{"id":"fc4c5383-d47e-42d8-addf-ffe626bd7527","arxiv_id":"2607.06972","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"HPR-SAM replaces manual prompts in SAM with hierarchical probabilistic anatomical representations, achieving state-of-the-art medical image segmentation on Synapse, LA, and PROMISE12 datasets.","lead":"This paper adapts the Segment Anything Model (SAM) for medical image segmentation by replacing manual prompts with a hierarchical probabilistic representation of anatomy. It achieves state-of-the-art results on standard benchmarks, suggesting that richer anatomical features matter more than prompt-generation tricks.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"SOTA claim on Synapse rests on a 0.30% Dice improvement over 12 test cases with no significance testing and worse HD95 than two competitors; the ranking is not statistically substantiated.","rationale":"The reader's verdict of CONDITIONAL with MODERATE confidence is appropriate. The reader correctly identified the marginal gains, missing significance tests, and worse HD95, but routed the weakest_assumption through the ablation study rather than the statistical fragility of the headline claim. The ablation is actually less concerning than the reader suggests: the full model's HD95 improvement over any two-branch ablation (6+ mm) is substantial, and the 0.77% Dice gain over DAR+LRR is reasonable for this benchmark. The more load-bearing issue is that the SOTA claim itself—0.30% over PG-SAM on 12 cases with no significance testing and worse boundary metrics—may not survive basic statistical scrutiny. This does not change the verdict (CONDITIONAL remains right), but it sharpens the specific condition that should be met: the authors need to demonstrate that their Synapse improvement is statistically real, not just numerically higher. The PROMISE12 few-shot result is the strongest evidence in the paper and provides independent support for the method's value, but it is secondary to the headline Synapse claim. The paper's conceptual framing (representation learning over prompt generator design) is interesting and the ablation structure is reasonable; the problem is purely empirical robustness of the central comparative claim.","tokens_in":12366,"tokens_out":2855,"duration_ms":169879,"concrete_test":"Re-run HPR-SAM and the two closest competitors (PG-SAM, H-SAM) with at least 5 random seeds on the Synapse test set. Report per-case Dice and HD95 as mean±std and perform paired Wilcoxon signed-rank tests (HPR-SAM vs PG-SAM, HPR-SAM vs H-SAM). If the Dice improvement p-value exceeds 0.05 or if the std overlaps the 0.30% margin, the Synapse SOTA claim is not statistically supported. Additionally, report per-case HD95 to confirm the boundary-quality deficit is systematic rather than driven by outlier cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that HPR-SAM achieves state-of-the-art on Synapse (85.09% Mean Dice). However, this rests on a 0.30% improvement over PG-SAM (84.79%) and 0.44% over H-SAM (84.65%), evaluated on only 12 test cases. No statistical significance tests, confidence intervals, or multi-seed variance estimates are reported. With 12 cases, a 0.30% mean Dice difference is well within plausible noise from random seed variation or minor data split perturbations. Compounding this, HPR-SAM's HD95 (11.94mm) is substantially worse than H-SAM (7.29mm) and PG-SAM (7.61mm), meaning the method trades boundary accuracy for a marginal region-overlap gain that may not be real. The paper attributes the HD95 gap to competitors' 'decoder-side refinement for boundary enhancement' but provides no evidence for this claim. The few-shot PROMISE12 result (+4.31% over H-SAM) is more convincing, but the headline SOTA claim on Synapse—the primary benchmark—is the weakest link. The ablation concern the reader raised (whether all three branches are necessary) is secondary: the full model does show a meaningful HD95 improvement over any two-branch combination (11.94 vs 17.74–18.08mm), and the 0.77% Dice gain over DAR+LRR is non-trivial in this context. The more fundamental issue is whether the reported SOTA ranking is statistically real at all.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes HPR-SAM, a prompt-free SAM adaptation framework for medical image segmentation that shifts focus from prompt generator design to anatomical representation learning. The framework introduces three probabilistic representation modules—DAR (global priors), MAR (intra-structure diversity), and LRR (local reliability)—whose predictions are fused via a Hierarchical Prediction Fusion (HPF) module. Experiments are conducted on Synapse (fully supervised), LA and PROMISE12 (few-shot). The method achieves 85.09% Mean Dice on Synapse, claiming state-of-the-art among SAM-based methods, and reports the best few-shot results on LA (84.65%) and PROMISE12 (81.26%). The core idea of modeling anatomical representations probabilistically rather than deterministically is well-motivated, and the ablation study in Table III provides reasonable evidence for component complementarity, particularly regarding HD95 improvement when all three branches are combined.","tokens_in":13411,"tokens_out":1483,"duration_ms":204889,"significance":"The paper addresses a relevant problem in prompt-free SAM adaptation for medical image segmentation. The reframing of prompt quality as a representation expressiveness problem is a reasonable conceptual contribution. The ablation study (Table III) is informative, particularly the HD95 progression showing that the full three-branch model achieves 11.94mm versus 17.74–18.08mm for any two-branch combination—a meaningful improvement that supports the hierarchical design. The release of anonymous code is a positive step toward reproducibility. The few-shot results on PROMISE12 (+4.31% over H-SAM) are more convincing than the Synapse headline claim. However, the significance of the Synapse SOTA claim is tempered by the small test set size (12 cases), the marginal Dice improvement, and the substantially worse HD95 compared to H-SAM and PG-SAM.","major_comments":[{"comment":"§IV-B, Table I: The headline SOTA claim on Synapse rests on a 0.30% Mean Dice improvement over PG-SAM (84.79% vs 85.09%) and 0.44% over H-SAM (84.65%), evaluated on only 12 test cases. No statistical significance tests, confidence intervals, or multi-seed variance estimates are reported. With 12 cases, a 0.30% mean Dice difference is within plausible noise from random seed variation or minor data split perturbations. The authors should either (a) report multi-seed variance and significance tests, or (b) moderate the SOTA claim to acknowledge that the Synapse ranking is not statistically substantiated. This is load-bearing because the Synapse SOTA claim is the primary headline in the abstract and contributions.","section":null},{"comment":"§IV-B, Table I: HPR-SAM's HD95 (11.94mm) is substantially worse than H-SAM (7.29mm) and PG-SAM (7.61mm). The paper attributes this to competitors' 'decoder-side refinement for boundary enhancement' but provides no evidence for this claim—neither a citation, an architectural analysis, nor an ablation isolating the effect of decoder refinement. If H-SAM and PG-SAM do use such modules, the authors should cite the specific components; if not, the attribution should be removed. As stated, the explanation is unsupported and the HD95 gap undermines the claim of 'competitive boundary localization.' The paper should either substantiate this attribution or present the HD95 gap as a known limitation.","section":null},{"comment":"§III-E, Eqs. (19)–(20): The HPF module uses learnable target-specific coefficients (α_c, β_c) that are optimized during training. The reader's report raises a concern about circularity—these coefficients are fitted on training data and then applied to test predictions. While this is standard for learnable fusion weights, the paper does not report the learned values of α_c and β_c, nor does it analyze whether they generalize across datasets or are dataset-specific. Since HPF is a core contribution and the fusion is target-wise, reporting the learned coefficients (or at least their distribution across targets) would strengthen the claim that the fusion is principled rather than overfit. Additionally, it is unclear whether α_c and β_c are shared across the three datasets or re-learned per dataset.","section":null}],"minor_comments":[{"comment":"§IV-A: Implementation details mention LoRA rank 4 and 200 epochs of AdamW, but do not specify the learning rate, weight decay, batch size, or data augmentation specifics. Please add these for reproducibility.","section":null},{"comment":"Table I: The HD95 value for H-SAM appears as '7.29' without clear separation from the Mean Dice value '84.65' in the same cell ('84.657.29'). Please use a line break or separate columns to clearly distinguish Mean Dice and HD95.","section":null},{"comment":"§III-B, Eq. (3): The text '在此处键入公式。' (Chinese for 'Type formula here') appears to be a leftover editing artifact. Please remove it.","section":null},{"comment":"§III-C, Eq. (8): The mixture weights ω_{c,m} are described as 'learnable' but it is not stated whether they are constrained to sum to 1 (e.g., via softmax). Please clarify the normalization.","section":null},{"comment":"§III-D, Eq. (16): The mutual likelihood formula uses ||μ_i − μ̄_c||² / (Σ_i + Σ̄_c) and log(Σ_i + Σ̄_c). If Σ_i and Σ̄_c are diagonal covariance matrices, this should be stated explicitly; if they are scalar variances, the notation should reflect that. Please clarify whether these are scalar, diagonal, or full matrix operations.","section":null},{"comment":"Fig. 1 caption: 'DOR' and 'MOR' appear as abbreviations for DAR and MAR. These do not match the terminology used in the rest of the paper. Please correct.","section":null},{"comment":"§IV-B: The paper states HPR-SAM 'consistently improves segmentation accuracy, surpassing ... H-SAM (+0.44%), and PG-SAM (+0.30%).' These margins are small; describing them as 'consistent' improvements overstates the evidence. Consider softer language.","section":null},{"comment":"References [9] and [23] appear to refer to PGP-SAM and PG-SAM respectively, which seem to be related works from overlapping author groups. The relationship between these works should be clarified for readers.","section":null}],"recommendation":"major_revision","confidential_remarks":"The Synapse benchmark with 12 test cases is a known limitation of this benchmark, and many papers in this space report results without significance testing. However, the combination of (a) marginal Dice improvement, (b) substantially worse HD95, and (c) no variance/significance reporting makes the headline SOTA claim difficult to support as-is. The few-shot results are more convincing and may actually be the stronger contribution. The authors could consider reframing the paper's emphasis toward the few-shot generalization story, which is more defensible. The anonymous code link is a positive signal, but I have not verified its contents."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for the careful and constructive review. The referee raises three major points: (1) the Synapse SOTA claim lacks statistical substantiation given the small test set and marginal Dice improvement; (2) the HD95 gap relative to H-SAM and PG-SAM is attributed to decoder-side refinement without evidence; and (3) the HPF learnable coefficients (α_c, β_c) are not reported or analyzed for generalization. We address each point below and commit to revisions for all three.","responses":[{"response":"The referee is correct that a 0.30% Mean Dice difference on 12 test cases cannot be presented as a statistically reliable improvement without further evidence. We will address this through both requested remedies. First, we will conduct multi-seed experiments (at least 3 seeds) and report mean ± standard deviation for both Mean Dice and HD95 on Synapse, along with paired statistical tests (e.g., Wilcoxon signed-rank test) against PG-SAM and H-SAM. Second, we will moderate the SOTA claim in the abstract and contributions to state that HPR-SAM achieves 'competitive' or 'best-reported' Mean Dice among SAM-based methods, explicitly noting that the margin on Synapse is small relative to the test set size and that statistical significance depends on the multi-seed analysis. We agree that the current phrasing overstates the strength of the evidence. The few-shot results on PROMISE12 (+4.31% over H-SAM) and LA (+0.71% over H-SAM) provide more robust evidence of improvement, and we will foreground these in the framing.","revision_made":"yes","referee_comment":"§IV-B, Table I: The headline SOTA claim on Synapse rests on a 0.30% Mean Dice improvement over PG-SAM (84.79% vs 85.09%) and 0.44% over H-SAM (84.65%), evaluated on only 12 test cases. No statistical significance tests, confidence intervals, or multi-seed variance estimates are reported. With 12 cases, a 0.30% mean Dice difference is within plausible noise from random seed variation or minor data split perturbations. The authors should either (a) report multi-seed variance and significance tests, or (b) moderate the SOTA claim to acknowledge that the Synapse ranking is not statistically substantiated. This is load-bearing because the Synapse SOTA claim is the primary headline in the abstract and contributions."},{"response":"The referee is correct that the current attribution is unsupported. Upon review, H-SAM (Cheng et al., CVPR 2024) employs a hierarchical two-stage decoder with a mask-guided attention mechanism that explicitly refines boundary details, and PG-SAM (Zhong et al., BIBM 2025) incorporates a fine-grained prior-guided refinement module. We will revise the manuscript to cite these specific components rather than making a generic attribution. However, we acknowledge that without an ablation isolating decoder refinement in our own framework, we cannot definitively attribute the full HD95 gap to this factor alone. Therefore, we will also present the HD95 gap as a known limitation of the current framework: HPR-SAM prioritizes representation-level improvements through probabilistic anatomical modeling while retaining the original SAM decoder without additional boundary refinement, which likely contributes to the higher HD95. We will remove the phrase 'competitive boundary localization' and replace it with an honest acknowledgment that boundary precision remains a limitation. We also note that the ablation study (Table III) shows the full three-branch model reduces HD95 from 27.98mm (baseline) to 11.94mm, indicating that the hierarchical representation itself substantially improves boundary quality even without a dedicated refinement module.","revision_made":"yes","referee_comment":"§IV-B, Table I: HPR-SAM's HD95 (11.94mm) is substantially worse than H-SAM (7.29mm) and PG-SAM (7.61mm). The paper attributes this to competitors' 'decoder-side refinement for boundary enhancement' but provides no evidence for this claim—neither a citation, an architectural analysis, nor an ablation isolating the effect of decoder refinement. If H-SAM and PG-SAM do use such modules, the authors should cite the specific components; if not, the attribution should be removed. As stated, the explanation is unsupported and the HD95 gap undermines the claim of 'competitive boundary localization.' The paper should either substantiate this attribution or present the HD95 gap as a known limitation."},{"response":"We agree that reporting the learned fusion coefficients is important for transparency and for substantiating the claim that HPF performs principled fusion. To clarify: α_c and β_c are re-learned per dataset, as they are target-specific and dataset-specific parameters optimized during training on each dataset's training split. We will add a table or figure reporting the learned α_c and β_c values for each anatomical target on the Synapse dataset (and where applicable, on LA and PROMISE12). This will allow readers to inspect whether the fusion weights reflect meaningful patterns—for example, whether targets with higher intra-structure variation (e.g., pancreas, stomach) receive higher α_c values, or whether targets with more ambiguous boundaries receive higher β_c values. We will also add a brief discussion of whether the learned coefficients suggest overfitting (e.g., extreme values concentrating all weight on a single branch) or principled balancing across branches. If the coefficients show dataset-specific patterns, we will discuss this as expected behavior given that anatomical characteristics differ across datasets (abdominal CT vs. cardiac MRI vs. prostate MRI).","revision_made":"yes","referee_comment":"§III-E, Eqs. (19)–(20): The HPF module uses learnable target-specific coefficients (α_c, β_c) that are optimized during training. The reader's report raises a concern about circularity—these coefficients are fitted on training data and then applied to test predictions. While this is standard for learnable fusion weights, the paper does not report the learned values of α_c and β_c, nor does it analyze whether they generalize across datasets or are dataset-specific. Since HPF is a core contribution and the fusion is target-wise, reporting the learned coefficients (or at least their distribution across targets) would strengthen the claim that the fusion is principled rather than overfit. Additionally, it is unclear whether α_c and β_c are shared across the three datasets or re-learned per dataset."}],"tokens_in":12412,"tokens_out":1359,"duration_ms":278672,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper reframes prompt-free SAM adaptation as an anatomical representation learning problem rather than a prompt generator design problem. That is a useful conceptual shift. The three-branch design (DAR, MAR, LRR) is a reasonable decomposition — global priors, intra-structure diversity, local reliability — and the ablation in Table III does show each component contributing. The few-shot PROMISE12 result (+4.31% over H-SAM) is the most convincing experiment. Code is anonymously released, which is good practice. The mathematical formulation is clean and internally consistent throughout Sections III-B through III-E. Credit is due for a well-structured, clearly written framework with a coherent motivating idea. The ablation study is thorough — all seven combinations are reported, which is more than most papers do. The HD95 improvement from the full three-branch model (11.94mm) versus any two-branch combination (17.74–18.08mm) is actually a meaningful gap that the reader and stress-test note both underweight. That is a real result. Now the soft spots. The Synapse SOTA claim rests on a 0.30% Dice improvement over PG-SAM on 12 test cases with no significance testing, no confidence intervals, no multi-seed variance. That is not a substantiated ranking. The HD95 on Synapse (11.94mm) is substantially worse than H-SAM (7.29mm) and PG-SAM (7.61mm), and the paper's explanation — that competitors use decoder-side boundary refinement — is an unsupported assertion. They need to either verify this or drop the claim. The ablation also shows MAR contributes only +0.14% Dice over DAR alone (82.74 to 82.88), which undercuts the argument that the three-way hierarchy is necessary rather than DAR plus LRR being the real workhorses. The HPF fusion coefficients (alpha_c, beta_c) are learnable per-class parameters optimized on training data, which is fine but should be acknowledged as a potential source of overfitting on small datasets. Hyperparameters (tau, M, loss weights, LoRA rank) are either unspecified or listed without justification. The reader's conditional verdict and moderate confidence are about right. The stress-test headline overstates things slightly — the HD95 ablation gap is genuine evidence the hierarchy helps — but the core concern about statistical significance on Synapse is correct and is the most important issue. This paper is for researchers working on SAM adaptation for medical imaging who care about representation design. It deserves a serious referee who can push back on the SOTA framing and demand significance tests. I would accept it for peer review with a major revision expectation.","headline":"Probabilistic representation learning for prompt-free SAM: solid idea, marginal gains, no significance testing","tokens_in":13443,"tokens_out":619,"would_cite":false,"duration_ms":71583,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Representation, not prompt design, limits SAM in medical segmentation","keywords":["medical image segmentation","Segment Anything Model","probabilistic representation learning","prompt-free segmentation","anatomical representation","Gaussian distributions","hierarchical prediction fusion","few-shot learning"],"falsifier":"Ablation showing that DAR alone, or DAR combined with only one other branch, matches the full model's performance within noise margins on held-out data.","tokens_in":12436,"feed_emoji":"","tokens_out":1375,"duration_ms":95138,"temperature":0.7,"pith_summary":"This paper argues that the bottleneck in prompt-free medical image segmentation with the Segment Anything Model (SAM) is not the sophistication of the prompt generator but the expressiveness of the underlying anatomical representation. Existing prompt-free SAM methods compress anatomy into deterministic prototypes — single feature vectors per organ — which the authors claim cannot jointly capture three properties: stable global anatomical priors, variation within a structure, and local reliability of evidence. The paper proposes the Hierarchical Probabilistic Representation (HPR) framework, which replaces deterministic prototypes with three probabilistic representation branches. Distributional Anatomical Representation (DAR) models each organ as a Gaussian distribution rather than a point, capturing uncertainty in organ identity. Multi-component Anatomical Representation (MAR) extends this to a mixture of Gaussians, allowing multiple sub-patterns within the same organ. Local Reliability Representation (LRR) parameterizes each image token as its own Gaussian and measures agreement with class-level reliability references, down-weighting unreliable observations from ambiguous boundaries or low-contrast regions. The three branches produce separate prediction maps that are combined by Hierarchical Prediction Fusion (HPF), a two-stage weighted combination with learnable per-organ coefficients. The entire framework remains compatible with SAM's original mask decoder. On the Synapse abdominal CT benchmark, HPR-SAM achieves 85.09% Mean Dice, surpassing prior SAM-based methods. Under few-shot settings (4 labeled scans for cardiac MRI, 3 for prostate MRI), it achieves the best scores on both the LA and PROMISE12 datasets.","feed_headline":"Representation, not prompt design, limits SAM in medical scans","feed_subtitle":"Replacing deterministic organ prototypes with hierarchical probabilistic representations lifts prompt-free segmentation accuracy to 85% Dice","key_machinery":"Three probabilistic representation branches: DAR (single Gaussian per organ for global priors), MAR (mixture of Gaussians per organ for sub-pattern diversity), LRR (per-token Gaussian with mutual-likelihood matching against class-level reliability references), combined by HPF (two-stage learnable weighted fusion of prediction maps).","core_discovery":"The central claim is a reframing of the prompt-free SAM problem: rather than designing better prompt generators, one should learn richer anatomical representations and let those representations drive prompt quality. The mechanism is probabilistic — replacing deterministic prototypes with Gaussian distributions and mixtures thereof — organized hierarchically across global, intra-structure, and local reliability scales. The ablation supports the claim that all three branches contribute, with the full model reaching 85.09% Mean Dice on Synapse versus 74.86% for the baseline without any of the three modules. DAR alone provides the largest single gain (+7.88%), while MAR and LRR add smaller but累积","pith_inferences":["The ablation shows DAR alone captures most of the gain (82.74% vs. 74.86% baseline), and the marginal contributions of MAR (+0.62% over DAR) and LRR (+1.58% over DAR+MAR) raise the question of whether a simpler single-branch probabilistic model could achieve comparable results with lower computational cost.","The framework's compatibility with SAM's original decoder means the probabilistic representations are projected into the same prompt space SAM expects — this suggests the representation improvement is orthogonal to decoder architecture and could stack with decoder-side refinements like those used by H-SAM or PG-SAM.","The LRR module's reliability estimation, if it truly identifies low-confidence regions, could be surfaced as an uncertainty map for clinical review, potentially flagging cases where automated segmentation should be verified by a radiologist."],"forward_implications":["If representation quality is the true bottleneck, then future prompt-free SAM methods should invest in representation learning rather than prompt generator architecture, potentially simplifying the design space.","The probabilistic representation approach could generalize to other foundation-model adaptation tasks beyond medical imaging, wherever deterministic prototypes are used to summarize variable visual categories.","The few-shot results (4 cardiac scans, 3 prostate scans) suggest that probabilistic representations may be particularly effective under data scarcity, where deterministic prototypes are most likely to overfit to limited examples.","The per-organ learnable fusion coefficients in HPF imply that different organs benefit from different representation branches — this could inform organ-specific adaptation strategies in clinical deployment."],"fun_headline_variants":["Probabilistic representations improve prompt-free SAM medical segmentation","From deterministic prototypes to probabilistic priors for SAM segmentation","Better representations, not better prompts, drive prompt-free SAM accuracy","HPR-SAM reframes prompt-free medical segmentation via probabilistic priors","Hierarchical probabilistic priors advance prompt-free SAM segmentation"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that all three probabilistic representation branches are necessary and complementary, but the ablation shows that DAR alone accounts for most of the improvement (from 74.86% to 82.74%), while MAR and LRR each add less than one percentage point on top of DAR. The marginal gains from the full hierarchy over simpler combinations (e.g., DAR+LRR at 84.32% vs. full model at 85.09%) leave open whether the added complexity is justified.","fun_headline_variants_meta":{"raw":{"variants":["Probabilistic representations improve prompt-free SAM medical segmentation","From deterministic prototypes to probabilistic priors for SAM segmentation","Better representations, not better prompts, drive prompt-free SAM accuracy","HPR-SAM reframes prompt-free medical segmentation via probabilistic priors","Hierarchical probabilistic priors advance prompt-free SAM segmentation"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1156,"prompt_tokens":531,"completion_tokens":625,"prompt_tokens_details":null},"tokens_in":531,"tokens_out":625,"duration_ms":15610,"temperature":1.0,"reasoning_tokens":636,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T00:40:13.093664+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Ablation showing that DAR alone, or DAR combined with only one other branch, matches the full model's performance within noise margins on held-out data.","supporting_citations":[],"review_version":1}