaluation (module A) · p. 5
- BronchoTop’s lumen detector achieves the best precision–recall balance on both datasets and runs in 17 ms per image, enabling real-time inference.Method M1 performs better on CVC-DBLumen, but its 57 ms inference time exceeds the pipeline’s roughly 30 ms per image.Source: B. Lumen detection evaluation (module A) · p. 5
- Sporadic false positives from the detector are filtered downstream by temporal tracking and switch verification.Source: B. Lumen detection evaluation (module A) · p. 6
C. Scope location probability estimation evaluation (module C).
The experiment holds lumen detection, tracking, and labeling constant while replacing the scope-localization step with alternative strategies. BronchoTop’s proposed probability-estimation module significantly outperforms the evaluated alternatives, especially on more challenging datasets.
C. Scope location probability estimation evaluation (module C).
- Baseline B1 moves the estimated scope location down the tree when fewer lumens are tracked in the current frame than in the previous frame.Source: C. Scope location probability… · p. 6
- Baseline B2 uses camera movement direction and unmatched tracked lumens to determine whether the scope stays in place or moves forward or backward.Source: C. Scope location probability… · p. 6
- BronchoTop’s scope-location probability estimation significantly outperforms the evaluated alternatives, with benefits particularly marked as datasets become more challenging.The comparison varies module C while sharing steps A and B across approaches.Source: C. Scope location probability… · p. 6
- Baseline B3 estimates location from tracked-lumen labels and the lumens theoretically visible at each bronchial-tree node.Source: C. Scope location probability… · p. 6
D. Topological Localization performance.
BronchoTop and its variant without Switch Verification both outperform baseline B4 on the simulated and real datasets, while B4 performs better on Phantom. Switch Verification improves BronchoTop’s accuracy and can reject erroneous location changes, though tracking-based errors can persist and require clinician correction.
D. Topological Localization performance.
- Both BronchoTop variants significantly outperform B4 on BronchoTop-Sim and BronchoTop-Real, while B4 achieves higher accuracy on Phantom.B4’s CNN requires supervised training on each target dataset; Phantom labels are available, but the trained model fails to generalize to real data.Source: D. Topological Localization performance · p. 6, Table IV · p. 7
- Adding Switch Verification increases BronchoTop accuracy by 21% and produces comparable results across sequences.The comparison is between the full A+B+C+D pipeline and the A+B+C ablation.Source: D. Topological Localization performance · p. 6, Table IV · p. 7
- In Sim2 and Real2, Switch Verification discards erroneous proposed location changes and maintains the ground-truth location LMB.Without module D, the estimated location changes from LMB to LLB6 in Sim2 and to TRA in Real2.Source: D. Topological Localization performance · p. 7, Figure 6 · p. 7
- A tracking-based location error can propagate to later predictions even when a subsequent transition is detected correctly, requiring manual clinician correction.In Real2, the later estimate is LMB rather than LLB6 after the earlier error.Source: D. Topological Localization performance · p. 7
VI. CONCLUSIONS
BronchoTop combines RGB-only, real-time topological localization with a generic bronchial-tree model, and the authors release the first public benchmark for this task. Future work targets deeper airway levels, robustness to erroneous location changes, and clinical or robotic integration.
VI. CONCLUSIONS
- BronchoTop is presented as a real-time, RGB-only framework for topological bronchoscopy localization using a generic bronchial-tree model and a four-stage pipeline.The pipeline includes lumen detection and tracking, lumen-branch association, probabilistic scope localization, and switch verification.Source: §VI · p. 7
- The authors report state-of-the-art localization accuracy and real-time performance on phantom, simulated, and real sequences, and release the first public topological-localization benchmark.Source: §VI · p. 7
- Future work will extend localization to deeper, more-complex airway levels, improve robustness to incorrect location changes and their effects on later estimates, and explore clinical and robotic integration.Source: §VI · p. 7