{"id":"3847cf7e-33c3-4424-9d83-6c4f63431e4e","arxiv_id":"2505.05136","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Bronchoscopy-only automatic stenosis severity estimation is demonstrated with lumen tracking, keyframe selection, and single-frame illumination-decline 3D reconstruction, validated against CT and expert estimates on a new 16-video dataset.","lead":"This paper presents a two-step computer vision pipeline that measures subglottic stenosis severity from bronchoscopy video alone, using the dark airway lumen and a single-frame 3D reconstruction. It also releases a small real-procedure dataset with CT and expert references, aiming to reduce diagnostic subjectivity and avoid repeated CT radiation during monitoring.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stenosis area is never validated against manual lumen annotations; the intensity-threshold mask could be measuring a shadow or tissue, and because it feeds Eqs. (1)-(3), the reported accuracy may not reflect anatomical stenosis severity.","rationale":"The reader's weakest assumption includes the claim that the darkest image region is the stenosed lumen rather than shadow or tissue, and my concern is a more concrete version of that: the paper provides no direct evidence that the binary threshold mask actually outlines the stenotic opening. The 'Correct Keyframes' metric only checks the temporal location of the selected frame, not the spatial accuracy of the segmentation. The 'Manual' baseline still uses the same threshold segmentation, so it only removes keyframe-selection error, not segmentation error. Because Eqs. (1)-(3) take the segmented region as the stenosis area, any systematic segmentation error propagates directly into the headline numerical results. This is the most load-bearing link between the image and the anatomical quantity the paper claims to measure. The training/evaluation overlap of LightDepth on sequence 1 and the Myer-Cotton reference-location mismatch are also real concerns, and the paper honestly acknowledges the latter, but they are secondary: even a perfectly trained depth model and a perfectly calibrated reference would not rescue a method that segments the wrong region. I therefore recommend keeping the reader's CONDITIONAL verdict, with the condition that segmentation-level validation be provided. If the proposed manual-mask comparison shows large disagreement, the verdict should be revised downward.","tokens_in":8374,"tokens_out":6436,"duration_ms":75267,"concrete_test":"Manually annotate the visible stenotic lumen (the residual opening) in the selected keyframe of each of the 16 sequences, ideally with two independent annotators. Compute the Dice/IoU between the manual lumen mask and the intensity-threshold mask used by the pipeline. Then recompute PSA and PSD using the manual masks in place of the threshold masks, and compare with the reported values. If mean IoU is below approximately 0.8, or if the manual-mask PSA differs from the reported PSA by more than 5 percentage points, the intensity-based segmentation is not a reliable proxy for the anatomical lumen and the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central measurement chain depends on the binary intensity-threshold segmentation (grayscale < 50, Section 3.2) to identify the stenosed lumen in the keyframe. The paper validates keyframe selection (Correct Keyframes) but never validates the segmentation itself against manual lumen delineations. The only baseline that shares this segmentation is the 'Manual' keyframe baseline, so it cannot distinguish a correct mask from a consistently biased one. If the darkest region is a shadow, a mucosal fold, or the dark lumen beyond a different structure, then Astenosis in Eq. (1), the diameter fit for PSD in Eq. (3), and therefore both PSA and PSD are systematically wrong. The reported agreement with CT and expert references is averaged over a small, single-center cohort and does not establish that the mask corresponds to the residual stenotic opening. This is the load-bearing link between image pixels and anatomy; without segmentation-level validation, the pipeline may be measuring a photometric artifact rather than the stenosis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a fully automated pipeline for estimating subglottic stenosis severity from bronchoscopy video. The method segments the darkest image region (intensity < 50), tracks it with an IoU tracker to detect when the bronchoscope passes the vocal cords, selects that frame as the keyframe, and applies LightDepth single-view depth reconstruction to measure the stenotic area and diameter ratios against a maximum reference area between the vocal cords and the stenosis. The authors contribute a new dataset of 16 real bronchoscopy videos from 11 patients with CT-based and expert-based references, and report a PSA MAE of 4.84% and a PSD MAE of 6.70%, with perfect keyframe selection on their method. They claim this is the first automated, robust SGS measurement from bronchoscopy images.","tokens_in":8621,"tokens_out":3970,"duration_ms":37275,"significance":"If the results hold, the work would provide a practical assistive tool that reduces subjectivity and CT radiation exposure, and the released benchmark would be a useful community resource. The use of external CT and expert references, repeatability analysis across repeated procedures, and comparison against keyframe-selection baselines are strengths. However, the current evidence is limited by the small cohort, the training/evaluation overlap for sequence 1, and the absence of validation of the intensity-mask segmentation that underlies the measurement; these gaps currently weaken the central claim of robustness. The paper also ships a public dataset, which is a tangible contribution to a data-scarce area.","major_comments":[{"comment":"LightDepth is trained on 2800 frames from sequence 1 and that same sequence is retained in the stenosis evaluation (Section 4.1, 'This sequence is used for LightDepth training, but is kept to evaluate...'). Sequence 1 also has CT ground truth. Because the depth model is evaluated on training data, the reported MAE for that sequence and the aggregate MAE may be optimistically biased, and the claim of consistency with CT is weakened. Please report results separately for sequence 1 versus held-out sequences, or retrain LightDepth on a split that excludes sequence 1 and re-run the evaluation.","section":"4.1"},{"comment":"The stenosed area Astenosis is defined by the binary intensity threshold (grayscale < 50) of Section 3.2, and this mask propagates directly into PSA and PSD via Eqs. (1)-(3). The paper validates keyframe selection (Correct Keyframes) but never validates the lumen mask against manual delineations or any independent segmentation. If the darkest region is a shadow or a mucosal fold, the reported agreement with CT and expert references could be coincidental. Please provide a segmentation-level validation (e.g., Dice against expert lumen annotations on a subset of keyframes) and a sensitivity analysis of PSA/PSD to the intensity threshold.","section":"3.3, Eqs. (1)-(3)"},{"comment":"The MAE values are reported without confidence intervals or per-sequence error distributions, and the paired t-tests (p=0.6039, p=0.2766) are underpowered with N=5 CT and N=12 expert comparisons; 'no statistically significant difference' is not evidence of agreement. Please report Bland-Altman limits of agreement and bootstrap confidence intervals for the MAE, and discuss the effect size relative to clinically meaningful differences.","section":"4.2, Table 2"},{"comment":"The healthy reference area Ak is measured between the vocal cords and the stenosis, whereas the CT ground truth is defined at the trachea's widest point below the stenosis and expert estimates follow the Myer-Cotton convention. This is acknowledged as a limitation, but it means the reported agreement in Table 2 is between two different anatomical quantities; the 'slight underestimation' could be entirely explained by this location mismatch. Please quantify the expected bias by computing both reference locations on the CT volumes for the five CT cases, or explicitly report the Myer-Cotton-equivalent index.","section":"5, Fig. 4"}],"minor_comments":[{"comment":"The name 'LigthDepth' in the Introduction (Section 1) is a typo; it should be 'LightDepth' as used in the rest of the manuscript.","section":"1"},{"comment":"The table mixes comma and period as decimal separators (e.g., '37,5' vs '4.84'); please use a consistent decimal separator throughout.","section":"Table 2"},{"comment":"The notation in Eq. (1) and the surrounding text, particularly 'X∋D', is confusing; please clarify what D denotes and how the set X3Dstenosis is obtained from the plane intersection.","section":"3.3"},{"comment":"The sentence 'When the tracker stops...' contains a stray closing parenthesis and could be rephrased for clarity: remove the extra ')' and state directly that tracking loss indicates the camera has passed the vocal cords.","section":"3.2"},{"comment":"The claim 'No public methods or datasets are currently available' is stronger than warranted given the semi-automatic methods cited in [11,12]; consider specifying 'no publicly available automated method or benchmark'.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The dataset release is valuable and the clinical motivation is strong. The main risk is that the current evaluation protocol (training/evaluation overlap, unvalidated segmentation, small N) overstates the robustness claim; I recommend requiring the fixes in the major comments before acceptance. Please also consider asking the authors to clarify the relationship of their PSA/PSD to the clinically used Myer-Cotton grading, since the current abstract may be seen as claiming equivalence that the method does not yet establish."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: the paper makes a real contribution: first public SGS bronchoscopy dataset and first automated bronchoscopy-only pipeline for subglottic stenosis severity. The illumination-decline idea is coherent and well-exploited in two places: keyframe selection and single-frame depth. Results are promising, and the repeatability numbers across same-patient procedures are encouraging.\n\nI agree with the stress-test note: the lumen segmentation is never validated against manual delineations. The intensity threshold mask is load-bearing, and if it picks up shadow or tissue instead of the residual lumen, both PSA and PSD are wrong. The Manual baseline only fixes the keyframe, so it does not disentangle mask error. Add a small manual annotation set and report Intersection-over-Union or Dice. That is the revision that matters most.\n\nThe train/eval overlap is also real: LightDepth is trained on sequence 1 and sequence 1 is still scored. The authors justify it by saying errors are similar, but a held-out sequence is needed. This is fixable; as written it weakens the accuracy claims.\n\nThe CT comparison is imperfect because the reference area is measured in the trachea while the pipeline measures just beyond the vocal cords. The authors admit the Myer-Cotton mismatch, but the same issue affects the CT ground-truth comparison and could bias the MAE.\n\nThe cohort is small and single-center, but for a first clinical dataset that is not a reason to reject. The paper is honest about limitations and does not overclaim, which matters.\n\nThis deserves peer review. A careful referee should push for segmentation validation and a held-out depth model. If those are addressed, the dataset alone would be useful to the community. I would cite the benchmark if I worked on endoscopic measurement.","headline":"A genuinely useful first benchmark and a clever illumination-decline pipeline for subglottic stenosis, but the evaluation has a train/eval overlap and an unvalidated lumen segmentation that need scrutiny.","tokens_in":654,"tokens_out":1145,"would_cite":true,"duration_ms":42692,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a fully automated bronchoscopy-only pipeline that measures subglottic stenosis severity from a single keyframe, using illumination decline to track the lumen and reconstruct the airway in 3D.","keywords":["bronchoscopy","subglottic stenosis","stenosis severity estimation","illumination decline","single-view depth estimation","airway lumen segmentation","IoU tracking","3D reconstruction"],"falsifier":"Run the pipeline on the healthy patients in the dataset (G, H, I), who have no stenosis: a correct method should produce near-zero stenosis indices, so a large reported $PSA$ or $PSD$ on those sequences would show that the darkest-region assumption is measuring shadow or tissue rather than lumen. A second check would be a phantom airway with known obstruction diameters placed at known distances from the camera; systematic deviation of estimated $PSA/PSD$ from the known geometry would falsify the claim that the single-frame 3D model preserves airway proportions.","tokens_in":8203,"feed_emoji":"🫁","tokens_out":10063,"duration_ms":90882,"temperature":0.7,"pith_summary":"This paper aims to turn a routine bronchoscopy video into an objective measurement of subglottic stenosis severity, replacing the subjective visual inspection that prior work has repeatedly shown to be unreliable. The proposed two-step pipeline follows the darkest region of the image—the airway lumen—until the scope passes the vocal cords, then builds a 3D model of the airway from a single keyframe and computes the percentage of obstruction. The authors claim this is the first automated method for this measurement that needs no traversal of the stenosed region, no CT scan, and no additional hardware. They support the claim with a new public dataset of real procedures, reporting stenosis estimates consistent with CT ground truth and expert judgment, and repeatable across repeated procedures on the same patient.","feed_headline":"Automated pipeline measures subglottic stenosis from bronchoscopy","feed_subtitle":"Using the darkest-pixel cue and one 3D frame, estimates match CT and expert grades and repeat across procedures.","key_machinery":"The load-bearing mechanism is the photometric inverse-square law: because the bronchoscope's light source and camera are co-located, image intensity $I$ falls with distance $d$ roughly as $I \\approx (1/d^2)^{1/\\gamma}$, so the darkest area marks the deepest visible point, the lumen. That cue drives both pipeline stages: an IoU tracker follows the darkest segment until its shape changes abruptly at the vocal cords, and a single-view self-supervised depth model trained on the same illumination-decline signal produces the 3D airway points $X_{3D}$ from the selected keyframe. Stenosed points are isolated as the darkest 3D set $X_{3D}^{\\text{stenosis}} = \\{X \\in X_{3D} \\mid X \\in P\\}$, where $P$ is the plane fitted to them, and the healthy reference area is the maximum over planes $Q_k$ perpendicular to the camera: $k = \\arg\\max_k A_k$. The ratio of these two areas or of fitted circle diameters yields the stenosis index. Keyframe selection and 3D geometry are therefore two uses of the same physical signal.","core_discovery":"The paper's central discovery, stated on its own terms, is that illumination decline in endoscopy carries enough geometric information to automate subglottic stenosis severity estimation from a single bronchoscopy frame. By segmenting and tracking the darkest image patch with an IoU tracker, the pipeline detects when the camera has just passed the vocal cords; the frame where tracking is lost becomes the keyframe. A single-view self-supervised depth model reconstructs the airway in 3D, the stenosed region is delineated as the darkest set of 3D points and intersected with a plane, and the healthy reference is taken as the largest plane section between the vocal cords and the stenosis. The stenosis index is reported both as an area ratio, $PSA = (1 - A_{\\text{stenosis}}/A_k)\\times 100$, and a diameter ratio, $PSD = (1 - d_{\\text{stenosis}}/d_k)\\times 100$. On the released dataset the pipeline selects correct keyframes for all 16 sequences, achieves mean absolute errors of 4.84% for $PSA$ and 6.70% for $PSD$, and shows same-patient differences of 1.79% for $PSA$ and 5.07% for $PSD$, compared with 7.5% average variation between expert estimates; the authors also note a slight underestimation of the stenosis index and a healthy-reference location that differs from the clinical Myer-Cotton grading.","pith_inferences":["The paper leaves implicit that the same illumination-decline cue could extend to other endoluminal strictures, such as tracheal or esophageal narrowing, wherever a co-located light and camera produce the same intensity-distance relation.","The paper does not draw the consequence that combining the scale-invariant ratio with the known bronchoscope diameter would yield absolute lumen diameters and a mapping onto the Myer-Cotton grading scale.","A testable extension the paper does not pursue is to use the albedo and normal maps the depth model already outputs to distinguish a true lumen from shadowed tissue, which would harden the darkest-equals-lumen premise.","The same-patient repeatability suggests a longitudinal monitoring tool, but the paper only shows two sequences per patient, so sensitivity to genuine stenosis progression over time is not established."],"forward_implications":["Stenosis severity can be measured during the same bronchoscopy pass used for exploration, without threading the scope through the narrowing, which shortens the procedure and removes the need for a CT scan for monitoring.","An automated, reproducible number replaces the subjective visual grade; the paper cites prior work in which physicians misclassified severity in 53% of cases, mostly by underestimation.","Because only a single keyframe is needed, the measurement can run in parallel with the clinician's assessment, taking about 7 seconds per case on a current GPU.","The healthy reference is measured between the vocal cords and the stenosis rather than in the trachea beyond it, so a direct mapping to the standard Myer-Cotton grade still requires calibration with expert-supervised data.","The released benchmark gives other research groups the first public data on which to train and compare automated stenosis estimators."],"supporting_citations":[{"why":"It documents that bronchoscopists lack a standardized tool for measuring central airway obstruction, the clinical gap this work targets.","marker":"[1]"},{"why":"It quantifies how often subjective visual grading misclassifies airway narrowing (53% of cases), motivating an objective automated estimate.","marker":"[2]"},{"why":"It supplies the single-view self-supervised depth model that turns the selected keyframe into the 3D airway used for stenosis measurement.","marker":"[3]"},{"why":"It provides the spotlight illumination model that grounds the intensity-decline assumption used for both lumen tracking and depth reconstruction.","marker":"[18]"},{"why":"It provides the IoU-based tracker that follows the darkest segment until the vocal cords are crossed, triggering keyframe selection.","marker":"[19]"},{"why":"It defines the clinical diameter-based grading for subglottic stenosis that the paper adapts into its PSD ratio and healthy-reference logic.","marker":"[23]"}],"fun_headline_variants":["Single bronchoscopy frame automates stenosis severity grading","Illumination-based AI measures airway narrowing from bronchoscopy video","No CT needed: single bronchoscopy frame grades stenosis severity","Automated stenosis estimation via illumination decline in endoscopy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that just after the tracked darkest patch changes shape the camera has passed through the vocal cords, and that in that keyframe the darkest image region is the stenosed airway opening rather than shadow or tissue, with the 3D shape computed from that single frame accurate enough that the area and diameter ratios match the real airway.","fun_headline_variants_meta":{"raw":{"variants":["Single bronchoscopy frame automates stenosis severity grading","Illumination-based AI measures airway narrowing from bronchoscopy video","No CT needed: single bronchoscopy frame grades stenosis severity","Automated stenosis estimation via illumination decline in endoscopy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000444,"raw_usage":{"total_tokens":2346,"prompt_tokens":1146,"completion_tokens":1200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":762,"completion_tokens_details":{"reasoning_tokens":1135}},"tokens_in":762,"tokens_out":1200,"duration_ms":9256,"temperature":1.0,"reasoning_tokens":1135,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:11:44.280899+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on the healthy patients in the dataset (G, H, I), who have no stenosis: a correct method should produce near-zero stenosis indices, so a large reported $PSA$ or $PSD$ on those sequences would show that the darkest-region assumption is measuring shadow or tissue rather than lumen. A second check would be a phantom airway with known obstruction diameters placed at known distances from the camera; systematic deviation of estimated $PSA/PSD$ from the known geometry would falsify the claim that the single-frame 3D model preserves airway proportions.","supporting_citations":[{"cited_title":"what do bronchoscopists do? Annals of the American Thoracic Society 12(1), 85–90 (2015) 10","cited_arxiv_id":null,"evidence_quote":"It documents that bronchoscopists lack a standardized tool for measuring central airway obstruction, the clinical gap this work targets."},{"cited_title":"Interactive cardiovascular and thoracic surgery 16(5), 655–660 (2013)","cited_arxiv_id":null,"evidence_quote":"It quantifies how often subjective visual grading misclassifies airway narrowing (53% of cases), motivating an objective automated estimate."},{"cited_title":"In: IEEE/CVF Int","cited_arxiv_id":null,"evidence_quote":"It supplies the single-view self-supervised depth model that turns the selected keyframe into the 3D airway used for stenosis measurement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the spotlight illumination model that grounds the intensity-decline assumption used for both lumen tracking and depth reconstruction."},{"cited_title":"GitHub (2020)","cited_arxiv_id":null,"evidence_quote":"It provides the IoU-based tracker that follows the darkest segment until the vocal cords are crossed, triggering keyframe selection."},{"cited_title":"Annals of Otology, Rhinology & Laryngology 103(4), 319–323 (1994)","cited_arxiv_id":null,"evidence_quote":"It defines the clinical diameter-based grading for subglottic stenosis that the paper adapts into its PSD ratio and healthy-reference logic."}],"review_version":1}