{"id":"5ee07edf-b7c1-4aed-bf23-cc851a091653","arxiv_id":"2507.13087","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"DiffOSeg uses a categorical diffusion model with probabilistic consensus labels and expert prompts to generate both group-consensus and expert-specific medical segmentations from one framework.","lead":"This paper presents a two-stage diffusion model for medical image segmentation that combines all raters' opinions into consensus masks and can also tailor outputs to individual experts using learned prompts. The authors report improved accuracy and diversity over prior methods on two public datasets, but the central training target is defined in a way that is difficult to reproduce.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage I's consensus label zc (Eq. 3) is not a valid categorical diffusion state as defined by Eq. (2): rows can be non-normalized counts, and the sampling distribution of W is unspecified, so the central method is under-specified.","rationale":"I agree with the reader's weakest_assumption: the consensus state zc is the load-bearing element of the paper's central claim, and as written it is not a well-defined input to categorical diffusion. The reader correctly identifies that Eq. (2) permits non-normalized weight vectors and that the distribution of W is unspecified. My stress-test confirms that even a charitable reading requires the authors to add a normalization constraint, a sampling distribution, and a row-normalization guarantee on the network output, none of which appear. This is a formal soundness problem, not a disagreement with consensus: it means the Stage I training objective in Eq. (5) may be undefined for the very target the method is designed to learn. The empirical comparisons in Tables 1 and 3 therefore cannot be interpreted as validating the proposed mechanism until this is fixed. Separately, I note that the abstract's claim of outperforming all methods on all evaluated metrics is contradicted by the paper's own Tables 1 and 2 (e.g., NPC Ds30 and NPC DA2/DA4 are below D-Persona); this is a secondary overclaim but does not change the primary concern. No formal verification is provided, and the listed repository has no commit hash, so the ambiguity cannot be resolved from the manuscript alone. The reader's REJECT verdict remains appropriate, and I see no adjustment needed.","tokens_in":8654,"tokens_out":7770,"duration_ms":89100,"concrete_test":"Analytically instantiate Eq. (2) with M=4, L=2, and a pixel where two experts label 0 and two label 1; take w=(1,1,1,1), giving zc row = (2,2). Compute π(z_t,z0) from Eq. (4) at t=1; if the resulting transition probabilities are not valid categorical probabilities (entries >1 or row sum ≠1), the stated consensus label is not a valid diffusion state and the Stage I objective is ill-posed. If the authors instead intend normalized simplex weights, re-run Table 1 with a fully specified W distribution and check whether the reported GED/D_soft numbers are reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The framework's central contribution is Stage I's probabilistic consensus target. In Eq. (3), zc = w × [z1,...,zM]^T with w drawn from S in Eq. (2). The categorical forward process q(z_t|z0) and reverse formula π(z_t,z0) in Eq. (4) are only probability distributions when each row of z0 is one-hot or, at minimum, a normalized distribution over the L labels. But Eq. (2) defines S = {w∈[0,1]^M: ||w||1 ≤ M} and the surrounding text identifies ||w||1 with the number of active experts, with 'Full Vote' at ||w||1 = M. Read literally, w=(1,...,1) is allowed, and for any pixel where experts disagree zc's row is a count vector with entries >1 and row sum M, not a probability vector. The transition probabilities in Eq. (4) then exceed 1 or fail to normalize, and the KL objective in Eq. (5) is not defined on valid categorical distributions. If the authors intended w to lie on the probability simplex and zc to be a soft label, they must state this, specify the distribution of W, and ensure \\hat p0=f_θ is row-normalized; the manuscript does none of these. This is not merely a cosmetic gap: every Stage I result in Tables 1 and 3 depends on this training target. The paper also indicates 'source code available' but gives no commit or artifact, which would otherwise help resolve the ambiguity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DiffOSeg proposes a two-stage diffusion-based framework for multi-rater medical image segmentation. Stage I constructs a 'probabilistic consensus' target by taking a weighted linear combination of expert one-hot masks and trains a categorical diffusion model to approximate the consensus distribution. Stage II injects expert-specific prompts, driven by a one-hot expert identity vector, to produce preference-aware segmentations. The method is evaluated on LIDC-IDRI and NPC-170 against several probabilistic and personalized segmentation baselines, reporting GED and Dice (soft) metrics, and claims state-of-the-art performance with 13.1M parameters, 54% fewer than D-Persona.","tokens_in":8899,"tokens_out":4483,"duration_ms":54599,"significance":"If the Stage I formulation were correctly specified, the two-stage design is a plausible and practically motivated combination of consensus modeling and personalization, and the 13.1M parameter count is a concrete efficiency advantage over existing multi-rater methods. The choice of two public datasets, the inclusion of multiple baselines, and the ablation studies are strengths. However, the central consensus target in Stage I is not a valid input to the categorical diffusion machinery as written, and the paper's headline claim of outperforming all methods on all metrics is contradicted by its own tables. These issues currently undermine confidence in the reported experimental results.","major_comments":[{"comment":"The consensus label zc = w × [z1,...,zM]^T is not a valid categorical diffusion state under the definitions given. Equation (2) allows any w in [0,1]^M with ||w||_1 ≤ M, including w=(1,...,1). For any pixel where expert masks disagree, the corresponding row of zc is then a count vector whose entries can exceed 1 and whose row sum equals M, so it is neither one-hot nor a normalized probability vector. The forward process q(z_t|z0) and reverse formula π(z_t,z0) in Eq. (4) are defined only for one-hot (or at least properly normalized) categorical states, and the KL objective in Eq. (5) is not defined for unnormalized counts. If the authors intended zc to be a soft label, they must specify the sampling distribution of W, normalize zc accordingly, and ensure that the network output p̂0 is row-normalized; none of these is stated. This issue is load-bearing because every Stage I result in Tables 1 and 3 depends on this training target.","section":"Section 2.1, Eqs. (2)-(3)"},{"comment":"The abstract claims that 'our model outperforms existing state-of-the-art methods across all evaluated metrics,' but the paper's own tables contradict this. In Table 1, on NPC-170, DiffOSeg (Stage I) reports Ds30 = 79.83, which is lower than D-Persona's 80.40. In Table 2, on NPC-170, DiffOSeg (Stage II) reports DA2 = 74.03 versus D-Persona's 74.60 and DA4 = 75.07 versus D-Persona's 75.17. The claim should be qualified to 'most metrics' or the reported numbers must be corrected.","section":"Abstract, Tables 1 and 2"},{"comment":"The text states that 'the denoising process begins with randomly sampled Gaussian noise,' but this is inconsistent with categorical diffusion, whose initial state should be sampled from a categorical distribution (e.g., uniform over the state space), not from a Gaussian. This inconsistency reinforces the concern that the categorical formulation is not implemented as described and further complicates reproducibility.","section":"Section 2.1, inference paragraph"}],"minor_comments":[{"comment":"The caption contains the typo 'stragety'; it should be 'strategy'.","section":"Figure 1 caption"},{"comment":"The caption contains the typo 'unncertainty'; it should be 'uncertainty'.","section":"Figure 2 caption"},{"comment":"The text says '9,0000 iterations' for LIDC-IDRI Stage II; this appears to be a typo for '90,000 iterations'.","section":"Section 3.1, Implementation Details"},{"comment":"References [7] and [8] are duplicates of the same Ho et al. paper, and references [14] and [15] are duplicates of the same Liao et al. paper; they should be merged or renumbered.","section":"References"},{"comment":"The paper states 'Source code is available' but provides only a repository URL with no commit hash, version tag, or artifact description; given the specification gaps in Eqs. (2)-(5), a pinned code release would be essential for verifying the method.","section":"Reproducibility"}],"recommendation":"reject","confidential_remarks":"The main issue is not experimental dishonesty but a load-bearing under-specification of the central Stage I objective. If the authors can supply a consistent definition of W, normalize the consensus target so that categorical diffusion applies, and correct the overclaim in the abstract, a resubmission could be considered. The editor may also wish to verify whether the code at the provided GitHub URL actually implements Eqs. (3)-(5) as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Useful direction, but the core Stage I consensus label is not a valid categorical diffusion state as written; the math needs to be fixed before the results are meaningful.\n\nThe paper's main idea is sensible: a two-stage diffusion model where Stage I produces a consensus distribution and Stage II personalizes to individual experts via prompts. That's a productive combination, and the parameter efficiency (13.1M vs. D-Persona's 28.46M) is a real plus. The ablations comparing random, average, and probabilistic consensus and different prompt encodings are the right kind of evidence, and the experiments cover both public datasets with multiple metrics.\n\nThe soft spot is load-bearing. In Eq. (2)-(3), z_c is defined as a weighted sum of one-hot expert masks with weights w in [0,1]^M and ||w||_1 <= M. If w=(1,...,1), then for any pixel where experts disagree, the row of z_c has entries possibly greater than 1 and row sum M, not 1. That is not a categorical distribution, yet the forward and reverse transitions in Eq. (4) and the KL objective in Eq. (5) assume z0 is one-hot or at least normalized. The distribution of W is never specified, so we cannot guess the intended normalization. This is not a cosmetic gap; it underlies every Stage I result in Tables 1 and 3. I agree with the stress-test note on this point.\n\nSecondary issues: the abstract claims \"outperforms existing state-of-the-art methods across all evaluated metrics,\" but Tables 1 and 2 show DiffOSeg below D-Persona on NPC-170 Ds30 and on DA2/DA4. No error bars are reported, and the code link has no commit hash, so the ambiguity can't be resolved by checking the artifact.\n\nBottom line: the direction is worth pursuing, and a revision that redefines the consensus label as a proper categorical distribution and specifies the weight sampling could salvage it. As written, I would not cite it, and my own review would be reject. But I'd still send it to peer review: the idea is timely, the experiments are reasonably broad, and the flaw is identifiable and fixable.\n\nRecommendation: send to review, but the reviewers should push for a corrected formulation and a toned-down abstract.","headline":"Stage I's consensus label is not a valid categorical diffusion state, so the method is under-specified despite a promising direction.","tokens_in":9464,"tokens_out":4909,"would_cite":false,"duration_ms":53331,"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":"DiffOSeg proposes a two-stage categorical-diffusion framework that generates both group-consensus and expert-specific segmentations from one model, beating prior multi-rater methods on nearly all evaluated metrics with 13.1 M parameters.","keywords":["multi-rater segmentation","medical image segmentation","annotation variability","categorical diffusion model","consensus-driven segmentation","preference-driven segmentation","expert-specific prompts","diffusion probabilistic model"],"falsifier":"Choose a pixel where two experts disagree and set the weights to $\\mathbf{w} = [0.5, 0.5]$. The consensus row becomes $[0.5, 0.5]$, which is neither a discrete label nor a normalized probability vector; substituting such a state into the categorical transition behind Eq. (4), whose posterior term relies on $\\delta_{z_0, z_t}$, leaves the KL objective in Eq. (5) undefined. A reader could check this directly in the released code, or more simply try to reproduce the reported GED and Dice$_{soft}$ numbers under any concrete choice of the distribution of $\\mathbf{w}$, since the paper fixes only the support set and never the sampling law.","tokens_in":8362,"feed_emoji":"🩻","tokens_out":17122,"duration_ms":147750,"temperature":0.7,"pith_summary":"DiffOSeg is an attempt to settle a standing dilemma in multi-rater medical image segmentation: systems so far have produced either a consensus (gold standard) segmentation or expert-specific ones, but not both from a single model. The paper argues that a two-stage categorical-diffusion framework can do both, training Stage I to sample from a probabilistically weighted blend of all experts' masks and Stage II to steer the same denoising backbone with per-expert prompts. If the claim holds, one model could give a clinician both the boundary the group of experts would agree on and the drawing any particular expert prefers, with realistic variation in both. On LIDC-IDRI and NPC-170, the paper reports better generalized energy distance and threshold-aware Dice than prior probabilistic methods in the consensus setting, and the best mean per-expert Dice in the preference setting, using about half of D-Persona's parameters.","feed_headline":"One diffusion model delivers consensus and per-expert segmentation","feed_subtitle":"Consensus and expert-preference prompts share one backbone, beating prior multi-rater methods on both datasets","key_machinery":"The load-bearing object is the two-stage categorical (multinomial) diffusion model with a probabilistic consensus start state; categorical diffusion means the state is a discrete label rather than continuous noise. Stage I's training state is the weighted combination $z_c = \\mathbf{w} \\times [z_1, \\dots, z_M]^T$, where $\\mathbf{w} \\in [0,1]^M$ with $\\|\\mathbf{w}\\|_1 \\le M$ realizes three scenarios, single vote, subgroup consensus, or full vote, and the reverse process follows the multinomial diffusion posterior of [9,29], trained by minimizing the KL divergence between the forward posterior $q(z_{t-1} \\mid z_t, z_0)$ and the predicted reverse distribution $p_\\theta(z_{t-1} \\mid z_t, I)$. Stage II's mechanism is the plug-in prompt block: global average pooling plus a linear-softmax layer generates fusion weights, a $3\\times3$ convolution modulates the prompt bank, and cascaded channel attention blocks inject the selected expert's prompt into the encoder features, so expert identity steers denoising without separate per-expert networks. Both stages share the same KL objective and a 13.1 M-parameter denoising UNet backbone.","core_discovery":"The central claim is that consensus and individual preference are not competing objectives that force a compromise; they can be assigned to two linked stages of one diffusion model. Stage I treats a weighted linear mixture of the $M$ experts' one-hot masks, $z_c = \\mathbf{w} \\times [z_1, \\dots, z_M]^T$, as the starting state of a categorical diffusion process, so repeatedly sampling the learned reverse process yields diverse masks that still respect the group's agreement. Stage II initializes from Stage I's weights and adds an expert identity, encoded as a one-hot vector plus a learnable prompt block, so the same backbone reproduces each expert's individual annotation style. On LIDC-IDRI the model improves consensus metrics over the strongest competitor (generalized energy distance, a distribution-diversity measure, drops roughly 37-45% depending on sample count) and raises mean per-expert Dice from 89.17 to 90.99; on NPC-170 it lowers GED30 from 0.2385 to 0.1822 and raises mean per-expert Dice from 76.19 to 76.88. The paper concludes that omni-view segmentation, both the group consensus and each expert's preference, is achievable in a single parameter-efficient framework.","pith_inferences":["Because expert identity is stored in a prompt rather than in separate weights, the prompt space may support interpolation: a new rater whose style sits between two known experts could plausibly be approximated by blending their prompts, a mechanism the paper leaves open when it lists few-shot adaptation as future work.","The weighted-consensus idea from Stage I is not tied to diffusion; applying the same single-vote, subgroup, or full-vote weighting as stochastic target augmentation for any multi-rater segmentation loss is a testable variant the paper does not explore.","The uncertainty maps in the paper are averages over many sample draws; a natural downstream use not discussed there is to flag regions where consensus-conditioned and expert-conditioned samples disagree, directing radiologists to the genuinely ambiguous boundaries."],"forward_implications":["A single diffusion backbone can serve both segmentation paradigms, so a clinical system no longer has to choose between a consensus model and a personalized model; a single training pipeline produces both.","The plug-in prompt block decouples expert identity from network weights, meaning adding or swapping a rater's style should cost only a prompt, not a retrained branch.","Under the paper's results, consensus fidelity (threshold-aware Dice) and diversity (generalized energy distance) improve together, which runs against the usual expectation that sampling diverse masks sacrifices accuracy.","The reported parameter count (13.1 M, about 54% below the main competitor) suggests the two-stage design stays within a modest computational budget, a practical condition for clinical deployment."],"supporting_citations":[{"why":"Supplies the multinomial (categorical) diffusion forward and reverse processes on which Eq. (4) is built.","marker":"[9]"},{"why":"Extends categorical diffusion to medical image segmentation; the reverse-transition derivation in Eq. (4) is referenced jointly to this work and [9].","marker":"[29]"},{"why":"D-Persona, the multi-rater method whose Stage I and Stage II results are the main baselines and whose protocol (including rank-simulated expert preferences) DiffOSeg follows.","marker":"[26]"},{"why":"Probabilistic U-Net, the principal probabilistic-segmentation baseline compared in Table 1.","marker":"[11]"},{"why":"PhiSeg, a stochastic segmentation baseline reimplemented by the authors for the Table 1 comparison.","marker":"[2]"},{"why":"Source of the LIDC-IDRI dataset, the four-expert CT benchmark used in the main experiments.","marker":"[1]"},{"why":"Source of the NPC-170 dataset, the four-radiologist MRI benchmark used in the main experiments.","marker":"[27]"},{"why":"CM-Pixel baseline for preference-driven segmentation and the protocol for ranking the four LIDC-IDRI annotators into simulated preference levels.","marker":"[31]"},{"why":"Pionono, a model of inter- and intra-observer variability that sets the per-expert Dice baseline in Table 2.","marker":"[22]"},{"why":"Supplies the cascaded Channel Attention Blocks used inside the plug-in prompt block's feature-modulation step.","marker":"[28]"}],"fun_headline_variants":["DiffOSeg: one diffusion model for consensus and per-expert segmentation","Diffusion model merges group consensus with each expert's style","Two-stage diffusion yields consensus and per-expert masks","Omni-view segmentation: consensus and per-expert in one diffusion model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the weighted average of the experts' one-hot masks is a legitimate starting state for the categorical diffusion process, even though those averaged rows are generally neither one-hot nor normalized to sum to one, and the paper never specifies the probability distribution from which the weight vector $\\mathbf{w}$ is sampled.","fun_headline_variants_meta":{"raw":{"variants":["DiffOSeg: one diffusion model for consensus and per-expert segmentation","Diffusion model merges group consensus with each expert's style","Two-stage diffusion yields consensus and per-expert masks","Omni-view segmentation: consensus and per-expert in one diffusion model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001107,"raw_usage":{"total_tokens":4632,"prompt_tokens":979,"completion_tokens":3653,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":3580}},"tokens_in":595,"tokens_out":3653,"duration_ms":28989,"temperature":1.0,"reasoning_tokens":3580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:30:39.626562+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a pixel where two experts disagree and set the weights to $\\mathbf{w} = [0.5, 0.5]$. The consensus row becomes $[0.5, 0.5]$, which is neither a discrete label nor a normalized probability vector; substituting such a state into the categorical transition behind Eq. (4), whose posterior term relies on $\\delta_{z_0, z_t}$, leaves the KL objective in Eq. (5) undefined. A reader could check this directly in the released code, or more simply try to reproduce the reported GED and Dice$_{soft}$ numbers under any concrete choice of the distribution of $\\mathbf{w}$, since the paper fixes only the support set and never the sampling law.","supporting_citations":[{"cited_title":"Advances in neural in- formation processing systems34, 12454–12465 (2021)","cited_arxiv_id":null,"evidence_quote":"Supplies the multinomial (categorical) diffusion forward and reverse processes on which Eq. (4) is built."},{"cited_title":"In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Extends categorical diffusion to medical image segmentation; the reverse-transition derivation in Eq. (4) is referenced jointly to this work and [9]."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"D-Persona, the multi-rater method whose Stage I and Stage II results are the main baselines and whose protocol (including rank-simulated expert preferences) DiffOSeg follows."},{"cited_title":"In: Proceedings of the 32nd ACM International Conference on Multimedia","cited_arxiv_id":null,"evidence_quote":"Source of the NPC-170 dataset, the four-radiologist MRI benchmark used in the main experiments."},{"cited_title":"In: Proceedings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Pionono, a model of inter- and intra-observer variability that sets the per-expert Dice baseline in Table 2."},{"cited_title":"In: International Workshop on Statistical Atlases and Computational Models of the Heart","cited_arxiv_id":null,"evidence_quote":"Supplies the cascaded Channel Attention Blocks used inside the plug-in prompt block's feature-modulation step."}],"review_version":1}