Source-linked AI summary
Investigating the trackability of landmarks in NBI and WLI
Jonas Below, Piermarco Gerli, Anna Jötten, Dennis Eggert, Axelle Felicio-Briegel, Veronika Volgger, Christian Betz, Alexander Schlaefer
TL;DR
Reliable landmark tracking is important for computer-assisted laryngeal interventions, but whether NBI improves automated tracking over WLI remains unclear. The study compares forward–backward consistency across 61 videos, two anatomical regions, and three trackers; NBI has lower error across trackers and regions, with significant reductions for vocal folds on two trackers.
Problem
Whether NBI improves automated landmark tracking over WLI, and whether effects vary by region or tracker, remains unclear despite NBI’s clinical use for enhancing vascular structures.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
Tracking remained feasible in both modalities with low failure rates for CoTracker3 and TrackOnR, supporting modality-aware tracking systems for NBI and WLI examinations.
Source: §4 · p. 4Forward–backward consistency does not establish anatomical correctness, and limited dataset size may reduce statistical power, especially for MPVP.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Reliable visual tracking of anatomical landmarks in endoscopic video is a key prerequisite for computer-assisted interventions in laryngeal surgery. While Narrow Band Imaging (NBI) is clinically established for enhancing vascular contrast, its effect on automated tracking relative to conventional White Light Imaging (WLI) remains unexplored. We present a quantitative comparison of landmark trackability in NBI and WLI using a forward-backward error framework applied to 61 endoscopic videos from 39 patients. Two regions of interest are evaluated: the vocal folds (VF) and mucosa with a prominent vascular pattern (MPVP). We compare a classical feature-based tracker (CSRT) against learning-based point trackers (CoTracker3-Online and TrackOnR) across both modalities. NBI consistently reduced the forward-backward error in both regions, reaching statistical significance for the vocal folds, while failure rates stayed low. These findings indicate that NBI can improve tracking robustness in laryngeal endoscopy, with implications for motion analysis, navigation, and robotic assistance.
1 Introduction
The study asks whether NBI improves landmark tracking over WLI and whether the effect depends on the tracked region or algorithm. This comparison matters because reliable landmark tracking supports computer-assisted interventions, motion compensation, and navigation.
1 Introduction
- Reliable tracking is challenging in laryngeal endoscopy because of camera motion, tissue deformation, illumination changes, and partial occlusions.Tracking vocal folds could support automated screening and robotic assistance.Source: §1 · p. 1
- NBI’s enhanced vascular texture could improve feature stability, but its altered appearance and stronger local contrast could also degrade some methods, leaving its overall effect uncertain.The work compares modalities across two ROIs and three trackers.Source: §1 · p. 1
- The study quantitatively compares NBI and WLI trackability across 61 videos from 39 patients, using two ROIs, two point trackers, and a baseline box tracker.Track consistency is evaluated with forward–backward error.Source: §1 · p. 1
2 Methods
The methods prepare paired modality-specific video segments, define vocal-fold and vascular-pattern mucosal ROIs, and estimate local tracking consistency by propagating candidates forward and backward through each subvideo.
2.1 Data processing
- Across 61 videos from 39 patients, modality-specific segments longer than 2 seconds and suitable for tracking were retained, then divided into overlapping 2-, 4-, and 6-second subvideos.Segments with extensive blur or rapid motion were excluded; a new set of subvideos began every 4 seconds. Figure 1 illustrates about 100 tracking candidates placed within the manually annotated ROIs.Source: §2.1 · p. 1, Figure 1 · p. 2
2.2 Experimental setup
- Each first frame was annotated for vocal-fold and prominent-vascular-pattern mucosal ROIs, which were tiled with distributed square boxes tracked as independent candidates.Point trackers followed each box centroid, while the box tracker followed its full extent; candidate counts used tolerance intervals, and only ROIs fully visible in the final frame were evaluated.Source: §2.2 · p. 2
2.3 Forward–Backward Tracking Consistency and Trackability Estimation
- Forward–backward error measures centroid discrepancy across reversed trajectories; spatially sampled mean errors form trackability maps, while zero-IoU recovery marks failed candidates.Mean errors exclude the final frame, where error is zero by definition, and are aggregated first across candidates and then at video level. The framework estimates local consistency without dense frame-wise annotations and applies to point and box trackers.Source: §2.3 · pp. 2–3
3 Results
Across trackers and ROIs, NBI yielded lower mFBE than WLI; VF reductions were significant for CoTracker3 and CSRT, while qualitative heatmaps show more contiguous low-error regions.
3 Results
- NBI yielded lower mFBE than WLI across all trackers and both ROIs; VF reductions were significant for CoTracker3 and CSRT (p = 0.013 for both).For MPVP, the reduction was observed for all trackers but was not statistically significant.Source: §3 · p. 3, Table 1 · p. 3
- For VF, CSRT failure rate was 61.0% ± 24.0% in NBI versus 78.1% ± 15.7% in WLI (p = 0.001); MPVP rates were nearly identical across modalities.CoTracker3 and TrackOnR had low failure rates in both modalities, with no significant modality differences.Source: §3 · p. 3, Table 2 · p. 3
- In selected heatmaps, NBI showed more contiguous low-error regions, whereas WLI appeared more heterogeneous.Figure 2 displays CoTracker3 trackability heatmaps for VF and MPVP, computed from 4-second subvideos with colors ranging from minimum to maximum mFBEtc.Source: §3 · p. 3, Figure 2 · p. 3
4 Discussion
Both NBI and WLI support landmark tracking, but NBI generally provides lower forward–backward error, with the clearest benefit for vocal-fold tracking. The discussion also notes that the consistency-based evaluation does not establish anatomical correctness.
4 Discussion
- NBI produced lower mFBE across trackers and ROI categories; vocal-fold reductions were significant for CoTracker3 and CSRT, but not TrackOnR.Failure rates stayed low for CoTracker3 and TrackOnR in both modalities; CSRT had substantially higher failure rates overall but improved significantly in NBI for VF.Source: §4 · pp. 3–4
- NBI heatmaps showed larger contiguous regions of low tracking error and more homogeneous consistency, while robust point trackers had low failure rates in both modalities.Landmark tracking could remain applicable when clinicians switch modalities, although image-appearance changes may require reinitialization.Source: §4 · p. 4
- The NBI benefit was clearest for VF; MPVP had greater variability and no statistically significant modality differences.The discussion relates MPVP variability to heterogeneous appearance and susceptibility to overexposure, reflections, motion blur, and local tissue changes.Source: §4 · p. 4
- The forward–backward analysis measures temporal self-consistency, not necessarily anatomical correctness, and the limited dataset may reduce statistical power, especially for MPVP.NBI and WLI were not recorded simultaneously with the same endoscopic view, so future studies should use predefined protocols to obtain more comparable segments.Source: §4 · p. 4
- Overall, the findings suggest NBI can improve natural-landmark tracking robustness in laryngeal endoscopy, particularly for the vocal folds.Tracking remains feasible in both modalities when robust trackers are used, supporting development of modality-aware tracking systems.Source: §4 · p. 4