{"id":"0bfbbb49-21ef-45f4-a3f5-cc7398d2b4a2","arxiv_id":"2506.15153","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free SAM/DINOv2 pipeline with Gaussian-sampled negative prompts achieves near state-of-the-art few-shot medical segmentation on CT and MRI abdominal data.","lead":"SynPo improves training-free few-shot medical segmentation by selecting better negative point prompts for SAM using a fused DINOv2-SAM confidence map. It matters because it suggests prompt engineering can match trained few-shot methods without fine-tuning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (9) mixes confidence maps: c_i is defined on the synergy map while μ,σ come from SynMap_neg, making the core negative-prompt selection ambiguous and potentially divergent from the evaluated method.","rationale":"The reader's weakest_assumption focuses on whether the negative-prompt insight generalizes beyond a single pilot and one CHAOS-MRI replacement test. My concern is more fundamental: the mathematical description of that selection mechanism is internally ambiguous. Section 2.3 defines every pixel p_i with confidence c_i on the synergy map, but Eq. (9) thresholds that c_i using μ and σ computed from SynMap_neg. The two maps are not on a common scale, and the paper does not state any normalization. Moreover, the apparent optimal band from Fig. 4 selects background-similarity values above the mean, which as written would pick background-like pixels, contrary to the textural claim that negatives should be placed inside the anatomical region. This is not a disagreement with community consensus; it is a within-paper technical inconsistency that prevents faithful reproduction. The central empirical claim, that training-free prompt engineering matches training-based SOTA, depends on this negative selection step being exactly what was implemented and evaluated. Until the authors provide code or a corrected, unambiguous formulation, the reported gains cannot be attributed to the described algorithm. I therefore preserve the reader's CONDITIONAL verdict: the paper should be accepted only after this ambiguity is resolved, ideally by releasing code and confirming which variant reproduces the quoted numbers. I note also that the CHAOS-MRI mean in Table 1 is arithmetically inconsistent (79.5 vs 81.15), but that is secondary to the selection-mechanism problem.","tokens_in":7606,"tokens_out":7089,"duration_ms":71335,"concrete_test":"Obtain the released code from the project page and inspect the implementation of Eq. (9): trace which tensor supplies c_i and from which tensor μ and σ are computed. Then recompute the Table 3 CHAOS-MRI ablation under both interpretations: (a) c_i from SynMap with μ/σ from SynMap_neg, (b) c_i from SynMap_neg with μ/σ from SynMap_neg. If the reported Dice of 68.90 is reproduced by only one interpretation, the paper text is inconsistent with the evaluated method; if neither matches, the reported ablation is not reproducible from the written description.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is the negative point selection strategy. In Section 2.3, p_i is defined as a pixel of the synergy map SynMap (target similarity), with c_i its confidence. Yet Eq. (9) thresholds c_i using μ and σ derived in Eq. (6) from SynMap_neg, the background-similarity map. Since SynMap and SynMap_neg are different distributions on different scales, using background μ/σ to select target-confidence values is mathematically unmotivated. The paper's claimed optimal band, [μ, μ+1.5σ] (Fig. 4, α=0, β=-1.5), selects values above the mean of SynMap_neg, i.e., pixels that are more background-like; this contradicts the stated insight that negative prompts should lie inside the organ just below target confidence. If the authors intended to select from SynMap_neg instead, the notation is wrong; if they intended SynMap, the threshold is scale-incompatible. Because no code is provided and α/β are not reported for the final experiments, the reader cannot determine which variant produced the ablated gains (+12.04 in Table 2, +13.18 in Table 3). The core empirical claim therefore rests on an underspecified selection mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SynPo, a training-free few-shot medical image segmentation method built on SAM and DINOv2. It combines DINOv2 semantic features and SAM spatial cues into a confidence map, selects positive point prompts from high-confidence pixels, and selects negative point prompts from a Gaussian-defined confidence band via K-means clustering, then refines the coarse mask. Experiments on Synapse-CT and CHAOS-MRI report mean Dice values of 79.91 and 81.15, claiming parity with training-based few-shot methods and large gains over PerSAM and ProtoSAM.","tokens_in":7860,"tokens_out":4423,"duration_ms":41437,"significance":"If the results hold, the paper would be significant: it suggests that prompt engineering alone, using frozen LVMs, can match training-based few-shot segmentation on abdominal CT and MRI. The strengths are the simple training-free framework, the concrete hypothesis that informative negative prompts lie inside the anatomical region below the target confidence, and an ablation isolating the negative-point strategy. However, as written, the empirical case is not yet persuasive: the headline CHAOS-MRI mean is arithmetically inconsistent, the negative-selection equation mixes confidence maps with incompatible statistics, and key hyperparameters and comparison protocols are undisclosed.","major_comments":[{"comment":"The reported mean Dice for CHAOS-MRI is internally inconsistent. The four organ scores in the SynPo row, 80.30 (spleen), 77.32 (liver), 77.32 (left kidney), and 83.04 (right kidney), average to 79.50, not the reported 81.15. The Synapse-CT row averages to 79.93, matching the reported 79.91, so the inconsistency is specific to CHAOS-MRI. This affects the abstract, Section 3.2, and the central claim of parity with GMRD (82.90). Please correct the arithmetic and re-check all summary statistics.","section":"Section 3.2, Table 1"},{"comment":"Eq. (9) selects negative pixels using c_i, defined as the confidence of a pixel p_i in SynMap, but thresholds c_i against μ and σ estimated from SynMap_neg in Eq. (6). SynMap and SynMap_neg are different maps with different value distributions; applying background-similarity statistics to target-similarity values has no stated justification. If the selection is intended on SynMap_neg, the notation p_i=(x_i,y_i,c_i) should be changed; if it is intended on SynMap, the thresholding rule needs a distributional argument or empirical calibration. This ambiguity is central because Eq. (9) is the core negative-prompt selection mechanism.","section":"Section 2.3, Eq. (9)"},{"comment":"The final configuration of the Point Selection Module is underspecified. The paper does not report the values of α, β, γ_1, γ_2, K_p, or K_n used for the results in Table 1 and the ablations in Tables 2 and 3. Fig. 4 sweeps α with β = α - 1.5 but does not state which point of the sweep was selected or whether the same values were used on Synapse-CT and CHAOS-MRI. Without these parameters, Eqs. (9)-(11) cannot be evaluated or reproduced, and the reported +12.04 and +13.18 gains cannot be traced to a specific selection rule.","section":"Section 2.3 and Section 3.3"},{"comment":"The comparison protocol does not support the strength of the claims. Except for PerSAM and SynPo, all baseline numbers are sourced from their original papers, so the comparison is not matched in folds, image preprocessing, or inference settings. No error bars are given for the training-based methods, and no significance test is reported. The statement that SynPo 'outperforms' training-based methods in spleen and right kidney on Synapse-CT should be softened or backed by a matched evaluation.","section":"Section 3.2, Table 1"},{"comment":"The ablation 'SAM + DINO + PSM' improves from 66.70 to 78.74 Dice, but this row combines the confidence-map fusion, positive-point selection, and negative-point selection. Since the central claim is about negative-prompt quality, the ablation should separate the negative-point selection from the positive-point and clustering components. Table 3 does this on CHAOS-MRI by replacing only PerSAM's negative selection, but the Synapse-CT ablation does not isolate the contribution of Eq. (9).","section":"Section 3.3, Table 2"}],"minor_comments":[{"comment":"In Eq. (10), the condition 'if |P_neg| ≤ γ_2·K_n' should refer to |P'_neg|, since P_neg is the final clustered set defined later in Eq. (11).","section":"Section 2.3, Eq. (10)"},{"comment":"In Eq. (3), the notation 'Ms' appears to denote the inverse of the support mask; using an explicit overbar, e.g., \\bar{M}_S, would avoid confusion with the original support mask M_S.","section":"Section 2.2, Eq. (3)"},{"comment":"The spatial dimensions h, w, H, and W are not consistently fixed; please clarify whether the confidence maps are 64×64 and how the point coordinates are rescaled when they are fed into SAM.","section":"Section 2.2 and Section 2.3"},{"comment":"The label 'RawData' in Fig. 4 is not defined in the text; please specify which dataset and preprocessing it refers to, and explain why it is plotted alongside CHAOS.","section":"Fig. 4"},{"comment":"There are several typographical issues, including 'yeiled' in Section 2.4 and a stray '3' in the ProtoSAM standard-deviation row of Table 1; a careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The CHAOS-MRI arithmetic error and the unresolved Eq. (9) issue make the current version unsuitable for publication as is. The idea is worth another round if the authors can release code or detailed parameters and re-run a matched comparison. I would also ask the editor to verify whether the numbers sourced from previous papers are compatible with the evaluation protocol used here."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea—pick negative prompts inside the organ, just below target confidence—is genuinely interesting, and the reported gains over PerSAM/ProtoSAM look large. But the paper has a central ambiguity in Eq. (9) plus an arithmetic slip, and until those are resolved I would not trust the headline numbers.\n\nWhat is new: fusing DINOv2 and SAM confidence maps is sensible; using a Gaussian band on that map to choose negative points is a concrete, training-free heuristic. The ablation replacing PerSAM's negatives with this strategy (+13.18 Dice on CHAOS) is the most convincing evidence. The method is training-free and competitive with training-based approaches, which matters for medical imaging.\n\nSoft spots, in order of severity. First, the CHAOS-MRI mean is wrong: 80.30, 77.32, 77.32, 83.04 average to 79.5, not 81.15. That needs a correction. Second, and more serious, Eq. (9) defines p_i as a pixel on the synergy map but thresholds its confidence using mu and sigma from SynMap_neg, a different distribution on a different scale. The final alpha and beta are not disclosed, and Fig. 4 suggests the best CHAOS setting is [mu, mu+1.5sigma], which on the background-similarity map would select more background-like pixels—contradicting the paper's \"inside the anatomical region\" rationale. Without code or disclosed hyperparameters, the reader cannot tell which variant actually produced the gains. This is not a nitpick; the negative-prompt selection is the paper's reason to exist. Third, hyperparameters appear to be tuned on evaluation data, and most baselines are quoted from other papers without matched folds or error bars. Those are common, perhaps acceptable, but should be acknowledged.\n\nCredit where due: the pilot experiment (Fig. 1(3)) is a nice sanity check, the confidence-map fusion is well motivated, and the reported results, if reproducible, would be a useful data point for training-free medical segmentation.\n\nBottom line: worth a serious referee, but not in current form. I would ask for code, a corrected Eq. (9) with explicit use of the right map, final hyperparameter values, and a rerun with validation-based selection. Then the claim \"comparable to training-based SOTA\" might hold. As it stands, I would be uncomfortable citing it.","headline":"A clever negative-prompt heuristic with real gains, but a load-bearing equation is underspecified and the CHAOS numbers do not add up.","tokens_in":8422,"tokens_out":4236,"would_cite":false,"duration_ms":40009,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Negative prompts placed inside the organ, just below target confidence, let training-free segmentation match trained methods.","keywords":["few-shot medical image segmentation","training-free prompt-based segmentation","negative prompt selection","confidence map synergy","Segment Anything Model","DINOv2 features","Gaussian confidence band","abdominal CT and MRI segmentation"],"falsifier":"Run SynPo on a new organ or modality with a control that draws negative points from the background or uniformly; if the inside-the-organ band $[\\mu - \\alpha\\sigma, \\mu - \\beta\\sigma]$ does not beat the control in mean Dice, the central claim is falsified. A direct minimal test is to replace Eq. (9)'s band with $[\\mu - 3\\sigma, \\mu - 2\\sigma]$ on the same datasets, since the margin narrowing or reversing would show the chosen band itself is load-bearing.","tokens_in":7399,"feed_emoji":"🩺","tokens_out":7302,"duration_ms":70331,"temperature":0.7,"pith_summary":"SynPo's central claim is that training-free few-shot medical segmentation fails not because the promptable segmenter is weak but because negative prompts are chosen badly. The paper proposes placing negative points inside the anatomical region, at confidence values a small distance below the target, and fusing DINOv2's semantic confidence map with SAM-ViT's spatial confidence map to make that placement reliable. On Synapse-CT and CHAOS-MRI, this prompt-only method reaches mean Dice of 79.91 and 81.15, surpassing the training-free baselines ProtoSAM and PerSAM and matching the training-based state of the art such as GMRD. If the claim holds, prompt engineering alone can substitute for training in low-annotation medical settings.","feed_headline":"Negative prompts inside the organ lift few-shot medical segmentation","feed_subtitle":"Prompt-only SynPo reaches 79.91 Dice on Synapse-CT and 81.15 on CHAOS-MRI, matching trained models.","key_machinery":"The load-bearing object is the Synergy confidence map, defined as $\\mathrm{SynMap} = \\delta_{S\\text{-}D}(S_{\\mathrm{SAM}}\\odot S_{\\mathrm{DINO}}) + \\delta_S S_{\\mathrm{SAM}} + \\delta_D S_{\\mathrm{DINO}}$, which multiplies the two confidence maps so that only pixels where both encoders agree retain high confidence while a weighted sum preserves each encoder's strengths. The negative-confidence distribution is fit to the fused map for background features, and the point selection restricts negative candidates to the interval $[\\mu - \\alpha\\sigma, \\mu - \\beta\\sigma]$. K-means then condenses the selected coordinates into the final positive and negative prompts. This machinery is what lets the paper turn fusion and point placement into performance.","core_discovery":"The discovery is a placement rule for negative prompts in point-promptable segmentation. Instead of choosing pixels least similar to the target, which tend to fall in background that SAM already ignores, SynPo models the distribution of negative confidence values with a Gaussian and draws negative points from the band $[\\mu - \\alpha\\sigma, \\mu - \\beta\\sigma]$, i.e. pixels that are less similar than the target but still within or near the anatomical region. These negative points, together with top-ranked positive points condensed by K-means, are fed to SAM. On the four-organ abdominal benchmarks the rule yields 79.91 mean Dice on Synapse-CT and 81.15 on CHAOS-MRI, outperforming training-free baselines (ProtoSAM 73.45 and PerSAM 62.34 on Synapse) and reaching the level of trained methods such as GMRD (78.52 on Synapse).","pith_inferences":["If the negative-band principle is model-agnostic, the same Gaussian-band selection should transfer to other promptable segmentation models; a cross-backbone test would separate the principle from SAM-specific behavior.","The paper fixes $\\alpha$ and $\\beta$ globally; per-organ tuning might raise scores further, or reveal that the band is really selecting boundary-adjacent pixels rather than simply 'less similar' pixels.","Because the synergy weight on the product term is 0.8, an ablation that varies $\\delta_{S\\text{-}D}$ would clarify how much of the gain comes from the fusion as opposed to the negative-prompt rule.","The pilot evidence for the core premise is qualitative; a systematic quantitative study across organs and modalities would test whether inside-the-organ negative points are always better than background negatives."],"forward_implications":["Training-free few-shot segmentation can reach trained-method accuracy on abdominal CT and MRI when negative prompts are selected from inside the anatomical region, suggesting that annotation cost can be cut without retraining.","The confidence-map fusion rule is transferable: the same fixed weights ($\\delta_{S\\text{-}D}=0.8$, $\\delta_S=\\delta_D=0.1$) work on both Synapse-CT and CHAOS-MRI, so the method does not need per-dataset retuning.","Replacing PerSAM's negative selection with SynPo's band rule adds 13.18 Dice on CHAOS-MRI, implying that negative prompt placement, not the positive points alone, is the dominant source of gain in this paradigm.","The Noise-aware Refine module adds only a smaller gain (+1.19 Dice) after the point-selection gains, so future work can focus on prompt placement rather than refinement."],"supporting_citations":[{"why":"Supplies the promptable segmentation model whose point prompts the method optimizes.","marker":"[6]"},{"why":"Supplies the DINOv2 semantic features used in the confidence-map fusion.","marker":"[8]"},{"why":"Supplies the SAM 2 encoder and decoder used to extract spatial features and to predict masks.","marker":"[10]"},{"why":"ProtoSAM is the strongest training-free baseline that SynPo must beat on the benchmarks.","marker":"[1]"},{"why":"PerSAM provides the training-free baseline and the negative-prompt strategy that SynPo replaces with a 13.18 Dice gain.","marker":"[18]"},{"why":"GMRD is the training-based state-of-the-art method that SynPo aims to match in performance.","marker":"[2]"},{"why":"ADNet supplies the evaluation protocol and a training-based comparison point for the experiments.","marker":"[4]"}],"fun_headline_variants":["Smarter negative prompts make training-free SAM match trained models","SynPo: negative points placed near the organ, not far, boost segmentation","Drop distant negative prompts, keep them close to the target","Training-free few-shot segmentation improved by Gaussian-sampled negatives","SynPo: pull negative prompts from a band near the organ, not random background"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that SAM is best guided by negative points placed inside the anatomical region at confidence values just below the target, rather than by background points; if that behavior does not hold for other organs or modalities, the gains collapse.","fun_headline_variants_meta":{"raw":{"variants":["Smarter negative prompts make training-free SAM match trained models","SynPo: negative points placed near the organ, not far, boost segmentation","Drop distant negative prompts, keep them close to the target","Training-free few-shot segmentation improved by Gaussian-sampled negatives","SynPo: pull negative prompts from a band near the organ, not random background"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2618,"prompt_tokens":908,"completion_tokens":1710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":1619}},"tokens_in":524,"tokens_out":1710,"duration_ms":13287,"temperature":1.0,"reasoning_tokens":1619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:41:48.639854+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SynPo on a new organ or modality with a control that draws negative points from the background or uniformly; if the inside-the-organ band $[\\mu - \\alpha\\sigma, \\mu - \\beta\\sigma]$ does not beat the control in mean Dice, the central claim is falsified. A direct minimal test is to replace Eq. (9)'s band with $[\\mu - 3\\sigma, \\mu - 2\\sigma]$ on the same datasets, since the margin narrowing or reversing would show the chosen band itself is load-bearing.","supporting_citations":[],"review_version":2}