{"id":"028c0ee5-c3a1-4d8d-b915-65910bf3a72d","arxiv_id":"2505.13746","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Fine-tuning a CLIP image encoder with learned prompts improves surgical phase recognition by 1 to 4 accuracy points on three datasets, though the comparison baseline does not control for CLIP pretraining.","lead":"This paper proposes using CLIP, a vision-language model, with learned text prompts to train the image encoder for surgical phase recognition. The method improves phase recognition by 1 to 4 accuracy points on three laparoscopic video datasets, but the comparison may not isolate the effect of the new technique.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains are not isolated from CLIP initialization: Table 3 compares prompt-learned CLIP ResNet-50 against ImageNet ResNet-50, so prompt learning is confounded with pretrained weights.","rationale":"The reader's weakest_assumption correctly identifies the pretrained-backbone confound, and my independent reading of Section 4 and Table 3 confirms it. This is the single most load-bearing issue because the paper's novelty claim rests on prompt learning, but the experimental design does not isolate prompt learning from the switch to CLIP initialization. A matched CLIP baseline without prompt learning would settle the question. The reader's CONDITIONAL verdict is therefore appropriate: the empirical claim is plausible but not yet supported. I do not see a reason to move the verdict to ACCEPT or REJECT; the requested matched baseline and repetition details are the right conditions. I also note that no code is released, but the code-release statement is conditional on acceptance and is secondary to the confound. No additional concern about internal inconsistency arose from the manuscript text.","tokens_in":7541,"tokens_out":2219,"duration_ms":22932,"concrete_test":"Retrain the proposed first stage with the CLIP-pretrained ResNet-50 but replace the prompt/text-encoder head with a linear classifier on the same 1024-dimensional feature, keeping all other hyperparameters (data, augmentation, epochs, loss, TCN stage) identical to the paper's setting. If this matched baseline achieves accuracy, Jaccard, and F1 scores within noise of ReSW-VL on Cholec80, Autolaparo, and m2cai16, then prompt learning contributes little beyond CLIP initialization and the central claim is unsupported; if the matched baseline clearly underperforms across all three datasets, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that prompt learning on a CLIP image encoder improves surgical phase recognition. The evidence is Table 3, which compares ReSW-VLi/ReSW-VLo against a 'conventional method' defined in Section 4 as an ImageNet-pretrained ResNet-50 fine-tuned on the same datasets. Two variables change between the arms simultaneously: the pretrained weights (CLIP vs ImageNet) and the first-stage training objective (prompt learning with a text-encoder logit head vs plain cross-entropy classification). The paper states that the hyperparameters are identical, but it does not include the missing control cell: a CLIP-pretrained ResNet-50 fine-tuned with the same cross-entropy objective and no prompt learning. Without that cell, the reported accuracy gains of roughly 1.0-4.3 points, Jaccard gains of 3.7-6.1 points, and F1 gains of 2.4-3.9 points cannot be attributed to prompt learning; they could be entirely due to CLIP's richer pretrained representations. Because prompt learning is the paper's stated contribution, this confound is load-bearing. The absence of repeated-seed variance (the reported standard deviations are over phases, not over training runs) further prevents assessing whether the gaps are significant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ReSW-VL, a two-stage method for surgical phase recognition. In the first stage, a CLIP ResNet-50 image encoder is fine-tuned with learned prompts while the text encoder is frozen; the logit is the inner product of the image feature with text features derived from phase prompts. Two variants are introduced: ReSW-VLi, which learns each phase's first prompt token independently, and ReSW-VLo, which learns a small number of reference tokens and interpolates the remaining phase tokens to preserve phase order. In the second stage, the image encoder is frozen and a causal TCN is trained on the resulting frame features. The method is evaluated on Cholec80, Autolaparo, and m2cai16 against a conventional baseline consisting of an ImageNet-pretrained ResNet-50 fine-tuned on the same datasets with the same TCN second stage. The reported results show consistent improvements in accuracy, Jaccard index, and F1 score for at least one proposed variant on each dataset.","tokens_in":7728,"tokens_out":4672,"duration_ms":44100,"significance":"If the reported gains were attributable to the proposed prompt-learning procedure, this would be a useful contribution: it is, to my knowledge, the first application of vision-language prompt learning to surgical phase recognition, and it addresses a genuine gap in the literature on representation learning for the spatial feature extractor. The evaluation uses three standard public datasets and multiple phase-wise video-wise metrics, and the two-stage protocol is described with enough detail to be reproduced. A notable strength is that the authors explicitly report standard deviations, although those deviations are over phases rather than over training runs. The main reason the significance is not yet established is that the central comparison confounds the pretrained weights with the prompt-learning objective, and no repeated-run statistics are provided to support the magnitude of the reported differences.","major_comments":[{"comment":"The comparison between the proposed methods and the conventional method changes two variables at once: the pretrained weights (CLIP versus ImageNet) and the first-stage training objective (prompt learning with a text-encoder logit head versus plain cross-entropy classification). Because the proposed method always starts from CLIP initialization, the reported gains of roughly 1.0-4.3 accuracy points, 3.7-6.1 Jaccard points, and 2.4-3.9 F1 points could be entirely due to the richer CLIP representation rather than to the prompt-learning procedure. The paper needs a missing control cell: a CLIP-pretrained ResNet-50 fine-tuned with the same cross-entropy objective and no prompt learning, run under otherwise identical hyperparameters. Without this cell, the central claim that prompt learning improves surgical phase recognition is not supported.","section":"Section 4, Table 3"},{"comment":"The reported standard deviations are over phases, not over training runs, and the manuscript provides no repeated-seed variance or significance tests. On Autolaparo, ReSW-VLi achieves an accuracy of 75.69 versus 75.68 for the conventional method, a difference of 0.01 percentage points that the text nevertheless summarizes as consistent outperformance. At minimum, the authors should report means and standard deviations over at least three training runs and perform a paired per-video significance test (for example, a paired bootstrap or Wilcoxon test) to establish that the observed gaps are not noise.","section":"Section 4, Table 3"},{"comment":"The only stated justification for setting n=3 in ReSW-VLo is 'Based on preliminary experiments,' with no details of those experiments, and the number m of learned prompt tokens is never reported anywhere in the manuscript. Since the relative advantage of ReSW-VLo over ReSW-VLi is dataset-dependent and is attributed to the sequentiality assumption, the paper should include a sensitivity analysis for n (for example, n=2, 3, and 5) and a stated value for m. Without this, the mechanism behind the proposed method and its dataset-dependent behavior remain unsupported.","section":"Section 3 and Section 4"},{"comment":"The conclusion that the proposed method 'consistently surpassed conventional methods' is broader than the evidence in Table 3, which compares only against the authors' reimplemented ImageNet ResNet-50 plus TCN baseline. Since Cholec80, Autolaparo, and m2cai16 are standard datasets with published results, the paper should include at least one comparison with a prior published method from Table 1 (for example, TeCNO or Trans-SVNet) under the same evaluation protocol, or should explicitly state why such a comparison is omitted. As it stands, the claimed superiority over 'conventional methods' is calibrated only against a single baseline.","section":"Section 1 and Section 4"}],"minor_comments":[{"comment":"The abstract expands CLIP as 'Convolutional Language Image Model'; the correct expansion is 'Contrastive Language-Image Pre-training'.","section":"Abstract"},{"comment":"The heading 'Propose Method' should be 'Proposed Method'.","section":"Section 3 heading"},{"comment":"The text refers to the 'M2cai dataset' while Table 2 uses 'm2cai16'; please use consistent naming throughout.","section":"Section 4"},{"comment":"The paper does not state which specific CLIP checkpoint was used (for example, OpenAI's CLIP ResNet-50); this information is needed for reproducibility.","section":"Section 4"},{"comment":"The captions of Figures 3 and 4 contain garbled text such as 'C ... R ...' and 'P ...'; these appear to be rendering artifacts and should be regenerated.","section":"Figures 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is clearly written and the proposed direction is potentially interesting, but the central empirical claim is currently confounded by the simultaneous change of pretrained weights and training objective. I do not see evidence of methodological misconduct; the issue is a missing control experiment rather than a circular or fabricated result. If the authors add the CLIP-without-prompt control, report repeated-seed statistics, and state the missing hyperparameters, the paper could become acceptable. I would also encourage requesting code release, since the manuscript states that code will be made available only upon acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a straightforward transfer of CLIP prompt learning to surgical phase recognition. What is actually new is the application: a CLIP image encoder fine-tuned with learned prompts, then frozen, then a causal TCN on top. On Cholec80, Autolaparo, and m2cai16, the authors report consistent gains over an ImageNet ResNet-50 baseline. That is a legitimate extension, and the qualitative figures show smoother phase predictions.\n\nThe paper does some things well. The two prompt variants are clearly explained. ReSW-VLo, which borrows ordinal interpolation from OrdinalCLIP (cited), is tested against ReSW-VLi, and the authors give a sensible explanation for why ordered prompts help on Cholec80 (phases progress monotonically) but not on Autolaparo (phases go back and forth). The experimental setup is mostly described with hyperparameters listed.\n\nNow the soft spots. The stress-test is right: the headline comparison is confounded. The baseline uses an ImageNet-pretrained ResNet-50 fine-tuned with plain cross-entropy. The proposed method uses a CLIP-pretrained ResNet-50 fine-tuned with prompt learning plus cross-entropy. Two things change at once. Without a control cell that fine-tunes the CLIP encoder without prompts, you cannot attribute the 1-4 point accuracy gains to prompt learning. It could be the CLIP weights. That is the load-bearing flaw, and it is fixable.\n\nAlso: no repeated-seed variance, so the reported standard deviations are across phases, not across runs. The gains may or may not be significant. n=3 for the ordered variant was chosen after preliminary experiments, which is a silent selection step. Prompt length m is mentioned but never given a value. CoOp is not cited for independent prompts; the method resembles it closely enough that the omission should be addressed. Code is promised only on acceptance.\n\nNone of this is fatal. The central idea is plausible and the domain transfer is sensible. But the current evidence does not establish that prompt learning is what helps. That missing control cell needs to be added, along with seed variance and complete hyperparameters.\n\nWho is this for? Researchers working on surgical phase recognition who want a quick test of CLIP-based feature extractors. It is not a methodological breakthrough, but it is a reasonable empirical note. I would send it to peer review because the confound is fixable and the application is useful to a specific community. I would not cite it in its current form, because the main claim is not yet isolated.\n\nRecommendation: do not desk reject, but the referee should ask for the matched baseline.","headline":"An incremental but sensible application of CLIP prompt learning to surgical phase recognition, undermined by a missing control cell that leaves the reported gains unassigned between prompt learning and CLIP initialization.","tokens_in":8308,"tokens_out":2030,"would_cite":false,"duration_ms":18468,"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":"Prompt-trained CLIP image encoder improves surgical phase recognition on three laparoscopic video datasets.","keywords":["surgical phase recognition","vision-language model","CLIP","prompt learning","representation learning","temporal convolutional network","laparoscopic video","surgical workflow"],"falsifier":"Retrain the proposed pipeline on Cholec80 with the CLIP ResNet-50 but use fixed random prompts (or no text branch at all) while keeping the same loss; if test accuracy stays at the ReSW-VLo level of about 85.6 percent, prompt learning is not what drives the improvement.","tokens_in":7275,"feed_emoji":"🩺","tokens_out":7318,"duration_ms":55742,"temperature":0.7,"pith_summary":"The paper sets out to show that how the CNN feature extractor is trained matters for surgical phase recognition, and that a vision-language pretrained encoder fine-tuned with learned prompts works better than the standard recipe of fine-tuning an ImageNet-pretrained ResNet-50. The proposed method, ReSW-VL, uses the image encoder of CLIP together with per-phase prompt vectors, then feeds the frozen encoder's features into a causal temporal convolutional network. On three laparoscopic surgery datasets, Cholec80, Autolaparo, and m2cai16, the method reports higher accuracy, Jaccard index, and F1 score than the conventional two-stage baseline. A sympathetic reading is that prompt learning is a practical and underexplored form of representation learning for surgical workflow analysis, with the ordered-prompt variant exploiting the sequential nature of many procedures.","feed_headline":"Prompt learning lifts surgical phase recognition up to 6 points","feed_subtitle":"Fine-tuning CLIP with learned phase prompts beats conventional baselines on three surgical datasets.","key_machinery":"The central mechanism is prompt learning on a CLIP model. A CLIP image encoder converts each frame into a $d$-dimensional feature vector, and a frozen text encoder converts $P$ learned prompt vectors into $P$ text features; the inner products of image and text features produce phase logits, and weighted cross-entropy trains the image encoder and the prompt tokens together. For ReSW-VLo, only $n$ reference phase tokens are learned and the remaining phase tokens are obtained by linear interpolation, which preserves the ordering of phases in text-embedding space. This prompt-learning objective is what the paper credits for reshaping the spatial feature space, and the same frozen features are later paired with a causal TCN for temporal modeling.","core_discovery":"ReSW-VL claims that fine-tuning a CLIP image encoder under a prompt-learning objective produces spatial features that are better suited to phase recognition than features from a conventionally fine-tuned CNN, under the same temporal model. Each surgical phase is assigned a prompt whose first token is a learnable embedding of the phase number; ReSW-VLo interpolates between a small set of learned reference tokens to encode the ordered progression of phases, while ReSW-VLi learns the phase tokens independently. After training, the image encoder is frozen and a causal TCN maps the frame-feature sequence to phase predictions. The reported results on Cholec80, Autolaparo, and m2cai16 show consistent improvements over the conventional ImageNet-pretrained ResNet-50 baseline, ranging from roughly 1.0 to 4.3 percentage points in accuracy, 3.7 to 6.1 points in Jaccard index, and 2.4 to 3.9 points in F1 score, with the best prompt variant depending on whether the dataset's phases advance monotonically.","pith_inferences":["A decisive control experiment would decouple pretraining from training method: apply the same prompt-learning objective to an ImageNet-pretrained ResNet-50, and apply plain fine-tuning to the CLIP ResNet-50, to see whether the reported gains come from prompt learning, CLIP initialization, or their interaction.","The ordered-prompt idea could generalize to non-linear phase graphs, for example by learning a phase-transition matrix or a low-dimensional manifold of phase tokens instead of linear interpolation.","Since the text encoder stays frozen and only a small number of prompt tokens are trained, the method may transfer to other procedural video domains (endoscopy, cataract surgery) with modest labeled data, almost by rewriting the phase vocabulary.","The paper evaluates on 1 fps frames; applying the same representation at higher frame rates or with a different temporal model might interact non-trivially with the frozen features, which the current experiments do not address."],"forward_implications":["On Cholec80, ReSW-VLo raises F1 from 77.89 to 81.81 and Jaccard from 62.36 to 68.45, the largest gains reported in the paper.","On m2cai16, both prompt variants improve F1 by about 2.3 to 2.4 points and Jaccard by about 3.5 to 5.7 points, showing the benefit is not limited to one dataset.","On Autolaparo, ReSW-VLi beats ReSW-VLo, which suggests the ordered-prompt prior only helps when phase numbers actually progress monotonically.","Because the image encoder is frozen in the second stage, the learned representation can be plugged into LSTM, TCN, or Transformer temporal models without retraining the spatial features.","The method provides an alternative to Transformer-only surgical phase recognition, keeping a CNN-plus-temporal-model design while improving the feature extraction stage."],"supporting_citations":[{"why":"Supplies the CLIP vision-language model whose image encoder is fine-tuned in the first stage.","marker":"[19]"},{"why":"Contributes the ordered prompt interpolation method that ReSW-VLo adopts for relating phase-number tokens.","marker":"[13]"},{"why":"Provides the causal TCN architecture used as the temporal model in the second stage and in the conventional baseline.","marker":"[1]"},{"why":"Supplies the Cholec80 dataset with seven surgical phase definitions used for training and testing.","marker":"[22]"},{"why":"Supplies the Autolaparo hysterectomy dataset, whose non-monotonic phase order tests the ordered-prompt variant.","marker":"[24]"},{"why":"Supplies the m2cai16 dataset as the third evaluation benchmark.","marker":"[21]"},{"why":"Defines the phase-wise video-wise precision, recall, and Jaccard metrics used for the reported comparisons.","marker":"[7]"},{"why":"Provides the median frequency balancing scheme used for the weighted cross-entropy loss in both stages.","marker":"[6]"}],"fun_headline_variants":["Prompt learning with CLIP boosts surgical phase recognition","ReSW-VL: vision-language model improves surgical workflow analysis","Surgical phase recognition up 6 points via CLIP prompt learning","Fine-tuned CLIP with phase prompts advances surgical recognition","Learned prompts for CLIP sharpen surgical phase classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed gain is attributed to prompt learning, but the experiments compare a CLIP-pretrained ResNet-50 to an ImageNet-pretrained ResNet-50, so the improvement could come entirely from the different pretrained weights rather than from the prompt-learning procedure.","fun_headline_variants_meta":{"raw":{"variants":["Prompt learning with CLIP boosts surgical phase recognition","ReSW-VL: vision-language model improves surgical workflow analysis","Surgical phase recognition up 6 points via CLIP prompt learning","Fine-tuned CLIP with phase prompts advances surgical recognition","Learned prompts for CLIP sharpen surgical phase classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1455,"prompt_tokens":934,"completion_tokens":521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":440}},"tokens_in":550,"tokens_out":521,"duration_ms":4626,"temperature":1.0,"reasoning_tokens":440,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:10:34.486843+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the proposed pipeline on Cholec80 with the CLIP ResNet-50 but use fixed random prompts (or no text branch at all) while keeping the same loss; if test accuracy stays at the ReSW-VLo level of about 85.6 percent, prompt learning is not what drives the improvement.","supporting_citations":[{"cited_title":"In: International conferen ce on machine learning","cited_arxiv_id":null,"evidence_quote":"Supplies the CLIP vision-language model whose image encoder is fine-tuned in the first stage."},{"cited_title":"Adva nces in Neural Informa- tion Processing Systems 35, 35313–35325 (2022)","cited_arxiv_id":null,"evidence_quote":"Contributes the ordered prompt interpolation method that ReSW-VLo adopts for relating phase-number tokens."},{"cited_title":"In: Medical Image Computing and Compu ter Assisted Intervention–MICCAI 2020: 23rd International Conference , Lima, Peru, October 4–8, 2020, Proceedings, Part III 23","cited_arxiv_id":null,"evidence_quote":"Provides the causal TCN architecture used as the temporal model in the second stage and in the conventional baseline."},{"cited_title":"In: International Conference on Medical Image Com- puting and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Supplies the Autolaparo hysterectomy dataset, whose non-monotonic phase order tests the ordered-prompt variant."},{"cited_title":"In: Proc eedings of the IEEE in- ternational conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Provides the median frequency balancing scheme used for the weighted cross-entropy loss in both stages."}],"review_version":1}