{"id":"fecbd873-3d9c-4f2a-b746-66f43b2cfb91","arxiv_id":"2506.01471","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A teacher-student video transformer with temporal consistency pseudo-labeling and prototype contrastive learning matches full-supervision surgical phase recognition using only 1/4 of Cholec80 labels.","lead":"Surgical phase recognition normally needs frame-by-frame expert annotations, which are costly. This paper shows a semi-supervised video transformer that, with a quarter of the labeled data, matches or beats full supervision on the public Cholec80 benchmark and improves accuracy on a private esophagectomy dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Cholec80 SOTA comparison is not yet supported: labeled subsets are drawn at random with no seed and no repeated sampling, so the 2–3 point gaps over SurgSSL could be split artifacts rather than method gains.","rationale":"The reader's weakest assumption is the load-bearing one. The method itself is plausible and internally consistent: the ablations in Table 1 show monotonic gains from temporal consistency regularization and prototype contrastive learning, and the code is released. However, the headline claims of SOTA on RAMIE and strong results on Cholec80 depend on comparisons to published numbers that are not matched to the paper's randomly chosen labeled subsets. Without seeds, repeated sampling, or split-level statistics, the reported gaps could be explained by which videos were selected, especially at 5 and 10 labeled videos. This is an addressable experimental-control issue, not a fundamental flaw in the architecture, so the appropriate verdict remains CONDITIONAL rather than ACCEPT or REJECT. The RAMIE result is additionally limited by the private dataset and placeholder citation, but the Cholec80 comparison is the more concrete and testable weakness.","tokens_in":8465,"tokens_out":7953,"duration_ms":90118,"concrete_test":"Re-run the Cholec80 experiments using the exact labeled splits from SurgSSL's original paper for 5, 10, and 20 labeled videos, or if those splits are unavailable, re-run both SurgSSL and SemiVT-Surge on 10 random labeled subsets per condition with fixed seeds and report mean±std accuracy across subsets. If the Ours-minus-SurgSSL gap shrinks to within one standard deviation of the subset distribution, the SOTA claim weakens; if the gap persists across all draws, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim's public half rests on Table 3, where labeled Cholec80 subsets are described only as 'randomly sample a subset of the training set.' No seed, no repeated sampling, and no split-level variance are reported. With only 5, 10, or 20 labeled videos, the specific videos chosen can dominate the result. The reported Ours-versus-SurgSSL differences (3.4, 3.8, and 2.1 accuracy points) are comparable to or smaller than the across-test-video standard deviations (7.0–7.7) in the same table, so the gap may fall within subset-selection noise. SurgSSL's numbers in Table 3 appear to be taken from the original paper, which likely used a different split protocol; without matching the exact labeled splits, the comparison is not controlled. The same issue affects the RAMIE results, where 20 unlabeled videos are 'randomly selected' from a pool of 70 and the dataset itself is private with a placeholder citation [2]. Thus the headline improvements on both datasets are not yet statistically grounded.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SemiVT-Surge, a semi-supervised video transformer for surgical phase recognition. The method uses a TimeSformer encoder in a student-teacher EMA framework; unlabeled clips are weakly and strongly augmented through short-term consecutive versus long-term random frame sampling, and a confidence-gated temporal consistency loss is applied. A prototype-based triplet loss, using EMA-updated class prototypes and teacher pseudo-labels, is added to structure the embedding space. Experiments on the private RAMIE dataset and the public Cholec80 dataset report accuracy gains over TeCNO, FixMatch, SurgSSL, and self-supervised pretraining baselines. The headline claims are state-of-the-art accuracy on RAMIE (83.4%) and accuracy exceeding the fully supervised TeCNO baseline with only 20 of 40 Cholec80 labels (90.4% versus 88.6%).","tokens_in":8709,"tokens_out":6814,"duration_ms":72180,"significance":"If the results are reproducible and the comparisons are well controlled, the paper would be a useful contribution to the under-explored area of semi-supervised surgical phase recognition. It combines several known semi-supervised ideas, including temporal sampling as augmentation, EMA teacher-student consistency, and prototype-based contrastive learning, in a way that is tailored to long surgical videos. The ablations in Table 1 are directionally consistent on both datasets, and the authors release code, which supports reproducibility. However, the central claims currently rest on comparisons whose labeled splits are not matched, no significance testing or repeated-split variance is reported, and one of the two evaluation datasets is private with a placeholder citation. These issues are addressable but must be fixed before the stated conclusions can be considered supported.","major_comments":[{"comment":"The comparison against SurgSSL on Cholec80 is not controlled. The paper states that for reduced-annotation experiments 'we randomly sample a subset of the training set as labeled data' but reports no random seed, no repeated sampling, and no split-level variance. With only 5, 10, or 20 labeled videos, the identity of the chosen videos can dominate performance, and the reported gaps over SurgSSL (2.1–3.4 accuracy points) may be within subset-selection noise. The standard deviations in Table 3 are computed across test videos, not across repeated labeled-subset draws, so they do not address this concern. In addition, the SurgSSL numbers appear to be quoted from the original paper, which likely uses a different split protocol. To support the state-of-the-art claim, the authors should either adopt the exact labeled splits used by SurgSSL or report mean and standard deviation over several random splits with fixed seeds, and should state whether the SurgSSL baselines were re-run under the same protocol.","section":"§3, Table 3"},{"comment":"The headline 4.9% accuracy improvement on RAMIE is not statistically supported. The test set contains only 9 videos, and no significance test or confidence interval is reported for the difference between the proposed method (83.4%) and the fully supervised TeCNO baseline (78.5%). The RAMIE dataset is also private, and reference [2] is a placeholder ('This paper is accepted by XXX but not yet available to the public'), so the result cannot be checked by reviewers or readers. Please provide a significance test (or a bootstrap confidence interval for the difference), and either make the dataset available or supply a detailed, de-identified dataset description and a complete citable reference.","section":"§3, Table 2"},{"comment":"The 'comparable to full supervision' claim is supported only against TeCNO, which is a different architecture. Table 3 compares Ours with 20 labeled videos (90.4%) to fully supervised TeCNO (88.6%), but the authors do not report the accuracy of their own TimeSformer-based model trained with all 40 labels. The apparent advantage may come from the video transformer backbone rather than from the semi-supervised framework. Please add a fully supervised upper bound of the same model trained with all labeled data, and phrase the abstract's claim in terms of that comparison.","section":"Abstract and §3, Table 3"},{"comment":"The ablations on Cholec80 are reported from a single run with a single randomly drawn labeled subset. The incremental gains between adjacent rows (e.g., 84.3±10.7 to 87.0±6.6 to 89.6±7.0 to 90.4±7.0) are small relative to the across-video standard deviations, and the text describes each addition as an improvement without repeated-trial or statistical support. At minimum, report the mean and standard deviation over multiple labeled-subset draws for the Cholec80 ablations.","section":"§3, Table 1"}],"minor_comments":[{"comment":"The unsupervised consistency loss is called L_U in Equation (1) but LReg in Algorithm 1 and in the text of Section 2.2; please use a single symbol throughout.","section":"§2.2 and Algorithm 1"},{"comment":"The word 'emperically' should be 'empirically'.","section":"§2.4"},{"comment":"Weak augmentation is described as 'random cropping, normalization, random rotation, and random crop', with 'random crop' appearing twice; also, the strong augmentation configuration 'rand-m9-n5-mstd0.8-inc1' is a RandAugment policy, not an AutoAugment policy. Please clarify the exact augmentation library and policy.","section":"§2.4"},{"comment":"The strong augmentation randomly selects T-1 frames from the full video history; since frames are extracted at 1 fps, this can span the entire surgery. Please state whether the selection is uniform over the whole history or constrained to a local temporal window, as this directly affects the interpretation of 'temporal consistency'.","section":"§2.4 and Algorithm 1"},{"comment":"The symbols L_Tri-L and L_Tri-U are used in the figure without being defined there; please refer explicitly to Equation (2) and the corresponding text in Section 2.3.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The placeholder reference [2] for the RAMIE dataset is a serious obstacle to external verification; one of the two evaluation pillars of the paper is currently inaccessible. Please also ask the authors to make their Cholec80 labeled-split seeds and repeated-split results available at revision time so that the comparison with SurgSSL can be checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate engineering contribution with a real statistical hole in the public half of its headline claim. If I were refereeing, I'd ask for controlled split comparisons before accepting.\n\nWhat's new: not the components—FixMatch consistency, EMA teacher, prototype contrastive learning are all published. The adaptation to surgical videos is the contribution: short-term vs long-term temporal sampling as weak/strong augmentation, and using the teacher's high-confidence pseudo-labels to update class prototypes. That is sensible and, on the evidence in the paper, it helps. The ablations are directionally consistent across RAMIE and Cholec80, and code is released. The RAMIE result—a semi-supervised video transformer beating a supervised TeCNO baseline by 4.9 points on 14 labeled plus 20 unlabeled videos—is the strongest empirical claim in the paper.\n\nSoft spots, in order. The Cholec80 comparison against SurgSSL in Table 3 is not controlled: labeled subsets are described only as \"randomly sample a subset,\" with no seed, no repeats, and no split-level variance. With 5–20 labeled videos, the choice of which videos are labeled can easily account for the 2–3 point gaps, especially since the per-video test standard deviations are 6–8 points. The authors also appear to use SurgSSL's published numbers, which may come from a different sampling protocol; the paper doesn't say. That's a load-bearing weakness for the claim \"comparable to full supervision with 1/4 labels,\" not because the method can't do that, but because this experiment doesn't yet show it. The RAMIE side has the opposite problem: the dataset is private, the citation is a placeholder [2], and the test set has only 9 videos. The 4.9-point gain over TeCNO is more plausible since it's within the same split, but the lack of significance testing matters with n=9.\n\nMinor: hyperparameters (threshold, margin, k, EMA) are dataset-specific and empirical; not a flaw, just no sensitivity analysis. The paper's self-citations are fine—they are prior work by the same group, cited appropriately.\n\nBottom line: the method is sound enough to deserve referee time, and the flaws are fixable. A revised version with seeded repeated splits, a proper match to SurgSSL's protocol, and significance testing would make the empirical claims much more credible. I'd send it to review and ask for that revision. For people working on surgical video understanding and label-efficient methods, this is a useful data point despite the statistical issues.","headline":"A sensible FixMatch-style extension to surgical video that likely works, but the headline Cholec80 SOTA comparison rests on uncontrolled random splits and a private dataset, so I'd want a revision before trusting the numbers.","tokens_in":9264,"tokens_out":2694,"would_cite":true,"duration_ms":29119,"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":"Semi-supervised surgical phase recognition can match fully supervised accuracy with a quarter of the annotations by using temporal consistency and class prototypes to leverage unlabeled video.","keywords":["surgical phase recognition","semi-supervised learning","video transformer","pseudo-labeling","temporal consistency regularization","contrastive learning","class prototypes","surgical video analysis"],"falsifier":"Re-run the Cholec80 comparison at 20, 10, and 5 labeled videos over at least ten random subset draws with fixed seeds and report mean and standard deviation; if the reported gaps over SurgSSL (90.4 vs 87.0, 88.8 vs 85.0, 85.3 vs 83.2) do not persist across most draws, the label-efficiency claim collapses to a dataset-luck artifact.","tokens_in":8293,"feed_emoji":"🏥","tokens_out":9808,"duration_ms":101775,"temperature":0.7,"pith_summary":"SemiVT-Surge is a semi-supervised video transformer for surgical phase recognition that claims to make unlabeled surgical video nearly as useful as labeled video. The paper argues that temporal consistency between weakly and strongly sampled clips, enforced through an exponential-moving-average teacher-student pair, plus contrastive learning against class prototypes, lets the model exceed fully supervised accuracy on a private RAMIE dataset by 4.9 percentage points. On the public Cholec80 dataset, the same method with only 20 labeled videos outperforms the previous semi-supervised baseline and matches or slightly exceeds fully supervised training that uses all 40 labeled videos. The motivation is practical: phase annotations require frame-by-frame review of long videos, so any method that cuts annotation cost while keeping accuracy matters for computer-assisted surgery.","feed_headline":"Video transformer uses unlabeled surgery footage to beat full labels","feed_subtitle":"Semi-supervised training with temporal pseudo-labels beats or matches fully supervised phase recognition.","key_machinery":"The load-bearing mechanism is a teacher-student video transformer pair with temporal sampling as the augmentation. A weak view (current frame plus the preceding 15 consecutive frames) is passed to an EMA-updated teacher, which produces pseudo-labels; a strong view (current frame plus 15 randomly chosen past frames) is passed to the student, and cross-entropy between the two predictions is counted only when the teacher's maximum probability exceeds a confidence threshold $\\delta$. In parallel, a triplet margin loss pulls normalized features toward their class prototype and pushes them away from the three nearest negative prototypes, with prototypes updated by exponential moving average. The temporal sampling is what makes the regularization surgical-domain specific: consistency across short- and long-term context encodes the fact that the same phase can recur and that phase transitions are anchored in the operative workflow.","core_discovery":"On the paper's own terms, the central discovery is that self-training a video transformer on unlabeled surgical footage need not be limited to pretraining: a pseudo-labeling loop with temporal consistency and prototype-based contrastive learning can lift phase-recognition accuracy directly. With labels only, the model reaches 78.7% accuracy on RAMIE; adding 20 unlabeled videos with the full objective raises it to 83.4%, above the fully supervised TeCNO baseline (78.5%) and above two large self-supervised pretraining baselines. On Cholec80, 20 labeled videos yield 90.4% accuracy versus 87.0% for SurgSSL and 88.6% for fully supervised TeCNO, and the same pattern holds at 10 and 5 labeled videos. The paper therefore claims that its objective is label-efficient and that it sets a benchmark for semi-supervised surgical phase recognition.","pith_inferences":["An implication the paper leaves implicit is a deployment path: hospitals with large archives of unlabeled surgical video could bootstrap phase-recognition models from just a few annotated cases of the same procedure.","Because the confidence threshold is set separately per dataset (0.6 for RAMIE, 0.8 for Cholec80), a natural extension is to make the threshold per class or per video, which could improve recall on underrepresented phases without changing the objective.","The same long-short temporal sampling scheme could transfer to other long-video workflow recognition tasks where phases recur and are defined by operative progress rather than by scene content."],"forward_implications":["Unlabeled surgical videos from the same procedure type can be used as a training signal during fine-tuning, not only for pretraining, which lowers the annotation cost of phase recognition systems.","With only a quarter of the labeled Cholec80 videos, the method reports higher accuracy than the fully supervised baseline, so fully manual annotation of every training video may not be necessary for this task.","Combining temporal consistency regularization with prototype-based contrastive learning gives larger gains than either component alone, according to the paper's ablation results.","Adding a causal temporal convolutional head on top of the transformer further improves long-term phase modeling, showing that the two forms of temporal modeling are complementary."],"supporting_citations":[{"why":"It supplies the TimeSformer video-transformer architecture whose class token serves as the feature embedding for both teacher and student.","marker":"[5]"},{"why":"It provides the fully supervised TeCNO baseline on both datasets and the causal temporal convolution head added in the final model.","marker":"[7]"},{"why":"Its Kinetics-400 weights initialize the encoder and accelerate convergence before semi-supervised training begins.","marker":"[15]"},{"why":"It is the semi-supervised SurgSSL baseline that SemiVT-Surge reports beating at 20, 10, and 5 labeled videos on Cholec80.","marker":"[19]"},{"why":"It provides the FixMatch baseline and the weak/strong consistency-with-confidence idea that this paper adapts to temporal sampling.","marker":"[20]"},{"why":"It supplies the mean-teacher EMA strategy used to update the teacher network and to stabilize pseudo-label targets.","marker":"[21]"},{"why":"It provides the public Cholec80 dataset with official train/test split and phase annotations used for the label-efficiency experiments.","marker":"[22]"},{"why":"It supplies the private RAMIE dataset with 27 labeled training videos and the unlabeled pool used for the main experiments.","marker":"[2]"},{"why":"It is one of the large self-supervised pretraining baselines that SemiVT-Surge outperforms on RAMIE.","marker":"[13]"},{"why":"It is the other large self-supervised pretraining baseline, trained on endoscopic clips, that SemiVT-Surge outperforms on RAMIE.","marker":"[23]"}],"fun_headline_variants":["Semi-supervised video transformer lifts surgical phase accuracy by 4.9%","Unlabeled surgical videos boost phase recognition beyond full supervision","Pseudo-labeling video transformer rivals full supervision with quarter labels","SemiVT-Surge: Semi-supervised transformer matches full labels on Cholec80","Self-training on unlabeled surgery beats fully supervised phase recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the randomly sampled labeled subsets used in the Cholec80 experiments are representative of the subsets used for the comparison baselines; because no seed or repeated sampling is reported, a single lucky draw of 5 to 20 videos could explain the claimed gains.","fun_headline_variants_meta":{"raw":{"variants":["Semi-supervised video transformer lifts surgical phase accuracy by 4.9%","Unlabeled surgical videos boost phase recognition beyond full supervision","Pseudo-labeling video transformer rivals full supervision with quarter labels","SemiVT-Surge: Semi-supervised transformer matches full labels on Cholec80","Self-training on unlabeled surgery beats fully supervised phase recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001354,"raw_usage":{"total_tokens":5493,"prompt_tokens":940,"completion_tokens":4553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":4459}},"tokens_in":556,"tokens_out":4553,"duration_ms":36168,"temperature":1.0,"reasoning_tokens":4459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:40:54.569837+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Cholec80 comparison at 20, 10, and 5 labeled videos over at least ten random subset draws with fixed seeds and report mean and standard deviation; if the reported gaps over SurgSSL (90.4 vs 87.0, 88.8 vs 85.0, 85.3 vs 83.2) do not persist across most draws, the label-efficiency claim collapses to a dataset-luck artifact.","supporting_citations":[{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part III 23","cited_arxiv_id":null,"evidence_quote":"It provides the fully supervised TeCNO baseline on both datasets and the causal temporal convolution head added in the final model."},{"cited_title":"Medical Image Analysis 73, 102158 (2021)","cited_arxiv_id":null,"evidence_quote":"It is the semi-supervised SurgSSL baseline that SemiVT-Surge reports beating at 20, 10, and 5 labeled videos on Cholec80."},{"cited_title":"Advances in neural information processing systems30 (2017) Semi-Supervised Video Transformer for Surgical Phase Recognition 11","cited_arxiv_id":null,"evidence_quote":"It supplies the mean-teacher EMA strategy used to update the teacher network and to stabilize pseudo-label targets."},{"cited_title":"In: This paper is accepted by XXX but not yet available to the public","cited_arxiv_id":null,"evidence_quote":"It supplies the private RAMIE dataset with 27 labeled training videos and the unlabeled pool used for the main experiments."}],"review_version":1}