{"id":"99135a95-d5a5-4e10-b208-f9873aeb07d2","arxiv_id":"2412.16503","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A first-frame-only supervised polyp segmentation method, PSDNet, reaches within roughly 1.6 to 3.1 Dice points of fully supervised performance on SUN-SEG while using one annotated frame per video.","lead":"This paper trains a colonoscopy video segmentation model using a manual mask for only the first frame of each video, cutting annotation cost by about 175x. It combines an off-the-shelf tracker with a learned teacher to generate pseudo labels for all later frames, and reports test accuracy within about 3% Dice of fully supervised training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The back-propagation quality score in §II-C is computed with the same XMem tracker that produces the propagative teacher's labels, creating a systematic bias risk; the paper never validates the score against ground-truth pseudo-label quality.","rationale":"The reader's weakest-assumption pinpoints the same load-bearing concern: the back-propagation scoring's use of XMem as an unbiased judge of both teachers' outputs. My analysis confirms that this is the most critical unproven element of the method. The empirical results on SUN-SEG are strong, and the ablation at least demonstrates that the semantic teacher and scoring contribute to the final performance. However, the scoring mechanism itself is never validated against ground truth; the paper relies on the assumption that back-propagation IoU correlates with true mask quality across both teachers. This is not an internal inconsistency, but it is a correctness risk that could invalidate the mechanism's generality. The proposed test would resolve the concern by measuring the scoring's fidelity on a fully labeled holdout set. Since the concern is addressable and the current evidence is suggestive but not decisive, the conditional-acceptance verdict from the reader remains appropriate; I would not harden it to accept or reject without this validation. I also note the abstract's 'minimum improvement of 4.5%' is inaccurate (the smallest gap over the best competitor is 3.4% in Table I), but that is a reporting issue, not a structural flaw.","tokens_in":8269,"tokens_out":5170,"duration_ms":67017,"concrete_test":"Partition the fully annotated SUN-SEG training clips into a train split (using only first-frame labels for PSDNet training) and a held-out validation split with full masks. Train PSDNet (or a single run with the same seed) on the train split. On the validation split, for every unlabeled frame, compute the propagative and semantic teacher predictions, their back-propagation IoU scores, and their true IoU against the full ground-truth mask. Then measure (a) the fraction of frames where the teacher with higher true IoU is selected by the back-prop rule, and (b) the Spearman rank correlation between back-prop score and true IoU separately for each teacher. If selection accuracy is near 50% or the correlation is markedly lower for the semantic teacher than the propagative teacher, the scoring is biased and the merging rule is not faithfully assessing quality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is the back-propagation scoring method (Eq. 2–3) for merging pseudo-labels from two teachers. This scoring assumes that feeding a candidate mask into XMem, propagating it backward to the first frame, and measuring IoU with the first-frame annotation is an unbiased proxy for mask quality. However, the propagative teacher's forward pseudo-labels pi are themselves produced by XMem. When pi is back-propagated, XMem is operating on its own prior output, likely yielding higher IoU not because pi is more accurate but because it is more temporally consistent with XMem's internal memory dynamics. The semantic teacher's outputs mi, by contrast, are produced by a segmentation network and may not align with XMem's object-tracking prior, so their back-propagation scores could be systematically depressed. If this bias exists, the merging rule would favor the propagative teacher even when the semantic teacher is more accurate, meaning the student could be trained on confidently wrong labels, and the reported gains over SOTA might not generalize to other trackers or datasets. The paper's ablation (Table II) shows a clear benefit from the semantic teacher and scoring, but does not isolate whether the scoring mechanism correctly identifies the better teacher per frame. Without a ground-truth validation of the scoring's fidelity, the central claim that PSDNet achieves near-fully-supervised performance through this specific merging mechanism is not securely established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces the First-Frame Supervised Video Polyp Segmentation (FSVPS) task, in which only the first-frame mask of each training video is annotated, and proposes PSDNet, a dual-teacher semi-supervised framework. A propagative teacher (XMem) produces pseudo-labels by forward propagation from the first frame, while a semantic teacher (an EMA copy of the student) produces per-frame segmentations. The two candidate masks are merged by a back-propagation scoring scheme that propagates each candidate backward to the first frame and measures its IoU with the first-frame annotation. The student is trained with a cross-entropy loss on the first-frame annotation and on the merged pseudo-labels. Experiments on SUN-SEG report consistent improvements over four sparse-frame video and four image-based semi-supervised baselines, and Dice within 1.6-3.1 points of the fully supervised upper bound. The paper also compares with zero-shot UVOS models. The central claim is that first-frame-only supervision can approach fully-supervised performance on the largest video polyp segmentation dataset.","tokens_in":8559,"tokens_out":6246,"duration_ms":54745,"significance":"If the results are reproducible, the paper is practically significant: it reduces annotation to one frame per video and reports large gains over existing semi-supervised baselines on SUN-SEG. The FSVPS task formulation is novel and likely to attract follow-up work. The manuscript also releases code, which supports reproducibility. However, the current evidence does not yet establish the mechanism underlying the gains: the back-propagation scorer is not validated as a faithful quality measure, and the empirical claim lacks statistical confidence intervals and a fully specified comparison protocol.","major_comments":[{"comment":"The abstract states that PSDNet improves over sparse-frame supervised state-of-the-arts with a 'minimum improvement of 4.5% in Dice', but Table I shows the smallest per-dataset improvement over the second-best method (TCCNet) is 3.4% on Seen-Easy (0.900 vs. 0.866). The value 4.5% is the average of the four per-dataset improvements. Please correct the abstract to say 'average improvement' or clearly report the aggregation being used, since the current headline claim is factually contradicted by the paper's own table.","section":"Abstract; Table I"},{"comment":"All quantitative results in Table I are point estimates from a single run, with no standard deviations, error bars, or significance tests. The reported improvements over the second-best method range from 3.4 to 6.2 Dice points, which could be within run-to-run variance for deep segmentation models. Given the central claim of significant superiority, please report mean +/- std over at least three random seeds and, if feasible, paired significance tests (e.g., per-video Wilcoxon signed-rank or paired t-test) against the best competitor.","section":"Section III-C, Table I"},{"comment":"The back-propagation quality score is computed by back-propagating the candidate mask with XMem, the same model that generates the propagative teacher's forward pseudo-labels. This creates a systematic bias risk: XMem's backward output is likely to be more consistent with its own forward memory dynamics than with the semantic teacher's segmentation output, so the score may favor the propagative teacher regardless of true segmentation quality. The ablation in Table II demonstrates an overall benefit of adding the back-propagation scorer but does not show that the scorer selects the better teacher per frame. Please validate the scorer by (a) measuring the correlation between the quality score and the IoU of each candidate mask against ground truth on a held-out annotated subset, or (b) comparing PSDNet's automatic selection against an oracle that always selects the candidate with higher ground-truth IoU.","section":"Section II-C, Eqs. (2)-(3); Table II"},{"comment":"The manuscript does not state how many labeled frames are used by each of the eight compared semi-supervised methods. Several of these methods, such as TCCNet and SSTAN, were designed for sparse-frame supervision rather than first-frame-only supervision; if they are given more than one labeled frame per video, the comparison is not apples-to-apples. Please specify the exact annotation budget for each competitor and, if necessary, re-implement them under the first-frame-only setting. Without this information, the headline claim of superiority over sparse-frame supervised state-of-the-arts is not fully established.","section":"Section III-C"}],"minor_comments":[{"comment":"There are several typos: 'foucs' should be 'focus' in Section I, 'absolution' should be 'absolute' in Section III-A, 'reisze' should be 'resize' in Section III-B, 'Tabel III' should be 'Table III' in Section III-E, and 'FSVOS' in the Conclusion should be 'FSVPS'.","section":"Section I; Section III-A; Section III-B; Section III-E; Conclusion"},{"comment":"The table uses the symbols '!' and '%' without defining them, and the first row 'None - 0.660' appears incomplete. Please use standard check marks and report all row and column values explicitly.","section":"Table II"},{"comment":"The notation 'Propi->1' is confusing because the subscript and arrow suggest a direction opposite to the backward mechanism described in the text. Please clarify the direction of propagation, for example by defining the backward pass explicitly.","section":"Eq. (2)"},{"comment":"The PSDNet column reports a single scalar '0.841 (Test set)' whereas the UVOS baselines appear to be summarized differently. Please report per sub-test set Dice for all methods, and clarify that PSDNet is trained on SUN-SEG while the UVOS baselines are zero-shot prompt-based methods.","section":"Table III"},{"comment":"The EMA weight is set to 0.999, but no sensitivity analysis is provided for this hyperparameter. A short ablation over a few values would strengthen the robustness of the approach.","section":"Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the code release is a positive point. The main risk is the unvalidated back-propagation scorer, which is the central novelty; if the additional validation experiments show that the score is not biased toward the propagative teacher, the paper could be acceptable. The abstract's 'minimum improvement of 4.5%' should be corrected regardless. The comparison with UVOS in Table III is somewhat tangential and could be condensed if space is needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing: this is one of the few papers where the task itself is the contribution. Reducing video polyp segmentation annotation to one frame per video is useful, and the back-propagation scoring idea is genuinely new: propagate a candidate mask backward to the labeled frame and use IoU with the first-frame GT as a quality score. That is clever and intuitive. The dual-teacher design is reasonable, the ablations are informative, and the reported results on SUN-SEG, if reproducible, support the claim that roughly 1/175 labels gets close to fully supervised performance. Credit where it is due.\n\nNow the soft spots. The abstract says a \"minimum improvement of 4.5%\" over SOTA; Table I shows a minimum gap of 3.4% (Seen-Easy). The average of the four gaps is 4.5%, so the intro's wording is fine and the abstract is simply wrong. That is fixable. No error bars or multiple seeds are reported, which matters on a medical dataset. The bigger issue is the methodological self-reference in the scoring: XMem both produces the propagative teacher's pseudo-labels and computes their back-propagation scores. It may systematically favor its own outputs over the semantic teacher's, so the merging rule could be selecting for temporal consistency rather than true segmentation quality. The paper never validates the score against actual mask accuracy, so this remains an open question. I do not think it sinks the paper—the ablation shows clear benefit from scoring, and even a biased selector can be useful—but it should be tested, e.g., by correlating scores with ground-truth IoU on a small annotated subset or by using an independent tracker for scoring. Minor points: the code link is not verified, and the annotation-budget arithmetic is approximate but fine.\n\nThe citation pattern looks solid and the comparison to prior semi-supervised and UVOS methods is fair. The paper is written for anyone working in label-efficient medical video segmentation or semi-supervised video segmentation; they will get a concrete method and a benchmark to beat.\n\nBottom line: this deserves a serious referee. The task is real, the method is non-trivial, and the empirical evidence is strong enough to justify the claim, caveats included. I would send it to review and ask the authors to correct the abstract and add the scoring validation before acceptance.","headline":"A genuinely new task and a clever pseudo-label scoring trick, but the abstract overstates the margin and the scoring's self-reference with XMem is not validated.","tokens_in":9044,"tokens_out":2603,"would_cite":true,"duration_ms":24322,"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":"PSDNet demonstrates that a single annotated frame per clip can train a video polyp segmenter to near-fully-supervised accuracy, by merging a tracker's forward-propagated masks with an EMA teacher's semantic masks using a…","keywords":["first-frame supervised video polyp segmentation","semi-supervised learning","teacher-student network","pseudo-label quality","back-propagation scoring","video object segmentation","colonoscopy","polyp segmentation"],"falsifier":"Compute, on SUN-SEG frames with held-out ground truth, the rank correlation between the backward-propagation IoU score and the true frame-level IoU of each candidate pseudo-label; a non-positive correlation would falsify the claim that the score selects the better teacher output.","tokens_in":8074,"feed_emoji":"🩺","tokens_out":12921,"duration_ms":98657,"temperature":0.7,"pith_summary":"This paper introduces first-frame supervised video polyp segmentation, a task setting where each training video provides exactly one annotated frame, and proposes PSDNet to make that sufficient. PSDNet is a teacher-student setup with two teachers: a universal tracker that propagates the first-frame mask forward, and an EMA copy of the student that produces stable semantic masks. The two sets of pseudo-labels are merged by a backward-propagation scoring rule that propagates each candidate mask back to the first frame and measures its IoU with the known annotation. On SUN-SEG, with roughly 1/175 of the frames labeled, the trained student beats the compared sparse-frame supervised methods by an average 4.5 Dice and lands within 1.6-3.1 Dice of the fully supervised upper bound. If this holds, the main annotation bottleneck for large-scale polyp video segmentation is removed.","feed_headline":"One labeled frame per video rivals full polyp supervision","feed_subtitle":"It beats sparse-frame methods by 4.5 Dice score on four test splits.","key_machinery":"The central mechanism is the back-propagation scoring method. For an unlabeled frame $i$, the propagative teacher produces a mask $p_i$ and the semantic teacher produces $m_i$; each is used to initialize the tracker at frame $i$ and track backward to frame 1, yielding round-tripped masks $e^{1,p_i}$ and $e^{1,m_i}$. Their IoU with the first-frame annotation $y_1$ defines quality scores $C_{p,i}$ and $C_{m,i}$, and the higher-scoring mask becomes the pseudo-label used in the student's cross-entropy loss. This turns the single known label into a temporal consistency check: a good pseudo-label should survive a round trip through the tracker and still align with the known mask.","core_discovery":"PSDNet's central claim is that one annotated frame per video is enough to train a video polyp segmentation model whose accuracy approaches the fully supervised setting. The propagative teacher supplies temporally coherent masks but drifts over long videos; the semantic teacher is an exponential moving average of the student and supplies time-invariant but less temporally anchored masks. The paper's key move is to use the tracker itself as a judge: for each unlabeled frame, each teacher's candidate mask is propagated backward to the first frame, and the IoU of the round-tripped mask with the first-frame ground truth becomes the quality score. The candidate with the higher score is selected as the pseudo-label for training the student. On SUN-SEG this procedure outperforms all compared sparse-frame supervised methods, with an average 4.5-point Dice gain and at least 3.4 points on every sub-test set, and leaves a gap of only 1.6-3.1 Dice relative to the fully supervised baseline.","pith_inferences":["The backward-propagation score is a general pseudo-label filter: any video object segmentation method with one reliable reference mask could apply round-trip consistency to reject drifting labels, independent of the polyp domain.","Since the propagative teacher doubles as the judge, the ceiling of the method is tied to tracker generalization; a tracker fine-tuned on colonoscopy motion and appearance could plausibly close the remaining 1.6-3.1 Dice gap.","With two or more annotated anchors per video, the same scoring idea could be extended to multi-anchor round trips, which might correct late-frame drift where both teachers degrade.","A testable scaling prediction follows: if annotation effort is the bottleneck, Dice should improve only slowly as the number of annotated frames grows from one to a few, until tracker or model capacity becomes limiting."],"forward_implications":["First-frame-only supervision is a viable regime for video polyp segmentation: annotation cost drops to one mask per clip (about 1/175 of SUN-SEG's frames) without collapsing accuracy.","Sparse-frame supervised methods, which annotate dozens of frames per video, are outperformed by a method using strictly fewer labels, by an average 4.5 Dice on SUN-SEG.","Pseudo-label quality in video self-training can be measured by round-trip consistency with a known anchor, not just by model confidence.","The same dual-teacher plus backward-scoring setup is expected by the authors to extend to other medical video segmentation tasks with scarce annotations."],"supporting_citations":[{"why":"Supplies the propagative teacher and the backward-propagation evaluator; forward tracking generates one pseudo-label stream and backward tracking produces the quality scores.","marker":"[19]"},{"why":"Provides the SUN-SEG benchmark and its four test splits on which the claimed Dice gains are measured.","marker":"[9]"},{"why":"Supplies the segmentation network used as the student and as the semantic teacher, and the shared backbone for all compared methods.","marker":"[11]"},{"why":"Gives the strongest sparse-frame supervised competitor; its per-split Dice is the direct reference for PSDNet's claimed improvements.","marker":"[15]"},{"why":"Provides natural-scene video object segmentation data used to pre-train the propagative tracker, supplying its zero-shot propagation ability.","marker":"[23]"},{"why":"Provides natural-scene video object segmentation data used to pre-train the propagative tracker, supplying its zero-shot propagation ability.","marker":"[24]"}],"fun_headline_variants":["One label per video rivals full polyp supervision","Single-frame supervisor beats sparse polyp methods","First frame only gives 4.5+ Dice over sparse","One annotated frame segments polyps in videos","Dual teachers trust one frame for polyp masks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a candidate pseudo-label's quality is faithfully measured by its IoU with the first-frame annotation after a backward trip through the same tracker, and that this test does not systematically favor temporally smooth but spatially wrong masks.","fun_headline_variants_meta":{"raw":{"variants":["One label per video rivals full polyp supervision","Single-frame supervisor beats sparse polyp methods","First frame only gives 4.5+ Dice over sparse","One annotated frame segments polyps in videos","Dual teachers trust one frame for polyp masks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000734,"raw_usage":{"total_tokens":3341,"prompt_tokens":1063,"completion_tokens":2278,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":2208}},"tokens_in":679,"tokens_out":2278,"duration_ms":17065,"temperature":1.0,"reasoning_tokens":2208,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:30:20.458031+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, on SUN-SEG frames with held-out ground truth, the rank correlation between the backward-propagation IoU score and the true frame-level IoU of each candidate pseudo-label; a non-positive correlation would falsify the claim that the score selects the better teacher output.","supporting_citations":[{"cited_title":"Video polyp segmentation: A deep learning perspective,","cited_arxiv_id":null,"evidence_quote":"Provides the SUN-SEG benchmark and its four test splits on which the claimed Dice gains are measured."},{"cited_title":"SALI: Short-term Alignment and Long-term Interaction Network for Colonoscopy Video Polyp Segmentation","cited_arxiv_id":"2406.13532","evidence_quote":"Supplies the segmentation network used as the student and as the semantic teacher, and the shared backbone for all compared methods."},{"cited_title":"Tccnet: Temporally consistent context-free network for semi- supervised video polyp segmentation","cited_arxiv_id":null,"evidence_quote":"Gives the strongest sparse-frame supervised competitor; its per-split Dice is the direct reference for PSDNet's claimed improvements."}],"review_version":1}