{"id":"2a4a2d07-338c-4711-9465-f7a22704e931","arxiv_id":"2506.14181","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Meta-SurDiff combines a classification diffusion model with meta-learned sample weighting and reports state-of-the-art results on five surgical video datasets, but the derivation of the reverse process contains a numerical inconsistency.","lead":"A team applies a diffusion model plus meta-learning to surgical videos, claiming state-of-the-art online phase recognition and uncertainty estimates on five datasets. The paper's core math has an error in the reverse-sampling formula, so the method as written may not match what was implemented.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reverse posterior coefficient γ2 in Eq. 9 does not match the forward process in Eq. 5–6, so Algorithm 2 does not sample the claimed diffusion posterior.","rationale":"The paper's central contribution is a 'classification diffusion model' that converts coarse phase representations into a precise frame-level phase distribution. That claim stands or falls on the reverse sampler in Algorithm 2. Checking the Gaussian algebra in Section 3.1: γ0 and γ1 in Eq. 9 match the standard conditional posterior of the forward process in Eq. 5–6, but γ2 must be 1 − γ0 − γ1. The printed γ2 is different; the t = 2 numerical example shows a discrepancy of about 0.30 in the coefficient, i.e. a roughly hundredfold error in this example. This is not a stylistic disagreement or a missing baseline: the method as specified is not the method being evaluated. The reader's REJECT verdict is therefore well supported. The other listed weaknesses—no released code, absent error bars on OphNet and NurViD, and sparse comparisons on those datasets—reinforce the rejection but are secondary. I leave the verdict unchanged.","tokens_in":16935,"tokens_out":9963,"duration_ms":92495,"concrete_test":"Re-derive the posterior from Eq. 5–6 and compare with Eq. 9 for a scalar case, e.g. β1 = 0.1, β2 = 0.2. Compute the closed-form posterior mean coefficients γ0 = β2√α1/(1 − α1α2), γ1 = √α2(1 − α1)/(1 − α1α2), γ2 = 1 − γ0 − γ1, then evaluate Eq. 9's γ2. If γ2 differs by about 0.30 for this schedule, Algorithm 2 does not implement the intended diffusion posterior. Optionally, implement Algorithm 2 and a Monte Carlo sample of the exact posterior for this scalar case and compare the resulting y1 marginals; any significant divergence confirms the mathematical objection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation 9's γ2 is inconsistent with the forward process of Eq. 5–6. Let a_t = √α_t and b_{t-1} = √̅α_{t-1}, with ̅α_t the cumulative product. From the Gaussian product in Eq. 7, the posterior mean must be γ0 y0 + γ1 y_t + γ2 z with γ0 = β_t b_{t-1}/(1-̅α_t), γ1 = a_t(1-̅α_{t-1})/(1-̅α_t), and γ2 = 1 − γ0 − γ1, because all three conditional means have affine interpolation form and the coefficients must sum to 1. The printed γ2 = 1 + ((√α_t − 1)(√α_t + √α_{t-1}))/(1 − ̅α_t) does not equal this. Concrete scalar check: β1 = 0.1, β2 = 0.2 gives α1 = 0.9, α2 = 0.8, ̅α2 = 0.72. The correct coefficients are γ0 ≈ 0.678, γ1 ≈ 0.319, γ2 ≈ 0.003, while the paper's γ2 ≈ 0.305. Thus Algorithm 2's update y_{t-1} = γ0 ŷ0 + γ1 y_t + γ2 z + √(γ3 β_t) ε injects a large z-dependent shift and is not sampling from the posterior of the stated forward process. Since the paper's central claim is precise frame-level distribution estimation and calibrated uncertainty, the mathematical grounding of the method fails as written; the empirical SOTA numbers cannot be attributed to the described diffusion model unless the sampler is corrected and the code is released.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Meta-SurDiff, a classification diffusion model conditioned on coarse phase representations from a ConvNext+LSTM backbone, trained with a meta-learning reweighting objective intended to handle frame ambiguity and unbalanced phase distributions for online surgical phase recognition. The method is evaluated on Cholec80, AutoLaparo, M2Cai16, OphNet, and NurViD with accuracy, precision, recall, Jaccard, and uncertainty metrics such as prediction interval width and a paired two-sample t-test. The abstract and conclusion claim new state-of-the-art performance and reliable frame-level uncertainty estimation.","tokens_in":1523,"tokens_out":3438,"duration_ms":132857,"significance":"If correct, the paper would provide a useful plug-in module for uncertainty-aware online surgical phase recognition, and the idea of conditioning the diffusion endpoint on a coarse prediction is reasonable. The meta-reweighting component largely follows Meta-Weight-Net, but the combination with a classification diffusion model is nonstandard. The breadth of datasets and the use of uncertainty metrics are positive features. However, the paper does not release code, and the core arithmetic of the diffusion posterior is incorrect, so the reported results cannot be attributed to the described method; in addition, the evaluation on two of the five datasets is not against current state-of-the-art baselines.","major_comments":[{"comment":"The coefficient gamma_2 in the reverse posterior mean is inconsistent with the forward process in Eqs. (5)-(6). With a = sqrt(alpha_t) and b = sqrt(alpha_{t-1}), where alpha_t denotes the cumulative product, the correct coefficient of z in the posterior mean is gamma_2 = (1-a)(1-b)/(1+ab), not the printed 1 + ((sqrt(alpha_t)-1)(sqrt(alpha_t)+sqrt(alpha_{t-1})))/(1 - alpha_t). A concrete check with beta_1=0.1 and beta_2=0.2 gives correct gamma_2 approximately 0.003, whereas Eq. (9) yields approximately 0.305. Because Algorithm 2 uses this gamma_2 in the update y_{t-1} = gamma_0 y_hat_0 + gamma_1 y_t + gamma_2 z + sqrt(gamma_3 beta_t) epsilon, the sampler as written does not draw from the posterior of the stated forward process.","section":"Section 3.1, Eq. (9) and Algorithm 2"},{"comment":"The denoising reconstruction step y_hat_0 = (1 / alpha_t) (y_t - (1 - sqrt(alpha_t)) z - sqrt(1 - alpha_t) epsilon_theta) is inconsistent with the forward marginal in Eq. (6). From Eq. (6), the correct inversion is y_hat_0 = (y_t - (1 - sqrt(alpha_t)) z - sqrt(1 - alpha_t) epsilon_theta) / sqrt(alpha_t), i.e., division by sqrt(alpha_t) rather than by alpha_t. As written, this line of Algorithm 2 is arithmetically incorrect and compounds the error in the gamma_2 coefficient.","section":"Algorithm 2"},{"comment":"The claim of state-of-the-art performance on OphNet and NurViD is not supported by the comparisons shown. Table 5 compares only against X-CLIP16 and X-CLIP32, and Table 6 only against SlowFast, C3D, and I3D, which are not current surgical-phase-recognition baselines for these benchmarks. Additionally, Table 6 reports 'All(162)' while Table 1 states that NurViD has 177 phases, and Appendix A says OphNet has 102 phases while Table 1 says 96; these inconsistencies need to be resolved and the results compared with the benchmarks' own published methods and metrics.","section":"Section 4.2.1, Tables 5 and 6"}],"minor_comments":[{"comment":"The expression for B appears to contain a typo: the coefficient multiplying y_t should be sqrt(alpha_t)/beta_t, not sqrt(alpha_{t-1})/beta_t, based on the standard Gaussian product and the subsequent gamma_1 in Eq. (9).","section":"Section 3.1, Eq. (7)"},{"comment":"The symbol alpha_t is used both for the per-step noise schedule (alpha_t = 1 - beta_t) and for the cumulative product, making the derivation hard to check; please use separate notation such as alpha_t and alpha_bar_t consistently.","section":"Section 3.1, Eqs. (5)-(9)"},{"comment":"Table 1 lists 'MeCai16' instead of 'M2Cai16', and the OphNet duration '5min37s' conflicts with Appendix A's description of 284.8 hours for the full dataset and 32 seconds for trimmed videos; please reconcile these values.","section":"Table 1 and Appendix A"},{"comment":"The notation 'at wt' for the meta-loss gradient is unclear, since the meta-loss is evaluated at the updated classifier parameters, not at wt; please clarify the point of evaluation in the gradient computation.","section":"Section 3.2, Eq. (12)"},{"comment":"The text says 'we select the diffusion timestep to be T=1000' and later reduces to T=100 with DDIM, while Appendix B says 'we set the number of timesteps as T=100'; please unify these statements to distinguish training diffusion steps from inference sampling steps.","section":"Section 4.2.1 and Appendix B"},{"comment":"The description of the paired two-sample t-test states that the null hypothesis is 'the top two maximum predicted values are the same,' which is not a standard t-test formulation; please define the paired differences and the null hypothesis explicitly.","section":"Appendix C, PTST"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft: the corresponding-author email is the literal placeholder 'email@example.com', and there are several unresolved inconsistencies (e.g., 'MeCai16', 177 vs. 162 phases for NurViD, and conflicting OphNet duration numbers). More importantly, the incorrect posterior coefficient in Eq. (9) and the incorrect denoising division in Algorithm 2 change the actual sampling procedure, so the reported empirical numbers cannot be attributed to the method as described without code release and rerun experiments. I therefore recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The gist: this is a sensible plug-in idea - a CARD-style classification diffusion refiner conditioned on coarse phase logits, trained with a Meta-Weight-Net reweighting objective - tested on five surgical video datasets. That combination is new, and the uncertainty evaluation (PIW plus paired t-tests) is a practical step for a safety-relevant domain. The empirical gains over ConvNext+LSTM baselines are plausible and the ablation study lines up.\n\nBut there is a load-bearing math problem. The stress-test note is correct: Eq. 8-9 give gamma0 and gamma1 that match the standard DDPM posterior, but gamma2 is not 1-gamma0-gamma1. For beta1=0.1, beta2=0.2, the correct gamma2 is about 0.003, while the paper's formula gives something like 0.03 or 0.3 depending on how you read the ambiguous alpha notation. Since Algorithm 2 literally uses that gamma2 to inject a z-dependent shift, the sampler does not draw from the posterior of the forward process defined in Eq. 5-6. This is not a cosmetic typo: the paper's central claim is precise frame-level distribution estimation and calibrated uncertainty, and the described generative process does not deliver that. Without the corrected sampler or released code, the SOTA numbers cannot be attributed to the claimed diffusion model.\n\nEverything else is softer. No code is public, OphNet and NurViD tables lack error bars, and the baseline coverage on those two datasets is thin. The meta-reweighting is not circular - the balanced meta set is an internal validation, fine - and the citations to CARD and Meta-Weight-Net are appropriate.\n\nVerdict: this deserves a serious referee, not a desk reject. The application is real, the components are known-good, and the math error, while central, looks fixable - provided the authors actually derive the right gamma2 and re-run the experiments. I would send it to review with instructions to demand a corrected posterior, the code, and complete experimental tables before anything is accepted. As written, though, the core uncertainty story does not hold up.","headline":"Useful empirical combination with a real math error in the reverse posterior; the sampler as written does not match the forward process, so the uncertainty claims need a corrected derivation and code before the SOTA results can be trusted.","tokens_in":671,"tokens_out":789,"would_cite":false,"duration_ms":80228,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Meta-SurDiff, a meta-learning-optimized classification diffusion model, achieves state-of-the-art online surgical phase recognition on five datasets.","keywords":["online surgical phase recognition","classification diffusion model","meta-learning reweighting","frame-level uncertainty estimation","imbalanced phase distribution","surgical video analysis","prediction interval width"],"falsifier":"Evaluate Eqs. 7-9 numerically: for a fixed small timestep such as $t=1$, take $\\mathbf{z}_i=0$ and $\\beta_1=10^{-4}$, compute the posterior mean $q(\\mathbf{y}_{t-1} \\mid \\mathbf{y}_t, \\mathbf{y}_0, \\mathbf{z}_i)$ directly from the forward transition in Eq. 5 and the marginal in Eq. 6, and compare it with Eq. 8. A mismatch in the $\\mathbf{z}_i$ coefficient would show that Algorithm 2 does not sample the posterior of the paper's own forward process, and the corrected update could then be tested to see whether the reported gains survive.","tokens_in":16731,"feed_emoji":"🩺","tokens_out":12649,"duration_ms":119921,"temperature":0.7,"pith_summary":"Online surgical phase recognition faces two reliability problems that most models ignore: individual frames can be visually ambiguous, and surgical phases are highly unbalanced in how often they occur. The paper's thesis is that both problems are best treated by replacing point predictions with frame-level distribution estimation. Meta-SurDiff feeds coarse phase scores from any video backbone into a classification diffusion model whose reverse process produces a per-frame predictive distribution, and it trains that diffusion model with a meta-learned per-frame reweighting objective. On five surgical-video datasets, the paper reports the best results among the compared online methods on accuracy, precision, recall, and Jaccard metrics. It also reports that correct predictions come with much narrower prediction intervals than incorrect ones, a step toward letting a recognition system flag uncertain frames for human review.","feed_headline":"Diffusion model tops five surgical phase-recognition benchmarks","feed_subtitle":"It also estimates frame-level uncertainty, so ambiguous moments can be flagged for human review.","key_machinery":"Two mechanisms carry the argument. The first is the classification diffusion model, a conditional DDPM whose forward process interpolates between the label embedding and a coarse representation, $q(\\mathbf{y}_t \\mid \\mathbf{y}_0, \\mathbf{z}_i) = \\mathcal{N}(\\sqrt{\\bar\\alpha_t}\\,\\mathbf{y}_0 + (1-\\sqrt{\\bar\\alpha_t})\\,\\mathbf{z}_i, (1-\\bar\\alpha_t)I)$, with the reverse update $\\mathbf{y}_{t-1} = \\gamma_0 \\hat{\\mathbf{y}}_0 + \\gamma_1 \\mathbf{y}_t + \\gamma_2 \\mathbf{z}_i + \\sqrt{\\gamma_3 \\beta_t}\\,\\boldsymbol{\\epsilon}$ whose coefficients come from the posterior derivation in Eq. 9. Repeating that update over $T$ steps turns a coarse backbone score into a sampled phase distribution. The second mechanism is the meta-learning reweighting objective: a two-layer MLP meta-weight net computes a weight for each frame's loss and is updated by alternating meta-training and meta-testing steps (Eqs. 12-13), so the diffusion model is not dominated by majority phases. Together these two mechanisms are what the paper credits for the accuracy and uncertainty-calibration gains.","core_discovery":"The paper's central discovery is that a diffusion model can be repurposed from a generator into a calibrator for online phase recognition. Standard DDPM diffuses data to Gaussian noise and learns to reverse that corruption; Meta-SurDiff instead diffuses the ground-truth phase label embedding $\\mathbf{y}_0$ toward a coarse phase representation $\\mathbf{z}_i$ produced by a backbone, $q(\\mathbf{y}_t \\mid \\mathbf{y}_{t-1}, \\mathbf{z}_i) = \\mathcal{N}(\\sqrt{1-\\beta_t}\\,\\mathbf{y}_{t-1} + (1-\\sqrt{1-\\beta_t})\\,\\mathbf{z}_i, \\beta_t I)$, so the reverse process must recover a sharp label embedding from a noisy estimate of it. The learned reverse model generates a full predictive distribution over phases for each frame, and a meta-weight network assigns larger training weights to frames from minority phases. The paper reports new best accuracy/precision/recall/Jaccard results on Cholec80, AutoLaparo, M2Cai16, OphNet, and NurViD, and reports that correct predictions show much narrower prediction intervals than incorrect ones.","pith_inferences":["The same pattern—diffuse the label toward a coarse streaming prediction, reverse-sample to get a distribution, then reweight with a balanced meta-set—is not surgery-specific and could transfer to online action segmentation or real-time procedure monitoring in non-medical video.","The PIW/t-test split suggests a concrete deployment rule the paper leaves implicit: accept a prediction when the top-two probability gap is statistically significant and the interval is narrow, and flag the rest for human review.","An ablation that swaps the diffusion posterior for a simpler noise-calibrated sampler would isolate how much of the reported gain comes from the generative refinement versus the meta-reweighting objective and backbone fine-tuning."],"forward_implications":["Attaching the classification diffusion module to a different backbone (ConvNeXt+LSTM, ResNet, ViT, or X-CLIP) improves that backbone's online phase recognition, so the method works as a plugin rather than a from-scratch architecture.","The uncertainty estimates give an operational quality signal: instances whose paired t-test rejects the equality of the top two predicted probabilities are recognized more accurately, so a system could automatically accept those and route the rest to an expert.","Because the temporal encoder is an LSTM and the diffusion conditioner uses only current and past representations, the method preserves the online constraint and does not need future frames.","With DDIM acceleration to 100 diffusion steps, inference runs at 76 ms per frame with 21.44M parameters on Cholec80, which is compatible with real-time surgical monitoring."],"supporting_citations":[{"why":"Supplies the DDPM forward/reverse process and noise-prediction training objective that the classification diffusion model adapts.","marker":"[18]"},{"why":"Supplies the conditioning scheme that diffuses toward a target embedding given an input, the direct antecedent of the classification diffusion model.","marker":"[16]"},{"why":"Supplies the meta-weight network that learns per-example weights, the basis of the reweighting objective.","marker":"[39]"},{"why":"Supplies the re-weighting meta-learning objective for imbalanced classification used to balance surgical phases.","marker":"[15]"},{"why":"Supplies the ConvNeXt+LSTM backbone used to produce the coarse phase representations and the batch-normalization discussion that motivates it.","marker":"[35]"},{"why":"Defines the online surgical phase recognition task and the Cholec80 training/testing protocol used in evaluation.","marker":"[21]"},{"why":"Supplies the DDIM accelerated sampling that reduces inference from 1000 to 100 diffusion steps for real-time use.","marker":"[40]"},{"why":"Provides the Cholec80 dataset used for the main online phase recognition experiments and the uncertainty analysis.","marker":"[45]"},{"why":"Provides the OphNet benchmark and its evaluation protocol, one of the five datasets used for the state-of-the-art claim.","marker":"[20]"},{"why":"Provides the NurViD benchmark and its phase/action evaluation protocol, one of the five datasets used for the state-of-the-art claim.","marker":"[19]"}],"fun_headline_variants":["Meta-learning diffusion model fixes surgical phase recognition","Diffusion plus meta-learning nails surgical phase benchmarks","Diffusion model reimagined for reliable surgical phase recognition","Uncertainty-aware diffusion model wins on five surgical datasets","Meta-SurDiff slashes ambiguity in surgical phase AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the algebra that converts the forward interpolation process into the closed-form reverse update used at test time; if that posterior derivation is incorrect, Algorithm 2's sampler does not implement the model the paper intends.","fun_headline_variants_meta":{"raw":{"variants":["Meta-learning diffusion model fixes surgical phase recognition","Diffusion plus meta-learning nails surgical phase benchmarks","Diffusion model reimagined for reliable surgical phase recognition","Uncertainty-aware diffusion model wins on five surgical datasets","Meta-SurDiff slashes ambiguity in surgical phase AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000279,"raw_usage":{"total_tokens":1716,"prompt_tokens":1060,"completion_tokens":656,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":580}},"tokens_in":676,"tokens_out":656,"duration_ms":6179,"temperature":1.0,"reasoning_tokens":580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:19:02.091285+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate Eqs. 7-9 numerically: for a fixed small timestep such as $t=1$, take $\\mathbf{z}_i=0$ and $\\beta_1=10^{-4}$, compute the posterior mean $q(\\mathbf{y}_{t-1} \\mid \\mathbf{y}_t, \\mathbf{y}_0, \\mathbf{z}_i)$ directly from the forward transition in Eq. 5 and the marginal in Eq. 6, and compare it with Eq. 8. A mismatch in the $\\mathbf{z}_i$ coefficient would show that Algorithm 2 does not sample the posterior of the paper's own forward process, and the corrected update could then be tested to see whether the reported gains survive.","supporting_citations":[{"cited_title":"Card: Classification and regression diffusion models.Advances in Neural Information Processing Systems, 35:18100–18115, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the conditioning scheme that diffuses toward a target embedding given an input, the direct antecedent of the classification diffusion model."},{"cited_title":"Meta-weight-net: Learning an explicit mapping for sample weighting.Advances in neural information processing systems, 32, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the meta-weight network that learns per-example weights, the basis of the reweighting objective."},{"cited_title":"Learning to re- weight examples with optimal transport for imbalanced classification.Advances in Neural Information Processing Systems, 35:25517–25530, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the re-weighting meta-learning objective for imbalanced classification used to balance surgical phases."},{"cited_title":"On the pitfalls of batch normalization for end-to-end video learning: A study on surgical workflow analysis.Medical Image Analysis, page 103126, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the ConvNeXt+LSTM backbone used to produce the coarse phase representations and the batch-normalization discussion that motivates it."},{"cited_title":"Endonet: a deep architecture for recognition tasks on laparoscopic videos.IEEE transactions on medical imaging, 36(1):86–97, 2016","cited_arxiv_id":null,"evidence_quote":"Provides the Cholec80 dataset used for the main online phase recognition experiments and the uncertainty analysis."},{"cited_title":"OphNet: A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding","cited_arxiv_id":"2406.07471","evidence_quote":"Provides the OphNet benchmark and its evaluation protocol, one of the five datasets used for the state-of-the-art claim."}],"review_version":1}