{"id":"30d8b4bb-ae21-4b61-8705-c183fc8bf8db","arxiv_id":"2507.16337","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"OP-SAM turns one labeled polyp image into iterative SAM prompts, reaching 76.93% IoU on Kvasir with no retraining.","lead":"This paper presents OP-SAM, a training-free framework that uses one annotated polyp image to automatically generate and refine prompts for SAM, segmenting polyps in new colonoscopy images. The authors report a 76.93% IoU on Kvasir, an 11.44 point gain over prior one-shot methods, which would remove much of the manual labeling burden in polyp screening.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"EPE's stopping rule relies on SAM2's predicted mask IoU, which is never calibrated on colonoscopy data; if miscalibrated, the core prompt-evolution gain and the headline IoU are not established.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing point: EPE's self-evaluation signal is SAM2's predicted IoU plus a coverage threshold, and neither is validated on the medical out-of-distribution setting. This matters because the paper's central claim is not merely that a frozen SAM can segment polyps, but that a fully automatic feedback-driven prompting loop outperforms prior prompting strategies. If the feedback signal is miscalibrated, the loop's behavior is not trustworthy, and the reported advantage from EPE could shrink or vanish under different datasets or support images. The paper provides no calibration curves, no repeat trials with error bars for the main tables, and no oracle-stop ablation that would separate the value of iterative prompting from the value of a particular stopping rule. I agree with the reader's conditional verdict: the method is coherent and the components are plausible, but the empirical headline needs this specific validation before the central claim can be fully accepted. I do not see an internal inconsistency or a reason to reject; the concern is testable and the existing evidence is suggestive rather than conclusive. Therefore the verdict should remain CONDITIONAL as the reader set it, pending the calibration/ablation check.","tokens_in":12269,"tokens_out":8808,"duration_ms":109825,"concrete_test":"Compute SAM2's predicted mask IoU versus true IoU across EPE rounds on held-out Kvasir and PolypGen images, and then rerun Algorithm 1 with the predicted-IoU stopping condition replaced by (a) an oracle stopping condition using true IoU against the annotated mask and (b) a coverage-only condition. If the IoU difference between predicted-stop and oracle-stop exceeds about 2 points on either dataset, the EPE benefit depends on a miscalibrated signal; if the difference is negligible, the concern is resolved. Report calibration curves and per-round stop counts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's distinctive contribution is EPE, and the ablation attributes +7.88 IoU on Kvasir to it. But Algorithm 1's stopping and expansion decisions (lines 12 and 28) use SAM2's predicted mask IoU (line 6) together with a hand-set threshold θ=0.85, while line 22 uses a negative-area threshold η that is never specified in §4.2. No evidence is given that SAM2's IoU predictions are calibrated on colonoscopy images; SAM2 was trained on natural images and can be overconfident or underconfident on out-of-distribution data. If the predicted IoU is overconfident, the loop prematurely terminates after the tight-prior prompt and EPE degenerates to a one-shot prompt; if underconfident, the loop over-expands and accumulates false positives. The reported 76.93% IoU and the 'surpasses random GT oracle' claim may therefore reflect a favorable calibration coincidence on Kvasir rather than a robust automatic evaluation signal. The thresholds were also analyzed and selected on the same Kvasir benchmark (Fig. 6), so the headline number is not an unbiased estimate. This is an empirical robustness risk, not a demonstrated failure, but it is load-bearing because EPE is the central novelty.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OP-SAM, a training-free one-shot polyp segmentation framework built on SAM2 and DINOv2. From a single annotated support image, it generates semantic priors via Correlation-based Prior Generation (CPG), fuses scale-cascaded priors via a reverse-transfer weighting scheme (SPF), and iteratively refines SAM2 prompts through Euclidean Prompt Evolution (EPE), which uses SAM2's internal mask-IoU predictions to decide when to stop or expand prompting. Experiments on five datasets report state-of-the-art IoU/Dice, including 76.93% IoU on Kvasir, which is claimed to surpass a random three-point ground-truth oracle.","tokens_in":12540,"tokens_out":3721,"duration_ms":39668,"significance":"The problem is practically important: reducing per-image annotation for polyp segmentation has clear clinical value, and a training-free pipeline that leverages foundation models is attractive. The paper's ablations (Tables 3–6) support the individual contributions of CPG, SPF, and EPE, and Table 7 shows some robustness to support-image selection. However, the central quantitative claims rest on a single support image with no variance estimates, and the EPE stopping rule depends on SAM2's mask-IoU calibration on out-of-distribution colonoscopy data, which is never validated. If these issues are addressed, the method could be a meaningful advance in one-shot medical segmentation; as presented, the headline numbers overstate the certainty of the result.","major_comments":[{"comment":"The main results, including the 76.93% IoU on Kvasir and the 'surpasses oracle' claim, are reported for a single fixed support image with no error bars or multi-support statistics. Table 7 provides mean and variance only for an unspecified dataset and only for a five-support evaluation, not for the headline comparisons. Please report mean±std over at least five support images for all datasets (or justify why one support image is sufficient) and state the selection procedure and seed.","section":"§4.1, Table 1"},{"comment":"The EPE stopping rule (Algorithm 1, lines 12 and 28) relies on SAM2's predicted mask IoU (line 6) and a fixed threshold θ=0.85, but the paper provides no evidence that SAM2's IoU predictions are calibrated on colonoscopy images. If the predictor is overconfident, EPE terminates prematurely and degenerates to a one-shot prompt; if underconfident, it over-expands and accumulates false positives. Moreover, Fig. 6 selects ϑt and ϑl on the Kvasir test set and the headline number is computed on the same data, so the evaluation is not unbiased. Please include a calibration analysis of SAM2's IoU on a held-out dataset (e.g., PolypGen centers) and report sensitivity of the final IoU to θ within a reasonable range.","section":"Algorithm 1, §4.2, Fig. 6"},{"comment":"The 'Oracle' baseline is defined as randomly choosing three prompt points from the ground-truth mask. This is a weak baseline because random points can cluster or miss large polyp regions, so the conclusion that EPE surpasses the oracle is not strong evidence for the prompting strategy. Please strengthen the oracle to a more informative protocol, such as points selected to maximize coverage of the GT mask, or a GT-derived bounding box, and compare against that.","section":"§4.3, Table 1"},{"comment":"Several hyperparameters used in the method are not specified, which harms reproducibility: the negative-area threshold η in Algorithm 1 line 22, the threshold τ used in the SPF reverse-transfer quality assessment in §3.2, and the maximum number of EPE rounds (mentioned only in prose as 'five rounds' but not given in Algorithm 1). Please report all numerical values and, where relevant, include them in the algorithm pseudocode.","section":"§4.2, Algorithm 1"}],"minor_comments":[{"comment":"The phrase 'surpassing the state-of-the-art by 11.44%' should specify that this is an IoU improvement on Kvasir, not a generic statement across datasets.","section":"Abstract"},{"comment":"The lesion-scaling procedure mentions that 'the blank gap between the zoomed-out polyp and original polyp is filled by inpainting,' but no inpainting method is described; please provide details or a reference.","section":"§4.2"},{"comment":"There is a typo: 'we we input' should read 'we input'.","section":"Table 6 caption"},{"comment":"The explanation of why the Euclidean distance transform (EDT) center is preferable to the bounding-box center would be clearer with a dedicated illustrative figure; the current 3rd-round example in Fig. 2 is difficult to parse.","section":"§3.3, Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable engineering contribution, but the empirical claims are brittle because the headline numbers come from one support image and the EPE stopping rule is tied to SAM2's unvalidated internal IoU. The threshold tuning on the Kvasir test set (Fig. 6) and then reporting the Kvasir result as the main achievement is a concern for a journal-level evaluation; the authors should provide an unbiased evaluation protocol, including calibration analysis and variance across support images. The oracle comparison as defined is too weak and should be replaced or supplemented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Candid take: OP-SAM is a coherent, training-free one-shot polyp segmentation pipeline that does something genuinely new—iteratively generating SAM prompts from a single annotated polyp image, using a coverage-and-IoU loop to decide when and where to prompt. The EPE module is the real novelty, and the paper earns credit for ablating it (Table 6) and for including a support-image variance study (Table 7), which many one-shot papers skip. The assembly of DINOv2 cross- and self-correlation, scale-cascaded priors with reverse-transfer weighting, and the EDT-based prompt point is new relative to PerSAM, Matcher, and ProtoSAM. The math is simple and parameter-free; I don't see internal contradictions.\n\nSoft spots, in proportion: the headline Kvasir number (76.93 IoU) is not as clean as it looks. The EPE thresholds (ϑt, ϑl, θ) are analyzed and selected on Kvasir (Fig. 6) and then reported on Kvasir, so that number is an in-sample optimistic estimate, not an unbiased evaluation. I'd trust the other datasets more, but the main table lacks error bars on query sets. The 'oracle' baseline is random points from GT, which is a weak oracle—a single well-chosen point or an oracle that places points by distance transform would make the 'surpasses oracle' claim less striking. Also, Algorithm 1 uses η (negative area threshold) but §4.2 never defines it; that's sloppy and should be fixed.\n\nThe stress-test point about SAM2's predicted mask IoU being uncalibrated on colonoscopy images is fair. The stopping rule relies on that predicted IoU, and if it's miscalibrated, EPE could terminate early or over-expand. I don't think it's load-bearing in the sense of breaking the method: even a degenerate EPE still leaves a strong one-shot prompt from CPG/SPF, and the ablation shows EPE adds value on Kvasir. But the paper should report a calibration check or at least sensitivity to θ. This is a fixable empirical gap, not a structural flaw.\n\nBottom line: worth a serious referee. The problem is relevant, the method is novel in its assembly, and the evaluation is above average for the subfield. The right outcome is a revise with requests for: repeated-trial error bars, a threshold selection protocol that avoids tuning on the test set, definition of η, and calibration analysis of the EPE stopping rule. My verdict: conditional accept, leaning positive.","headline":"Genuinely new iterative prompting for one-shot SAM polyp segmentation, with solid ablations; the Kvasir headline number is over-tuned and the EPE stopping rule needs a calibration check, but it deserves a real review.","tokens_in":13078,"tokens_out":2512,"would_cite":true,"duration_ms":26983,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single annotated polyp can automatically prompt SAM2 to segment unseen polyps, achieving 76.93% IoU on Kvasir without any training.","keywords":["one-shot segmentation","polyp segmentation","SAM2 prompting","training-free","DINOv2 feature correlation","scale-cascaded prior fusion","Euclidean prompt evolution","medical image segmentation"],"falsifier":"Run the EPE loop over a colonoscopy image set with expert polyp masks, and for each accepted round compare SAM2's predicted quality score with the true overlap against the expert mask; if many rounds that pass the 0.85 threshold have true overlap far below 0.85, the stopping rule is the weak link.","tokens_in":12060,"feed_emoji":"🩺","tokens_out":15079,"duration_ms":129650,"temperature":0.7,"pith_summary":"OP-SAM is built around the claim that one labelled polyp image is enough to segment polyps in new colonoscopy images automatically, without any training or fine-tuning. From a single support image and mask, it generates patch-level semantic priors with DINOv2 features, fuses three scale-augmented versions of the support, and then lets SAM2's own quality estimates guide where to add the next prompt point. The paper reports 76.93% IoU (intersection over union) on Kvasir, 11.44 points above the best existing one-shot method, and also above a random three-point oracle that draws prompts from the ground truth. Across five public datasets plus an extreme-size subset, the method reports consistent gains over prior one-shot SAM approaches. If the result holds, a clinician would only need to annotate one polyp per deployment setting to get automated, generalizable segmentation.","feed_headline":"One annotated polyp segments the rest, no training","feed_subtitle":"OP-SAM turns one annotated mask into automatic SAM2 prompts and beats prior one-shot methods by 11.44% IoU on Kvasir.","key_machinery":"Three coupled components carry the argument. Correlation-based Prior Generation (CPG) computes a query-support cross-correlation matrix from DINOv2 patch features, transfers the support mask through it, then refines the result by repeated multiplication with a Sinkhorn-normalized self-correlation matrix, turning a rough label transfer into a smooth semantic prior. Scale-cascaded Prior Fusion (SPF) creates enlarged and shrunk copies of the support lesion, generates a prior for each scale, and weights the three priors by a confidence IoU obtained by reverse-transferring each candidate prior back onto the known support mask. Euclidean Prompt Evolution (EPE) is the feedback loop that finally prompts SAM2, a promptable image segmentation model: at each round it computes the Euclidean distance transform of the still-uncovered confident prior region, prompts at the point farthest from the background, and iterates until SAM2's predicted mask IoU and prior coverage both clear the 0.85 threshold, with a negative prompt available when the prediction spills outside the loose prior.","core_discovery":"The paper's central claim is that in one-shot SAM-based polyp segmentation, prompt placement can matter more than the amount of ground-truth information, and that a training-free feedback loop can outperform hand-crafted prompt strategies. The proposed pipeline transfers the support mask to the query by cross-correlating DINOv2 patch features, refines the coarse transfer with repeated self-correlation and Sinkhorn normalization, and repeats this for enlarged and shrunk versions of the support lesion; the three priors are weighted by a reverse-transfer confidence IoU that measures how well each prior reconstructs the known support mask. The fused prior is then consumed by an iterative prompting algorithm that adds one Euclidean-distance-transform-centered prompt per round, keeps the accumulated prompt list, and stops when both SAM2's predicted mask IoU and prior coverage exceed 0.85. On Kvasir the authors measure 76.93% IoU, 11.44 points above the best prior one-shot method and above their random three-point ground-truth oracle; ablations attribute the largest single gain to the iterative prompting rule, which drops to 58.04% IoU when the prompt center is replaced by a bounding-box center.","pith_inferences":["Editorial inference: the same cascade—correlation priors from a frozen encoder, adaptive multi-scale fusion, and SAM2's predicted-IoU feedback—could in principle transfer to other lesion types or anatomical structures that SAM2 can already segment, with a single annotated exemplar per class.","Editorial inference: the reported 'surpasses oracle' result is conditional on the oracle being three randomly placed points; a stronger oracle that places prompts optimally from ground truth would be a harder benchmark, so the comparison should be read as evidence about prompt placement rather than about information content.","Editorial inference: because the loop treats SAM2's predicted mask IoU as a trusted stopping signal, the method's margin over baselines may shrink on domains where that prediction is poorly calibrated; a calibration study of the 0.85 threshold would separate prompt-placement gains from prior-generation gains.","Editorial inference: a natural testable extension is to feed EPE's accepted prompts back as pseudo-labels for a small fine-tuning step; a training-free method that already performs well suggests these prompts carry usable supervision, but the paper does not test that upper bound."],"forward_implications":["On Kvasir, OP-SAM's reported 76.93% IoU exceeds the random three-point ground-truth oracle's 71.88% IoU, so the authors conclude that prompt placement matters more than having ground-truth points to sample from.","The scale-cascaded fusion with confidence weighting beats naive averaging: the paper's support-configuration ablation shows one support image with two scale augmentations plus SPF reaches 61.47% prior IoU, ahead of three separate support images with naive averaging at 37.88%.","The EPE stopping rule, powered by SAM2's predicted mask IoU and prior coverage, accounts for the largest single effect in the paper's ablations: replacing its Euclidean-distance-transform center with a bounding-box center collapses Kvasir IoU from 76.93% to 58.04%.","Across five randomly chosen support images, the method's IoU variance is 0.48, compared with 12.57 for PerSAM, 2.55 for Matcher, and 2.80 for ProtoSAM, indicating that performance does not hinge on a lucky support choice.","On the extreme-size Kvasir-H subset, OP-SAM reports 57.31% IoU, 10.26 points above the second-best method, which the authors tie to the scale-cascaded prior fusion."],"supporting_citations":[{"why":"SAM2 is the promptable segmentation model that EPE prompts iteratively and whose predicted mask IoU drives the stopping rule.","marker":"[27]"},{"why":"DINOv2 is the frozen feature encoder whose patch embeddings and self-correlation supply CPG's semantic priors.","marker":"[22]"},{"why":"Provides the self-correlation and Sinkhorn-normalization refinement step that turns the coarse cross-correlation prior into a smoother one.","marker":"[15]"},{"why":"PerSAM is the one-shot SAM prompting baseline whose single-point strategy EPE improves on and outperforms.","marker":"[40]"},{"why":"Matcher is the all-points feature-matching baseline that EPE's iterative prompt selection is designed to beat.","marker":"[17]"},{"why":"An issue report documenting that excessive SAM prompt points degrade output, which motivates EPE's iterative not-all-prompts-at-once design.","marker":"[18]"},{"why":"ProtoSAM is the multi-scale one-shot medical segmentation baseline that motivates SPF and provides the bounding-box-center prompting alternative in ablations.","marker":"[2]"},{"why":"SegGPT is the in-context segmentation baseline that OP-SAM must surpass as a generalized alternative to specialist one-shot methods.","marker":"[35]"},{"why":"IFA is the cross-domain few-shot specialist baseline used to establish OP-SAM's advantage under domain shift.","marker":"[21]"},{"why":"Kvasir-SEG is the primary benchmark on which the headline 76.93% IoU and the oracle comparison are measured.","marker":"[7]"}],"fun_headline_variants":["One polyp mask, no retraining: OP-SAM auto-prompts SAM","OP-SAM turns one annotation into top polyp segmentation","Single mask, iterative prompts: OP-SAM outperforms SOTA","From one polyp to automatic SAM prompts, OP-SAM excels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the model's own estimate of segmentation quality, together with the fixed coverage threshold of 0.85, reliably tells when a polyp has been fully covered; if that estimate is miscalibrated on colonoscopy images, the loop stops early on under-segmented polyps or accepts noisy expansions, and the iterative-prompting gains shrink.","fun_headline_variants_meta":{"raw":{"variants":["One polyp mask, no retraining: OP-SAM auto-prompts SAM","OP-SAM turns one annotation into top polyp segmentation","Single mask, iterative prompts: OP-SAM outperforms SOTA","From one polyp to automatic SAM prompts, OP-SAM excels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":2144,"prompt_tokens":1032,"completion_tokens":1112,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":1039}},"tokens_in":648,"tokens_out":1112,"duration_ms":11520,"temperature":1.0,"reasoning_tokens":1039,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:11:44.517692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the EPE loop over a colonoscopy image set with expert polyp masks, and for each accepted round compare SAM2's predicted quality score with the true overlap against the expert mask; if many rounds that pass the 0.85 threshold have true overlap far below 0.85, the stopping rule is the weak link.","supporting_citations":[{"cited_title":"Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the self-correlation and Sinkhorn-normalization refinement step that turns the coarse cross-correlation prior into a smoother one."},{"cited_title":"Bad results with multiple point prompts","cited_arxiv_id":null,"evidence_quote":"An issue report documenting that excessive SAM prompt points degrade output, which motivates EPE's iterative not-all-prompts-at-once design."},{"cited_title":"Seggpt: Towards seg- menting everything in context","cited_arxiv_id":null,"evidence_quote":"SegGPT is the in-context segmentation baseline that OP-SAM must surpass as a generalized alternative to specialist one-shot methods."},{"cited_title":"Cross-domain few-shot segmentation via iterative support-query correspon- dence mining","cited_arxiv_id":null,"evidence_quote":"IFA is the cross-domain few-shot specialist baseline used to establish OP-SAM's advantage under domain shift."},{"cited_title":"Kvasir-seg: A segmented polyp dataset","cited_arxiv_id":null,"evidence_quote":"Kvasir-SEG is the primary benchmark on which the headline 76.93% IoU and the oracle comparison are measured."}],"review_version":1}