{"id":"0d1f56a2-36c9-4631-9201-645e00f27846","arxiv_id":"2507.18082","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TextSAM-EUS combines learned text prompts with LoRA-tuned SAM to segment pancreatic tumors in endoscopic ultrasound automatically, reaching 82.69% Dice on a public dataset.","lead":"A lightweight version of SAM, guided by learned text prompts and small trainable adapters, segments pancreatic tumors in endoscopic ultrasound without any manual clicks at inference. On a public 18-patient dataset it reaches roughly 82.7% Dice, near or above strong supervised baselines while tuning only about 1.7 million parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set selection of the final hyperparameters makes the reported 82.69% Dice and the p<0.05 comparisons optimistically biased; the headline claim needs a validation-frozen evaluation.","rationale":"The reader's weakest assumption identifies the test set as the source of design choices, and Section 4.5 confirms this directly: all four ablations are reported on the held-out test set, and the final configuration is the best-performing combination there. This is the most load-bearing weakness because it affects the point estimate itself, not just the significance machinery. The paper otherwise has substantial merits: the method is clearly described, the code is public, the dataset is public, and the parameter-efficient design is a reasonable contribution even if the numerical claim needs to be re-anchored. The per-frame t-test issue is a second, related threat to the 'significantly outperforms all automatic baselines' claim; it compounds the selection bias but does not replace it. A validation-frozen protocol with patient-level statistics would settle both concerns. Since the reader's proposed condition is exactly the repair needed, I agree with the reader and see no reason to change the conditional verdict.","tokens_in":12891,"tokens_out":8660,"duration_ms":89206,"concrete_test":"Select all hyperparameters (LoRA rank, context-token length, prompt depth, and refinement-prompt choice) using only the two-patient validation split by DSC or validation loss, freeze the chosen configuration, then evaluate the frozen model on the four-patient test set. If the validation-selected configuration differs from the reported one, or its test Dice/NSD falls materially below 82.69/85.28, the headline result is optimistically biased.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on 82.69% Dice and 85.28% NSD being an unbiased estimate of TextSAM-EUS on the test distribution. That condition is violated in Section 4.5: every ablation—LoRA rank (Table 3), context-token length (Table 4), prompt-injection depth (Table 5), and refinement-prompt choice (Table 6)—is explicitly evaluated on the held-out test set, and the final configuration (r=16, b=4, t=12, box+centroid) is the best-scoring combination on that same set. Reporting metrics for a configuration selected on the test set turns the headline numbers into selection-optimized upper bounds rather than unbiased predictions. The selection effect is not negligible: the test-set comparisons move DSC by up to about 2.2 points (depth 9 vs 12), about 1.4 points (LoRA r=4 vs 16), and about 0.7 points (refinement on/off). Additionally, the significance claim uses two-tailed paired t-tests over 4,185 frames drawn from only 4 test patients; per-frame tests ignore within-patient correlation and cannot support the across-patient superiority claim. The claim can be repaired by freezing all hyperparameters on validation data and reporting patient-level or cluster-robust intervals on the test set.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TextSAM-EUS, a parameter-efficient adaptation of the Segment Anything Model (SAM) for automatic pancreatic tumor segmentation in endoscopic ultrasound (EUS). The method combines LoRA-based fine-tuning of SAM's image encoder and mask decoder with learnable text prompt tokens processed through a frozen BiomedCLIP text encoder, replacing manual geometric prompts with learned text-driven prompts at inference. An iterative refinement stage derives box and centroid prompts from the initial text-driven mask and feeds them back into SAM. On the public Endoscopic Ultrasound Database of the Pancreas, the authors report 82.69% Dice and 85.28% NSD for the fully automatic variant, and 83.10% Dice and 85.70% NSD when manual prompts are supplied, while tuning only 0.86% (1.69M) of SAM's parameters. Ablations study LoRA rank, context-token length, prompt-injection depth, and refinement-prompt choices. The paper claims significant improvements over all automatic baselines and over several manual-prompted SAM variants.","tokens_in":13210,"tokens_out":3458,"duration_ms":38065,"significance":"If the reported evaluation were unbiased, the contribution would be practically valuable: a text-driven, prompt-free SAM adaptation for a noisy, low-contrast ultrasound modality, with a very small trainable parameter footprint and a clear baseline comparison against both fully supervised and foundation-model methods. The code release and use of a public dataset are strengths that support reproducibility. The central novelty, integrating CoOp-style prompt learning with SAM via a biomedical text encoder, is plausible and timely. However, the evaluation protocol contains a load-bearing flaw: the held-out test set is used for hyperparameter and design selection, and the statistical comparisons ignore within-patient correlation. These issues directly affect the headline accuracy numbers and the 'significant outperformance' claim, so the current evidence does not yet support the stated state-of-the-art conclusion.","major_comments":[{"comment":"All four ablations (LoRA rank, context-token length, prompt-injection depth, and refinement-prompt choice) are explicitly evaluated on the held-out test set: 'All results are reported on the held-out segmentation test set.' The final configuration (r=16, b=4, t=12, box+centroid) is then the best-scoring combination on that same test set, and the reported 82.69% Dice and 85.28% NSD come from that configuration. This makes the headline numbers selection-optimized estimates, not unbiased predictions of performance on the test distribution. The problem is compounded in the comparison with baselines: baseline hyperparameters follow the recommendations of their original papers, while TextSAM-EUS's hyperparameters are selected on the test set. The paper should freeze all hyperparameters using only the validation split (or use nested cross-validation) and report test metrics for the frozen configuration; Section 4.3's statement that 'model selection is based on the lowest validation loss' does not resolve this, because the ablation choices are not made on validation data.","section":"Section 4.5, Tables 3-6"},{"comment":"The significance claims are based on two-tailed paired-sample t-tests over 4,185 frames drawn from only 4 test patients. Treating each frame as an independent observation ignores the strong within-patient correlation of EUS frames, so the effective sample size for the comparison is much closer to 4 than to 4,185, and the reported p-values are anti-conservatively biased. The paper should report patient-level metrics (e.g., per-patient Dice and NSD) and use a patient-level or cluster-robust significance test, with appropriate correction for the multiple pairwise comparisons against the many baselines. Without this, the statement that TextSAM-EUS 'significantly outperforms all automatic baselines' is not supported.","section":"Section 4.4, statistical analysis"},{"comment":"The text says the automatic TextSAM-EUS and MedSAM yield 'nearly identical DSC (82.69 vs. 82.66; p = 0.906)', but then claims to 'significantly outperform all automatic baselines'. MedSAM is not automatic, since it is supplied with ground-truth bounding boxes, so the claim is internally consistent only under the paper's definition of 'automatic'. However, the comparison also shows that MedSAM achieves significantly higher NSD (85.75 vs. 85.28; p = 0.035). The Discussion should state this discrepancy clearly rather than describing the two as nearly identical across both metrics; the current wording overstates the closeness of the result.","section":"Section 4.4, MedSAM comparison"}],"minor_comments":[{"comment":"The preliminary ablation in which 'swapping the SAMUS encoder into our framework yielded a maximum of 67.53% DSC' is mentioned only in prose and not reported in any table or with experimental details; either add the full setup and results to Section 4.5 or remove the claim.","section":"Section 5"},{"comment":"The phrase 'the first attempt to incorporate prompt learning in SAM-based medical image segmentation' is strong given the authors' own MedCLIP-SAM family [23,24] and the existence of text-guided SAM variants such as CC-SAM [10]; the novelty claim should be qualified to specify what exactly is new relative to these prior text-driven SAM adaptations.","section":"Abstract and Introduction"},{"comment":"The description of the prompt-injection depth is ambiguous: 'a depth of 12 layers (threaded through the first 12 Transformer layers)' could be read as two different quantities. Clarify that BiomedCLIP has 12 blocks total and that t=12 means the learnable tokens are updated through all blocks.","section":"Section 4.3"},{"comment":"SwinUNet's NSD of 10.22 ± 6.64 with a Dice of 59.69 is an extreme outlier relative to its DSC; a brief explanation of this behavior (e.g., boundary tolerance effects) would help the reader interpret the table.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The evaluation-protocol issue is the main reason for major revision rather than acceptance. The central method is plausible and the code release is a plus, but the headline numbers and significance claims rest on test-set-based model selection. I would encourage the editor to require a validation-frozen evaluation and patient-level statistical analysis before considering the paper for publication. The 'first' novelty claim also needs tightening relative to the authors' own earlier MedCLIP-SAM work, which may be a scope or novelty-disclosure concern for the editor's awareness."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid application paper with code and a public benchmark. The method — CoOp-style learnable context tokens through BiomedCLIP, a small adapter into SAM's prompt encoder, LoRA on the image encoder and mask decoder, and iterative refinement using auto-extracted geometric cues — is coherent and genuinely automatic at inference. Tuning 0.86% of SAM's parameters to reach 82.69% Dice on the public EUS dataset is a useful result, especially for low-resource medical segmentation. The ablations are thorough and the related work is handled honestly, including the group's own earlier text-driven SAM variants.\n\nThe main soft spot is exactly what the stress-test flags. Section 4.5 reports every ablation on the held-out test set, and the final configuration (LoRA rank 16, 4 context tokens, depth 12, box plus centroid refinement) is the best-scoring combination on that same set. The selection swings are not negligible — up to about 2.2 points in DSC for depth, 1.4 for LoRA rank, 0.7 for refinement — and the claimed advantage over the strongest automatic baseline, AutoSAM, is only about 1.4 points. So the headline 'significantly outperforms all automatic baselines' is optimistically biased as written. The fix is straightforward: freeze the configuration on validation data before touching the test set.\n\nThe second issue is statistical. The p < 0.05 comparisons use paired t-tests over 4,185 frames from four test patients. That treats frames as independent, which they are not. Patient-level or cluster-robust inference is needed before claiming superiority over baselines. With only four test patients, the significance claims would likely be much weaker.\n\nMinor points: the 'first attempt to incorporate prompt learning in SAM' phrasing needs qualification given MedCLIP-SAM and CC-SAM already use text-driven SAM prompts, though the CoOp-style learnable-token twist is new. The citation pattern is fine; citing your own earlier work as prior art is legitimate here.\n\nThe central idea holds up, and the numbers are probably in the right ballpark, but the evaluation protocol needs to be corrected before the SOTA claim can be trusted. I would send this to peer review — it is worth refereeing — and ask for a validation-frozen evaluation and patient-level statistics. For a reading group, it is a good example of both a useful medical SAM adaptation and a subtle evaluation pitfall.","headline":"A competent parameter-efficient text-driven SAM adaptation for EUS tumor segmentation, with a real but fixable evaluation flaw: the final hyperparameters are selected on the test set, and the significance tests are frame-level.","tokens_in":13717,"tokens_out":2710,"would_cite":true,"duration_ms":31298,"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":"Text prompts alone can drive SAM to 82.69% Dice on pancreatic EUS tumor segmentation, with no manual points or boxes at inference.","keywords":["text prompt learning","Segment Anything Model","pancreatic tumor segmentation","endoscopic ultrasound","low-rank adaptation","BiomedCLIP","context optimization","parameter-efficient fine-tuning"],"falsifier":"Freeze the final configuration (LoRA rank 16, 4 context tokens, depth 12, box plus centroid refinement) using only the training and validation splits, then evaluate it once on a test set that has never influenced any choice. If the automatic variant's Dice drops below AutoSAM or AutoSAMUS, or the p < 0.05 superiority claim reverses, the central claim is refuted. A second check is to report per-patient confidence intervals to see whether the roughly 1.4–1.65% Dice gap is stable across patients.","tokens_in":12737,"feed_emoji":"🩺","tokens_out":6106,"duration_ms":63685,"temperature":0.7,"pith_summary":"TextSAM-EUS claims that a Segment Anything Model, adapted only through low-rank updates and learnable text tokens, can segment pancreatic tumors in endoscopic ultrasound without any manual geometric prompts at inference. On the public Endoscopic Ultrasound Database of the Pancreas, the fully automatic variant reports 82.69% Dice and 85.28% NSD, and the paper states this significantly outperforms all automatic baselines it compared against (p < 0.05) while tuning 0.86% of SAM's parameters. A variant given one ground-truth point and a perturbed box reports 83.10% Dice and 85.70% NSD, matching or exceeding MedSAM, which receives ground-truth boxes. The practical stake is that a clinician could obtain tumor delineation by specifying a class name, avoiding the radiological judgment that geometric prompting requires.","feed_headline":"Text prompts alone hit 82.7% Dice on pancreatic EUS tumors","feed_subtitle":"A lightly tuned SAM segments tumors with no manual points or boxes, matching a model that gets ground-truth boxes.","key_machinery":"The load-bearing object is the learned text prompt rather than a geometric prompt: a small set of context tokens (b = 4) concatenated with the class name 'tumor' and optimized through the first t = 12 Transformer layers of the BiomedCLIP text encoder, then mapped by an MLP adapter into SAM's prompt encoder. LoRA with rank 16 on the SAM image encoder and mask decoder makes the base model trainable with only 1.69 million parameters, about 0.86% of SAM's total. The mask decoder fuses the text-derived tokens with dense image embeddings, and an iterative refinement step adds a bounding box and centroid extracted from the initial prediction, giving the final 82.69% Dice automatic result.","core_discovery":"The central discovery is that language can replace geometry as the prompt modality for SAM in a noisy, low-contrast medical imaging setting. TextSAM-EUS threads learnable context tokens through the first 12 layers of the BiomedCLIP text encoder, projects the resulting embedding into SAM's prompt-token space with a two-layer MLP adapter, and feeds those text tokens to SAM's mask decoder alongside LoRA-adapted image features. After an initial text-only mask, it extracts the predicted mask's bounding box and centroid and feeds them back for two refinement iterations. The paper's headline assertion is that this automatic, text-driven pipeline reaches 82.69% Dice and 85.28% NSD on the public pancreatic EUS dataset, significantly outperforming all automatic baselines it compares against (p < 0.05), and that with manual geometric prompts it reaches 83.10% Dice and 85.70% NSD, matching or exceeding the strongest prompt-based baseline, MedSAM. The paper also claims this is the first use of prompt learning in SAM-based medical segmentation, and that the success shows text-to-ViT feature alignment carries more value for EUS than adding an external CNN encoder path.","pith_inferences":["Because the text encoder is frozen except for the context tokens, a testable extension is to check whether the same learned prompt tokens transfer across EUS scanners or patients; if they do, adapting to a new lesion type could reduce to retraining only the context tokens.","The comparison to MedSAM is asymmetric, since MedSAM receives ground-truth boxes while TextSAM-EUS receives none; an editorial consequence is that a head-to-head with MedSAM using identical predicted boxes would clarify whether text guidance or box guidance is the stronger prior.","A stricter evaluation protocol that holds out a separate fold for hyperparameter selection could reveal whether the reported 1.4–1.65% Dice gap over AutoSAM and AutoSAMUS persists; this is directly testable with the same public dataset."],"forward_implications":["If the claim holds, automatic EUS tumor segmentation no longer requires a human to place points or boxes; the same frozen SAM can be redirected to a new target by changing the learned text prompt and fine-tuning a very small parameter set.","The parameter footprint of 1.69 million trainable parameters means the method could run on clinical hardware with a single GPU, a different regime from full SAM fine-tuning or training a U-Net from scratch.","Text-only prompting already reaches 82.00% Dice before refinement, and manual geometric prompts add only about one point of Dice, suggesting language guidance carries most of the discriminative information needed for this binary task.","The reported significant superiority over AutoSAM and AutoSAMUS (p < 0.05) implies that, for EUS, learning text-to-feature alignment is a better use of adaptation capacity than learning automatic point or box generators."],"supporting_citations":[{"why":"Supplies the public Endoscopic Ultrasound Database of the Pancreas and the patient-wise split on which every reported DSC and NSD is measured.","marker":"[16]"},{"why":"Defines the Segment Anything Model architecture that TextSAM-EUS adapts, including the image encoder, prompt encoder, and mask decoder.","marker":"[21]"},{"why":"Provides the pretrained BiomedCLIP text encoder whose context tokens are learned to produce biomedical text embeddings.","marker":"[41]"},{"why":"Defines low-rank adaptation, the mechanism that lets TextSAM-EUS tune only 0.86% of SAM's parameters.","marker":"[12]"},{"why":"Supplies the context-optimization method for learning prompt tokens inside a frozen text encoder.","marker":"[20]"},{"why":"MedSAM is the strongest manual-prompt baseline that TextSAM-EUS claims to match or exceed, and it is evaluated with ground-truth bounding boxes.","marker":"[30]"},{"why":"SAMUS and AutoSAMUS are automatic-prompt SAM baselines for ultrasound that TextSAM-EUS claims to significantly outperform with far fewer trainable parameters.","marker":"[26]"},{"why":"SAMed is the LoRA-based SAM baseline that TextSAM-EUS outperforms, supporting the claim that deep text prompting adds value beyond LoRA adaptation alone.","marker":"[39]"}],"fun_headline_variants":["Text-prompt SAM hits 82.7% Dice on EUS pancreatic tumors","No boxes, no points: text-only SAM outscores supervised DL on EUS","First text-prompt learning for SAM in medical imaging: 82.7% Dice on EUS","Only 0.86% of SAM tuned, text prompts reach 82.7% Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the held-out test set was not used to choose the final configuration; the paper reports ablation results on that same test set, so if those design choices were made on it, the 82.69% Dice and the significance comparisons are optimistically biased.","fun_headline_variants_meta":{"raw":{"variants":["Text-prompt SAM hits 82.7% Dice on EUS pancreatic tumors","No boxes, no points: text-only SAM outscores supervised DL on EUS","First text-prompt learning for SAM in medical imaging: 82.7% Dice on EUS","Only 0.86% of SAM tuned, text prompts reach 82.7% Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3351,"prompt_tokens":1064,"completion_tokens":2287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":2193}},"tokens_in":680,"tokens_out":2287,"duration_ms":16186,"temperature":1.0,"reasoning_tokens":2193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:17:18.211869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Freeze the final configuration (LoRA rank 16, 4 context tokens, depth 12, box plus centroid refinement) using only the training and validation splits, then evaluate it once on a test set that has never influenced any choice. If the automatic variant's Dice drops below AutoSAM or AutoSAMUS, or the p < 0.05 superiority claim reverses, the central claim is refuted. A second check is to report per-patient confidence intervals to see whether the roughly 1.4–1.65% Dice gap is stable across patients.","supporting_citations":[{"cited_title":"Endoscopic ultrasound database of the pancreas","cited_arxiv_id":null,"evidence_quote":"Supplies the public Endoscopic Ultrasound Database of the Pancreas and the patient-wise split on which every reported DSC and NSD is measured."},{"cited_title":"Lungren, Tristan Naumann, Sheng Wang, and Hoifung Poon","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained BiomedCLIP text encoder whose context tokens are learned to produce biomedical text embeddings."},{"cited_title":"Lora: Low- rank adaptation of large language models","cited_arxiv_id":null,"evidence_quote":"Defines low-rank adaptation, the mechanism that lets TextSAM-EUS tune only 0.86% of SAM's parameters."},{"cited_title":"Segment anything in medical images","cited_arxiv_id":null,"evidence_quote":"MedSAM is the strongest manual-prompt baseline that TextSAM-EUS claims to match or exceed, and it is evaluated with ground-truth bounding boxes."},{"cited_title":"Be- yond adapting sam: Towards end-to-end ultrasound image segmentation via auto prompting","cited_arxiv_id":null,"evidence_quote":"SAMUS and AutoSAMUS are automatic-prompt SAM baselines for ultrasound that TextSAM-EUS claims to significantly outperform with far fewer trainable parameters."}],"review_version":1}