{"id":"efa8edba-59a4-4550-9fe4-05eaa4503d8b","arxiv_id":"2507.19074","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A self-training teacher-student framework raises pulmonary vessel segmentation precision on COPD CT scans from 88.0% to 90.3% while keeping Dice overlap nearly unchanged.","lead":"The authors trained a model to outline blood vessels in CT scans of COPD patients using a handful of manually drawn examples, then had the model label additional scans by itself. This semi-supervised approach reduces false-positive detections and yields vessel measurements that differ across COPD severity stages.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Precision gain may come from extra pseudo-labeled data, not from the precision-based selection strategy: the design lacks an ablation that trains on randomly selected or all unlabeled images.","rationale":"The reader's conditional verdict already asks for error bars, baseline comparisons, and code release. My stress-test converges on the missing baseline comparison as the most load-bearing concern: the claimed benefit of precision-based pseudo-label selection is confounded with the amount of additional pseudo-labeled data used in Semi1 and Semi2. The reader's weakest_assumption focused on the circularity of the selection metric, which is related but distinct; I partially agree because the selection reference is indeed the teacher's own prediction rather than human labels, yet the more decisive flaw is the absence of a random/all-data control. The concrete random-selection ablation would settle whether the selection strategy matters. The paper has strengths: a clear clinical motivation, a reasonable interactive annotation pipeline, and a reproducible-enough training description, but the headline precision improvement is not yet causally or statistically secured. Since this concern does not change the reader's overall conditional stance but sharpens the requested revisions, the verdict remains CONDITIONAL/UNCHANGED.","tokens_in":16436,"tokens_out":5952,"duration_ms":67536,"concrete_test":"Run a matched ablation on the same 12 labeled and 103 unlabeled split: train a student with the same architecture, loss, augmentation, and training schedule, but select the 80 unlabeled images for pseudo-labeling randomly instead of by precision/stability, then evaluate on the same 10 test scans. If this random variant achieves a test precision within the bootstrap 95% confidence interval of Semi2's 0.903, the precision-based selection strategy is not the cause of the reported gain and the central mechanism claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the proposed precision-based pseudo-label selection strategy reduces false positives, with Semi2 improving precision from 0.880 to 0.903. The experimental design, however, cannot isolate this mechanism. Fully uses 12 labeled images; Semi1 adds 40 precision-selected pseudo-labeled images; Semi2 adds another 40. The precision improvement could simply reflect training on 80 additional images, or the benefit of a second self-training iteration, rather than the selection criterion. No control trains a student on 40 or 80 randomly selected pseudo-labeled images, on all 103 unlabeled images, or on images selected by a different criterion such as Dice only. Without such an ablation, the causal attribution to precision-based selection is unsupported. Additionally, the selection reference is the teacher's own best-model prediction, not human ground truth, so a teacher with systematic false positives could select images where those errors are consistent across checkpoints, and the student could inherit them. The reported test evaluation uses only 10 patients and no per-patient error bars or significance test, so the word 'significantly' in the abstract is not backed statistically. Both issues are testable, but the missing control is the more direct threat to the mechanism claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a semi-supervised self-training framework, Semi2, for pulmonary vessel segmentation in non-contrast CT images of COPD patients. A teacher model is first trained on 12 interactively annotated CT scans; it then generates pseudo-labels for 103 unlabeled scans, from which 40 are selected per iteration using a precision-based criterion (Mean Precision > 0.9 in iteration 1; Mean Precision > 0.95 and Mean Dice > 0.85 in iteration 2). The student is retrained on labeled plus selected pseudo-labeled data, and the process is repeated once more. Using nnFormer as the backbone and HRA loss, the authors report that Semi2 improves precision from 0.880 (fully supervised) to 0.903 on a 10-patient test set, with Dice and IoU remaining nearly unchanged. The paper also applies the segmentation to quantify pulmonary vessel parameters across GOLD stages 1–4, reporting statistically significant differences in several metrics such as TBV, BV5/TBV, and radius-bin counts. The code is promised to be released.","tokens_in":16720,"tokens_out":3902,"duration_ms":39677,"significance":"If the reported precision gain is real and caused by the proposed pseudo-label selection strategy, the work has practical value for reducing false positives in pulmonary vessel segmentation, which is clinically relevant for COPD quantification. The interactive annotation pipeline and the downstream vessel-parameter analysis across GOLD stages are useful contributions. The paper also makes a machine-checkable claim (precision 0.903 vs 0.880) and commits to code release, which aids reproducibility. However, the central mechanism claim is not yet supported by the experimental design: no control isolates the effect of precision-based selection from simply adding more pseudo-labeled data or performing more training iterations. The statistical evidence is also thin, as the test set comprises only 10 patients and no uncertainty or significance testing is reported for the segmentation metrics.","major_comments":[{"comment":"The claimed benefit of precision-based pseudo-label selection is not isolated. The comparison is Fully (12 labeled images) versus Semi1 (12 + 40 selected pseudo-labeled) versus Semi2 (12 + 80 selected). Without controls that train on 40 or 80 randomly selected pseudo-labeled images, on all 103 unlabeled images, or on images selected by a different criterion (e.g., Dice only), the precision improvement from 0.880 to 0.903 could be due simply to additional training data or to the extra iteration, rather than to the precision-based selection strategy. Add such ablations with matched amounts of pseudo-labeled data.","section":"Section 3.5 and Table 2"},{"comment":"The abstract's word 'significantly' is not backed by statistical testing. The segmentation metrics are reported only as point estimates on a 10-patient test set, with no standard deviations, confidence intervals, or significance tests. Provide per-patient results (e.g., box plots) and a paired statistical test (e.g., Wilcoxon signed-rank) comparing Fully and Semi2 for precision, Dice, IoU, and sensitivity.","section":"Abstract and Table 2"},{"comment":"There is an inconsistency between the pseudocode and the text. Algorithm 1 assigns D_u2 = D_u - D_u1 and then trains the student on (D_l ∪ D_u1 ∪ D_u2), implying that all 63 remaining unlabeled images are used in the second iteration. Section 3.5 states that only 40 reliable pseudo-labels are selected in the second iteration with specified Mean Precision and Mean Dice thresholds. Clarify whether the remaining 23 images are discarded, and modify the pseudocode to include the selection step for both iterations.","section":"Algorithm 1 and Section 3.5"},{"comment":"The pseudo-label selection is self-referential: 'reliable' is defined by agreement among the teacher's own checkpoints and by precision computed against the teacher's best-model prediction, which is the same precision metric later reported as the gain. A teacher with systematic false positives could select images where those errors are consistent across checkpoints, and the student could inherit them. Although the final test evaluation is against human annotations, this confound limits the mechanistic claim that precision-based selection reduces false positives. Add an analysis that validates the selected pseudo-labels against human annotations on a subset, or explicitly discuss this limitation.","section":"Section 3.5 and Discussion"}],"minor_comments":[{"comment":"Equation (1) is typeset imprecisely; the summation limits and the definition of the index n are unclear. Please rewrite the equation and define all symbols.","section":"Section 3.4"},{"comment":"There are typos, including 'Bonfroni' instead of 'Bonferroni' and 'GOLD' without a grade number in the ANOVA sentence ('GOLD 1, GOLD 2, GOLD, and GOLD 4').","section":"Section 5.4"},{"comment":"The caption mentions 'a, b, and c' but the figure legend does not clearly explain these labels; please annotate the figure or explain the subprocesses in the caption.","section":"Figure 6"},{"comment":"The vessel parameter calculation is said to be based on VesselVio, but the paper does not specify which segmentation model (Fully, Semi1, or Semi2) generates the vessel trees used for the COPD analysis. Please state this explicitly, as the parameter values may depend on the model.","section":"Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper combines existing components (nnFormer, HRA loss, ST++-style self-training) into a medical application, so the novelty is incremental. The decisive issue is the missing ablation for the pseudo-label selection mechanism; without it, the central claim is not established. The reported 2.3% precision gain on 10 patients is also likely fragile. If the authors add the suggested controls and statistical analysis, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you want a clean example of how a decent applied semi-supervised learning paper can over-claim its mechanism. The authors adapt ST++ to pulmonary vessel segmentation in COPD, curate a 125-patient single-center dataset with interactive annotation, and show that two rounds of self-training with precision-filtered pseudo-labels nudge precision from 0.880 to 0.903 on a 10-patient test set. The downstream GOLD-grade vessel parameter analysis is a plus and appears to be a genuine new application.\n\nWhat the paper does well: the annotation workflow is described carefully, the baseline comparison among nnFormer, UNETR, Swin UNETR, and nnU-Net is straightforward, and the reported precision gain, though small, is directionally consistent across iterations. The vessel parameter results are statistically analyzed with ANOVA and post-hoc tests, which is more than many segmentation papers do. The authors also openly acknowledge the trade-off: Dice and sensitivity dip slightly as precision rises.\n\nThe soft spots are real. The biggest one, which the stress-test note correctly identifies, is the missing ablation. Semi1 adds 40 precision-selected pseudo-labeled images, Semi2 adds another 40, but there is no control that trains on 40 or 80 randomly selected pseudo-labels, or on all unlabeled images, or on a different selection metric. So the 2.3% precision improvement might just be the benefit of more training data or a second iteration, not the precision-based selection rule. The paper's central causal claim is therefore not supported by the experimental design. Second, the test set is 10 patients; there are no per-patient confidence intervals or significance tests, so the abstract's 'significantly' is not backed. Third, selecting pseudo-labels by agreement with the teacher's own best checkpoint and by Mean Precision means the teacher's systematic errors can be self-consistent and inherited; this doesn't make the final evaluation circular (human labels are used at test time), but it weakens the inference that the selected labels are anatomically better.\n\nThe vessel parameter analysis is exploratory but adequate; the statistics are present, the story is coherent, and the limitations section acknowledges single-center data. I would not treat the BV5/TBV trend as a new biomarker without external validation.\n\nBottom line: this is a competent, useful engineering contribution with a valuable new dataset, but the mechanism claim needs a proper ablation before publication. It deserves a serious referee, not desk rejection, and the revision should ask for the random-selection control, per-case error bars, and a significance test on the 2.3% precision gap. I would bring it to a reading group for medical imaging folks, but I would not cite it in my own work as evidence for precision-based pseudo-label selection.","headline":"A decent applied extension of ST++ to pulmonary vessel segmentation, undermined by a missing ablation that leaves the precision-gain mechanism unproven.","tokens_in":17197,"tokens_out":2653,"would_cite":false,"duration_ms":26823,"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":"Iterative self-training lifts pulmonary vessel segmentation precision on COPD CT scans to 90.3%, and the resulting vessel metrics separate GOLD severity levels.","keywords":["semi-supervised learning","self-training","pulmonary vessel segmentation","COPD","CT imaging","pseudo-label selection","teacher-student model","vessel quantification"],"falsifier":"On the 10 fully annotated test scans, have two radiologists independently count false-positive vessel fragments in the fully supervised, Semi1, and Semi2 outputs; if the 2.3-point precision gain is not matched by a drop in these fragment counts, the reported improvement is an artifact of the precision-based pseudo-label filter rather than a real reduction in false vessels.","tokens_in":16293,"feed_emoji":"🫁","tokens_out":9721,"duration_ms":86107,"temperature":0.7,"pith_summary":"The paper aims to show that a semi-supervised teacher-student loop can segment pulmonary vessels from CT scans of COPD patients almost as well as full supervision while using far fewer human annotations. The central claim is that two rounds of self-training—called Semi2—raise segmentation precision from 88.0% to 90.3% on a 10-patient test set, with Dice overlap nearly unchanged, by selecting pseudo-labels that the teacher's checkpoints agree on and that pass a precision threshold. If true, the practical payoff is a precise vessel segmenter built from only 12 labeled scans plus unlabeled ones, and vessel measurements that distinguish disease severity. The paper also reports that vessel metrics computed from Semi2 outputs, including a drop in BV5/TBV from 0.5701 to 0.5213 as GOLD grade rises, are statistically different across severity groups, which it offers as evidence of clinical usefulness.","feed_headline":"Self-training lifts lung-vessel precision to 90.3%","feed_subtitle":"Precision-filtered pseudo-labels hold overlap accuracy steady and make COPD vessel metrics severity-sensitive.","key_machinery":"The load-bearing mechanism is the pseudo-label reliability filter, not the network architecture itself. For each unlabeled CT scan, the teacher's $K$ saved checkpoints each output a pseudo-mask, and the paper computes a stability score $s_i = \\sum_{j=1}^{K-1}\\operatorname{meanIoU}(M_{ij}; M_{iK})$ between checkpoint predictions; scans whose pseudo-labels pass a mean-precision threshold are declared reliable and enter the student's training set. The student is trained on the 12 labeled scans plus the selected pseudo-labeled scans with weak or strong augmentation and a hard-region-adaptation loss that concentrates cross-entropy on difficult vessel borders, then the student becomes the teacher for the next iteration. Choosing by precision rather than Dice is what suppresses false-positive vessel voxels, which the paper argues matters most for downstream COPD quantification.","core_discovery":"On its own terms, the paper establishes that precision-filtered self-training is the active ingredient behind a 2.3-percentage-point precision gain in pulmonary vessel segmentation. A teacher model trained on 12 labeled non-contrast CT scans produces pseudo-labels for 103 unlabeled scans; for each scan, the teacher's saved checkpoints yield a stability score, and only the 40 pseudo-labels with mean precision above 0.9 (first round) or mean precision above 0.95 and mean Dice above 0.85 (second round) are kept. After two iterative retraining rounds, the student reaches 90.3% precision on the 10-scan test set versus 88.0% for the fully supervised teacher, while Dice moves only from 0.855 to 0.850 and IoU from 0.747 to 0.740. The same segmentations feed a vessel-quantification step, producing total blood volume, surface area, BV5/TBV, and radius-binned segment counts; these show statistically significant differences between GOLD 1-4 groups, including a fall in the small-vessel fraction BV5/TBV and a rise in the count of the largest radius bin R(3-4) as severity increases.","pith_inferences":["Because the selection threshold (mean precision) is the same quantity reported as the gain, part of the reported improvement may be the model optimizing the selection metric rather than learning anatomically better vessels; a skeleton-based precision metric or radiologist review of false-positive fragments would separate the two.","The fall in sensitivity from 0.832 to 0.804 across iterations shows the precision gain is partly a precision-recall tradeoff; a future version could add a per-scan recall floor to protect the smallest distal vessels.","The same stability-score selection could be applied to other branching structures, such as airways or coronary arteries, since the filter does not encode vessel-specific anatomy.","The BV5/TBV decline across GOLD grades could be amplified by segmentation bias, because emphysematous low-density regions change the error pattern; external datasets or pathology correlation would test whether the biomarker is real or an artifact of the segmenter."],"forward_implications":["A precision-focused self-training loop can cut annotation cost: 12 labeled CT scans plus unlabeled scans outperform a fully supervised model trained on the same 12 labels in precision (90.3% vs 88.0%) while keeping Dice and IoU nearly flat.","Vessel metrics from the improved segmentations separate GOLD severity groups on several parameters, with BV5/TBV falling from 0.5701 (GOLD 1) to 0.5213 (GOLD 4).","The count of the largest vessels, radius bin R(3-4), rises with severity from 7 to 17 segments and is significantly different at GOLD 4, pointing to vascular remodeling rather than only pruning.","Because the selection criterion targets precision, the same scheme is a template for tubular-structure segmentation tasks where false positives are the dominant clinical cost."],"supporting_citations":[{"why":"Supplies the self-training reliable-pseudo-label selection strategy that the method adapts.","marker":"[41]"},{"why":"Supplies the transformer-based U-Net architecture used as the teacher and student model.","marker":"[17]"},{"why":"Supplies the pre-trained model whose outputs seed the interactive annotation process.","marker":"[42]"},{"why":"Supplies the hard-region-adaptive loss used during training.","marker":"[43]"},{"why":"Supplies the algorithm that turns segmented vessels into volume, surface, and radius-bin metrics.","marker":"[46]"},{"why":"One of the fully supervised baselines whose lower precision motivates the self-training comparison.","marker":"[44]"},{"why":"Another fully supervised transformer baseline used to position the architecture choice.","marker":"[20]"}],"fun_headline_variants":["Self-training boosts lung-vessel precision by 2.3%","Vessel segmentation gains 2.3% precision via self-training","Precision-filtered pseudo-labels lift vessel segmentation","COPD vessel metrics sharpen with semi-supervised segmentation","Semi-supervised vessel segmentation improves precision to 90.3%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pseudo-labels selected because they agree with the teacher's own checkpoints and pass a precision threshold are reliable enough to teach a student that will also perform better on human-annotated test images.","fun_headline_variants_meta":{"raw":{"variants":["Self-training boosts lung-vessel precision by 2.3%","Vessel segmentation gains 2.3% precision via self-training","Precision-filtered pseudo-labels lift vessel segmentation","COPD vessel metrics sharpen with semi-supervised segmentation","Semi-supervised vessel segmentation improves precision to 90.3%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1284,"prompt_tokens":1100,"completion_tokens":184,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":716,"completion_tokens_details":{"reasoning_tokens":98}},"tokens_in":716,"tokens_out":184,"duration_ms":2309,"temperature":1.0,"reasoning_tokens":98,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:01:00.954862+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the 10 fully annotated test scans, have two radiologists independently count false-positive vessel fragments in the fully supervised, Semi1, and Semi2 outputs; if the 2.3-point precision gain is not matched by a drop in these fragment counts, the reported improvement is an artifact of the precision-based pseudo-label filter rather than a real reduction in false vessels.","supporting_citations":[{"cited_title":"3D U-Net: learning dense volumetric segmentation from sparse annotation,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained model whose outputs seed the interactive annotation process."},{"cited_title":"Linking convolutional neural networks with graph convolutional networks: application in pulmonary artery -vein separation,","cited_arxiv_id":null,"evidence_quote":"Supplies the hard-region-adaptive loss used during training."},{"cited_title":"The CT pulmonary vascular parameters and disease severity in COPD patients on acute exacerbation: a correlation analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer-based U-Net architecture used as the teacher and student model."},{"cited_title":"A Pulmonary Vascular Extraction Algorithm from Chest CT/CTA Images,","cited_arxiv_id":null,"evidence_quote":"Supplies the algorithm that turns segmented vessels into volume, surface, and radius-bin metrics."},{"cited_title":"Learning tubule -sensitive cnns for pulmonary airway and artery -vein segmentation in ct,","cited_arxiv_id":null,"evidence_quote":"One of the fully supervised baselines whose lower precision motivates the self-training comparison."},{"cited_title":"Naviairway: a bronchiole-sensitive deep learning -based airway segmentation pipeline for planning of navigation bronchoscopy,","cited_arxiv_id":null,"evidence_quote":"Another fully supervised transformer baseline used to position the architecture choice."}],"review_version":2}