{"id":"f9b3b979-b6c6-4169-96e5-d0b461ed5af7","arxiv_id":"2504.12215","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-stage coarse-to-fine lung tumor segmentation pipeline with anatomy-aware component filtering improves Dice and Hausdorff metrics over a single-stage baseline on private and public CT datasets.","lead":"This paper builds a two-stage system that first finds tumor candidates in full CT scans, filters them using lung anatomy rules, and then finely segments each remaining region. On a private Orlando dataset, it raises Swin UNETR Dice from 0.469 to 0.645, though the uncertainty-aware loss did not improve Dice.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates anatomy-aware post-processing from ROI refinement; the Table I gains may come from cropping alone, so the paper's central claim is not yet established.","rationale":"The reader's weakest assumption focuses on whether the Section III-F filters discard true tumors (a recall/safety concern). That is a legitimate external-validity issue, but it is not the most load-bearing problem for the paper's central claim. The more fundamental gap is internal validity: the experimental design never isolates the post-processing module. The headline result compares the full system to a no-ROI baseline, so even if every filter is perfectly safe, the reported gains could be due solely to cropping and second-stage refinement. The paper's Table III correlation is suggestive but not causal, as component count strongly correlates with tumor size and difficulty. The authors have released code, so the missing ablation is readily testable. I keep the verdict CONDITIONAL (UNCHANGED from the reader) because the overall two-stage result is still plausible and the requested ablation would either strengthen or refute the attribution. I disagree with the reader's weakest_assumption because it identifies a secondary risk (filter false negatives) rather than the primary evidential gap (no within-pipeline ablation of the filter).","tokens_in":8714,"tokens_out":5339,"duration_ms":57866,"concrete_test":"Re-run the fully automated pipeline on the Orlando Health dataset with the Section III-F filters disabled: extract the Top-1 connected component from the raw coarse prediction without applying the lung-overlap (>80%), surface-distance (≤5 voxels), or size (>50 voxels) criteria, then feed that component's tight bounding box to the same ROI-0margin second-stage model. Compare Dice and HD95 to Table I. If the no-filter variant matches 0.6447 Dice / 8.89 mm HD95, the anatomy-aware post-processing is not responsible for the gain; if it drops substantially, the filters are validated as the causal component.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that anatomy-aware component filtering plus ROI refinement produces the large Dice/HD95 improvements in Table I (0.469 → 0.6447 Dice; 187.97 → 8.89 mm HD95). However, the only reported comparison is the full two-stage pipeline against a single-stage full-CT baseline. There is no ablation that removes the Section III-F post-processing filters (lung-overlap threshold, surface-distance rule, size cutoff) while keeping the second-stage ROI refinement. Without such a condition, the improvement could be caused entirely by ROI cropping and high-resolution refinement; the anatomy-aware filtering may be neutral or even harmful, and the paper's correlational evidence (Table III, component reduction vs. gain) is confounded because cases with fewer components are likely easier tumors. The paper itself acknowledges in Section V that the NSCLC dataset was 'more sensitive to ROI and component thresholds' and that there are 'potential concerns about the generalizability of heuristic post-processing,' which underscores that the effect of these heuristics is not isolated. Consequently, the load-bearing assertion that anatomy-aware post-processing contributes to the headline gains is unsupported by the current experiments.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a two-stage, fully automated tumor segmentation framework for thoracic CT. A first-stage full-volume model produces a coarse prediction; connected components are then filtered with anatomy-aware heuristics (lung-overlap threshold, distance-to-lung-surface rule, minimum component size, and optional top-K selection) to form ROIs. A second-stage ROI model, optionally trained with an uncertainty-aware loss, produces the final segmentation. The authors report that on a private Orlando dataset the fully automated Top-1, ROI-0margin pipeline improves Swin UNETR from Dice 0.469/HD95 187.97 mm to Dice 0.6447/HD95 8.89 mm (Table I), with smaller gains on the public NSCLC-Radiomics dataset (Table VI). They also report a negative correlation between remaining component count and performance (Table III) and state that the uncertainty-aware loss improves spatial calibration.","tokens_in":8903,"tokens_out":3480,"duration_ms":35898,"significance":"If the headline results hold, the framework offers a practical, modular, post-hoc refinement strategy that can be grafted onto existing segmentation models, with publicly released code and a reproducible experimental configuration on a public dataset. The large reported gain on the private dataset is clinically meaningful and the anatomical filtering rules are transparent. However, the current evidence is not yet sufficient: the improvement is not isolated from ROI cropping, no error bars or patient counts support the quantitative claims, and the uncertainty component is not shown to deliver the claimed calibration benefit. With additional ablations and statistical reporting, this could become a useful contribution to cascaded medical-segmentation pipelines.","major_comments":[{"comment":"The central claim that anatomy-aware post-processing contributes to the headline gains is not supported because no ablation isolates the Section III-F filtering from the second-stage ROI refinement. The comparison in Table I contrasts the full pipeline (coarse model + III-F filters + fine ROI model) with a full-CT baseline. A condition that keeps the ROI model but removes the lung-overlap, surface-distance, and size filters (e.g., selecting the largest component of the raw coarse prediction) is necessary to attribute the improvement to anatomy-aware filtering rather than to cropping and high-resolution refinement alone. Without such a condition, the Table I gains could be produced entirely by the second stage.","section":"Section IV, Table I; Section III-F"},{"comment":"No error bars, confidence intervals, or patient counts are reported for any quantitative result, and the size of the private Orlando dataset is not stated anywhere in the manuscript. The headline Dice change from 0.469 to 0.6447 therefore cannot be distinguished from run-to-run variance, especially given the much smaller improvement on NSCLC in Table VI (SwinUNETR 0.4731 to 0.4860). The authors should report the number of test cases, per-case distributions, and paired significance tests (or per-patient bootstrap CIs) for the main comparisons.","section":"Section IV, Tables I-VII"},{"comment":"The claim that the uncertainty-aware loss improves spatial calibration is unsupported. In Table V the uncertainty-aware loss gives slightly worse Dice (0.6321 vs 0.6447) and HD95 (8.92 vs 8.89) than the standard loss, yet the text states it 'improved spatial calibration and boundary interpretability.' No calibration metric (e.g., expected calibration error, reliability diagram, Brier score) is reported anywhere. Either remove this claim or provide a quantitative calibration evaluation on held-out data.","section":"Section III-D, Table V"},{"comment":"The correlation evidence in Table III is confounded and is overinterpreted. The statement that 'these findings confirm that anatomical and uncertainty-informed postprocessing ... enhances overall accuracy' does not follow from a correlation between component count and performance, because cases with fewer connected components are plausibly easier tumors (small, isolated, well-defined) regardless of post-processing. To support the causal claim, the authors should compare per-case final Dice with and without the post-processing filters, or otherwise show that the filters specifically remove components that would lower the final segmentation quality.","section":"Section IV, Table III"}],"minor_comments":[{"comment":"The phrase 'followed bb conclusion' appears in the organization paragraph; this is a typo for 'followed by the conclusion.'","section":"Section V"},{"comment":"The adaptive loss equation is not numbered, and U(x) is described only as 'the variance of multiple forward passes'; the number of MC dropout samples and the variance computation (per-voxel across samples?) should be specified.","section":"Section III-D"},{"comment":"The row label 'S4 ROI Model' is unexplained; the architecture and training details of this model should be defined in the caption or text.","section":"Table V"},{"comment":"The notation 'K=1' in the Orlando Health rows is not defined; the caption should explain that K denotes the number of retained components after post-processing.","section":"Table VI"},{"comment":"The text says the comparison includes 'the LungMask library [18], auto-generated masks, and human-annotated ground truth,' but Table IV appears to list only two rows ('Original Lung Mask' and 'New Lung Mask'). Please align the table with the described comparisons.","section":"Section IV, Table IV"},{"comment":"The sensitivity analysis reported in the text covers component size thresholds from 50 to 150 voxels, but the claim in Section V that 'changes in lung overlap and voxel size thresholds had minimal impact' is not fully supported; no lung-overlap threshold sweep is shown. Please report the overlap-threshold sensitivity or qualify the claim.","section":"Section IV and V"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a conference paper with a promising practical idea, but the experimental design does not yet establish the central attribution claim. The authors should be encouraged to add the missing ablation and statistical detail; the paper is within the scope of the venue but needs a substantive revision to become reliable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Readable, honest, and the pipeline is a reasonable combination of known pieces: coarse full-CT segmentation, lung-mask-informed filtering (overlap, surface distance, size), top-k component selection, then a second-stage ROI model, with MC-dropout uncertainty reweighting in the loss. The specific recipe is new as a whole, and they provide code and evaluate on two datasets with three backbones. I give them credit for reporting that the uncertainty-aware loss did not improve Dice (Table V) and for a sensitivity check on the size threshold.\n\nThe soft spot is the one the stress-test flags, and it is real. Table I compares the single-stage full-CT baseline to the complete two-stage pipeline. There is no condition that keeps the ROI refinement while removing the Section III-F filters, or vice versa. So the improvement from 0.469 to 0.6447 Dice could come mostly from the cropping and higher-resolution refinement; the anatomy-aware filtering might be neutral or even harmful. Table III's correlation between component reduction and performance is confounded—cases with fewer components are probably easier tumors. The paper also lacks patient counts, error bars, and split details, and it claims better 'spatial calibration' from the uncertainty-aware loss without reporting any calibration metric. In fact, Table V shows Dice drops slightly with that loss. The filtering thresholds are partly hand-set and only partially sensitivity-tested; the 'All Valid' condition in Table II badly degrades HD95, which makes me wonder how much of the benefit is top-k selection alone.\n\nThe weakest assumption the reader flags is legitimate: components with low lung overlap and no near-surface adjacency are discarded, which could drop true chest-wall or multifocal tumors. The paper acknowledges this in Section V but doesn't quantify the sensitivity cost.\n\nVerdict: the central two-stage claim is plausible but not yet established. It deserves peer review, not a desk reject, because the method is directly usable, the experiments are concrete, and the missing ablation is fixable. The referee should ask for (i) an ablation with and without the post-processing filters, (ii) patient-level statistics, and (iii) either a calibration metric or a softened claim. If the authors deliver those, this becomes a solid niche contribution.","headline":"Sensible cascade, credible numbers, but the central claim that anatomy-aware filtering drives the gain is unproven because no ablation separates it from ROI cropping.","tokens_in":9446,"tokens_out":3484,"would_cite":false,"duration_ms":35295,"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":"Filtering, not a bigger model, lifts tumor Dice from 0.47 to 0.64","keywords":["tumor segmentation","coarse-to-fine","uncertainty-guided","anatomy-aware post-processing","component selection","CT imaging","Monte Carlo dropout","cascade segmentation"],"falsifier":"Run the pipeline on a held-out set rich in chest-wall, pleural, and multifocal tumors and count true positive components that fail both retention criteria (overlap at or below 80% and either distance to lung surface above 5 voxels or volume at or below 50 voxels). If the filter removes any ground-truth tumor component, recall drops and the claim that filtering preserves clinically relevant tumors is falsified; a cleaner quantitative test is comparing filtered versus unfiltered recall on that subset.","tokens_in":8493,"feed_emoji":"🫁","tokens_out":7024,"duration_ms":65473,"temperature":0.7,"pith_summary":"This paper argues that a fully automated two-stage pipeline—coarse whole-volume tumor segmentation followed by region-of-interest refinement—can substantially improve lung tumor delineation in CT, provided the coarse outputs are filtered with anatomical constraints before refinement. On the primary private dataset, the best configuration raises Swin UNETR Dice from 0.4690 to 0.6447 and cuts the 95th-percentile Hausdorff distance from 187.97 mm to 8.89 mm; smaller gains appear on the public NSCLC dataset. The authors attribute most of the gain to component selection: reducing the number of candidate regions forwarded to the second stage correlates strongly with better Dice and Hausdorff scores. Uncertainty-aware losses did not improve Dice or HD95 over a standard Dice-plus-cross-entropy loss, but they improved boundary calibration and interpretability. If the claim holds, existing segmentation models can be upgraded with a modular post-hoc pipeline without architectural changes.","feed_headline":"Filtering, not a bigger model, lifts tumor Dice from 0.47 to 0.64","feed_subtitle":"Two-stage CT pipeline cuts 95th-percentile Hausdorff distance from 188 mm to 8.9 mm.","key_machinery":"The load-bearing mechanism is the stage-one post-processing module combined with tight ROI cropping. Connected components in the coarse tumor mask are merged by 3D dilation, then filtered by anatomical criteria: volumetric overlap with a lung mask (components below 80% overlap are rejected unless they lie within 5 voxels of the lung surface and contain more than 50 voxels), followed by optional top-k selection. Surviving components are cropped with zero or sixteen voxel margins and passed to a second-stage model trained with the adaptive loss $L_{\\text{adaptive}} = \\alpha(x)\\,L_{\\text{Dice}} + (1-\\alpha(x))\\,L_{\\text{CE}}$ with $\\alpha(x)=\\exp(-U(x))$, where $U(x)$ is the variance of Monte Carlo dropout forward passes. The filtering acts as what the authors call a spatial attention mechanism: fewer, larger, lung-adjacent candidates make the ROI model's job easier and lower predictive uncertainty.","core_discovery":"The paper's central claim is that treating segmentation as a surrogate for localization, then refining only the winning candidate region after anatomy-aware filtering, yields more accurate and clinically plausible tumor masks than full-volume segmentation alone. On the primary dataset, the fully automated Top-1, ROI-0margin pipeline lifts Swin UNETR from Dice 0.4690 and HD95 187.97 mm to Dice 0.6447 and HD95 8.89 mm; on the NSCLC dataset the same pipeline improves Swin UNETR Dice from 0.4731 to 0.4860. The authors further report a strong monotone correlation between component-count reduction and HD95 improvement ($\\rho = -0.83$, $p < 0.0001$), which they read as evidence that the post-processing filter, not the second-stage network, drives most of the gain.","pith_inferences":["A testable extension beyond the paper is a pipeline with stage-one filtering but no stage-two ROI model, which would isolate how much of the Dice gain comes from filtering alone versus ROI refinement.","The 80%-overlap, 5-voxel, and 50-voxel thresholds are likely dataset-specific; on cohorts with many chest-wall or mediastinal tumors, the retention rule will discard true extrapulmonary components, so a threshold-sweep calibration curve is the natural next experiment.","The correlation between component count and quality may be confounded by tumor difficulty: small isolated tumors are both easier to segment and naturally produce one component, so component count could partly be a proxy for tumor size or site.","The MC-dropout uncertainty maps could be reused beyond loss reweighting, for example to spatially weight the final mask or to flag uncertain boundaries for clinician review, which the paper only gestures at."],"forward_implications":["Top-1 component selection beats Top-2 and All Valid on the primary dataset, so for single-dominant tumors extra components behave as noise; multifocal cases must be handled by relaxing top-k or explicitly detecting satellite lesions.","Tight ROI-0margin crops outperform ROI-16margin crops, meaning the second-stage model benefits from reduced surrounding context once the component is well localized.","The strong negative correlation between component count and Dice/HD95 implies that a stage-one model can be improved mainly by filtering its false-positive components, without retraining.","The framework generalizes to the public NSCLC dataset but with smaller gains, indicating that lesion size and focality modulate the benefit of component filtering.","Because standard Dice-plus-cross-entropy loss matches or beats the uncertainty-aware loss on Dice/HD95, the practical contribution of uncertainty modeling here is calibration and interpretability, not headline metric improvement."],"supporting_citations":[{"why":"Supplies the Swin UNETR architecture and pre-trained encoder weights used for all baseline and stage-one models.","marker":"[10]"},{"why":"Provides the MC-dropout uncertainty quantification approach that the ROI models build on.","marker":"[17]"},{"why":"Supplies the automatic lung segmentation masks used by the overlap and surface-distance filters.","marker":"[18]"},{"why":"Grounds the Monte Carlo dropout variance computation that defines the uncertainty term in the adaptive loss.","marker":"[14]"},{"why":"Establishes the cascade coarse-to-fine segmentation strategy that the two-stage ROI refinement extends.","marker":"[5]"}],"fun_headline_variants":["Anatomy-aware filtering lifts tumor Dice from 0.47 to 0.64","Post-processing, not model size, boosts tumor segmentation Dice","Coarse-to-fine with anatomy priors cuts HD95 from 188 mm to 8.9 mm","Uncertainty-guided cascade lifts tumor Dice from 0.47 to 0.64"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every true tumor component will pass at least one retention rule—more than 80% lung overlap, or adjacency within 5 voxels of the lung surface combined with size above 50 voxels—so that no clinically meaningful tumor is discarded before stage two.","fun_headline_variants_meta":{"raw":{"variants":["Anatomy-aware filtering lifts tumor Dice from 0.47 to 0.64","Post-processing, not model size, boosts tumor segmentation Dice","Coarse-to-fine with anatomy priors cuts HD95 from 188 mm to 8.9 mm","Uncertainty-guided cascade lifts tumor Dice from 0.47 to 0.64"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000326,"raw_usage":{"total_tokens":1818,"prompt_tokens":932,"completion_tokens":886,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":795}},"tokens_in":548,"tokens_out":886,"duration_ms":8498,"temperature":1.0,"reasoning_tokens":795,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:34:58.287789+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a held-out set rich in chest-wall, pleural, and multifocal tumors and count true positive components that fail both retention criteria (overlap at or below 80% and either distance to lung surface above 5 voxels or volume at or below 50 voxels). If the filter removes any ground-truth tumor component, recall drops and the claim that filtering preserves clinically relevant tumors is falsified; a cleaner quantitative test is comparing filtered versus unfiltered recall on that subset.","supporting_citations":[{"cited_title":"Self-supervised pre-training of swin transformers for 3d medical image analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the Swin UNETR architecture and pre-trained encoder weights used for all baseline and stage-one models."},{"cited_title":"Self-supervised learning for organs at risk and tumor segmentation with uncertainty quantification,","cited_arxiv_id":null,"evidence_quote":"Provides the MC-dropout uncertainty quantification approach that the ROI models build on."},{"cited_title":"Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem,","cited_arxiv_id":null,"evidence_quote":"Supplies the automatic lung segmentation masks used by the overlap and surface-distance filters."},{"cited_title":"nnu-net: Self-adapting framework for u-net-based medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Establishes the cascade coarse-to-fine segmentation strategy that the two-stage ROI refinement extends."}],"review_version":1}