{"id":"6b17d216-03e0-4f26-8f24-9ba794e70991","arxiv_id":"2608.08191","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A bandit-based prompt-selection framework improves boundary metrics over fixed prompting in ultrasound segmentation, but the reported evaluation may rely on ground-truth prompts at test time.","lead":"BAP-MOS adapts a frozen SAM/MedSAM model to multi-organ ultrasound segmentation by learning, per organ and during fine-tuning, whether a box, point, or combined prompt gives the best boundary agreement. The paper reports large boundary-metric gains over conventional baselines, but it does not describe how prompts are obtained at test time, which makes the reported comparisons hard to interpret.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central evaluation depends on ground-truth-derived prompts at test time, with no specified inference-time prompt source, so the reported boundary gains may be oracle-prompted and the U-Net/nnU-Net comparison is not a test of a deployable automatic method.","rationale":"I read the paper in good faith. The framework is internally coherent: UCB-Tuned per-organ selection, a bounded composite reward, and a TPE outer loop are well specified, and the ablation in Table II does show that prompt scheduling can move boundary metrics substantially while Dice stays flat. The code release is a genuine asset, and I do not question the authors' intent. However, the reader's weakest-assumption identification is exactly the load-bearing risk I find. The paper never states how test prompts are produced, and its own prompt-geometry construction consumes GT masks. That makes the central evaluation unverifiable as a claim about an automatic segmentation method. My independent reading also noticed a supporting inconsistency: Table I reports Point-prompt decoder-only multi-organ HD95 of 32.89 and MSD of 12.06, which is far outside the range of all other reported MSD/HD95 numbers and suggests either a metric-unit or bookkeeping error; while not the central objection, it lowers confidence in the reported measurements. The external PFUS1 result lacks a fixed-prompt MedSAM comparison, so it cannot establish the generalization benefit of adaptive prompting. Because the key premise is unstated and testable only by running the code without GT prompts, the reader's REJECT verdict remains appropriate; I would not change it.","tokens_in":13035,"tokens_out":3583,"duration_ms":37603,"concrete_test":"Inspect the released GitHub repository's evaluation entrypoint and run BAP-MOS on the held-out TRUS test slices with all GT masks withheld. If the code requires box/point coordinates from GT masks to build prompts via Equations (5)-(6), record that inference is oracle-dependent. Then provide test prompts from an automatic detector (e.g., bounding boxes and centroids from a U-Net prediction on the same image) and compare Dice/HD95/MSD against Table III's BAP-MOS(SAM) row; if metrics degrade materially or prompt construction fails, the central comparison is not a standalone segmentation result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adaptive prompt allocation improves boundary-sensitive multi-organ segmentation. For that claim to hold as stated, prompts used at test time must be obtainable without ground-truth masks, or the method must be framed as interactive with an external prompt provider. Section IV-B-2 states that ground-truth masks are used only to construct prompt coordinates, and Equations (5)-(6) require the organ mask M_o to place the negative ring; no inference-time prompt generator is described. Section V-B reports Dice, HD95, and MSD on held-out TRUS slices without stating how test prompts were generated. If those test prompts come from GT masks, BAP-MOS is not a standalone segmentation method at inference: it is a fine-tuned promptable decoder evaluated with oracle-provided prompts. The headline comparison against U-Net and nnU-Net, which receive only images, is then apples-to-oranges, and the claimed ~48% HD95 and ~45% MSD reductions are not evidence for a deployable method. The PFUS1 external row cannot repair this because no fixed-prompt MedSAM baseline is reported there, so the adaptive mechanism is not isolated as the cause of the gain there either. This is the load-bearing weakness: the paper's strongest quantitative evidence is consistent with the hypothesis that the gains come from GT-derived test prompts rather than from adaptive prompt selection per se.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BAP-MOS, a framework that adaptively selects per-organ prompt types (box, point, or combined) for promptable segmentation backbones such as SAM and MedSAM. A Tree-structured Parzen Estimator (TPE) outer loop tunes prompt-selection parameters and a UCB-Tuned inner loop adapts prompt choices during decoder-only fine-tuning. The authors introduce an organ-scaled negative-point ring and a bounded Dice--MSD--HD95 composite reward. Evaluation on a pooled TRUS benchmark reports Dice 0.982, HD95 0.482, and MSD 0.204, with a large improvement over conventional baselines, and an additional external PFUS1 result is reported. The central claim is that adaptive prompt allocation improves boundary-sensitive multi-organ ultrasound segmentation without modifying the foundation-model backbone.","tokens_in":13319,"tokens_out":6253,"duration_ms":65774,"significance":"If the method were fully specified and evaluated as an automatic segmentation approach, the per-organ bandit formulation and the boundary-sensitive reward would be a useful contribution to promptable medical image segmentation. The paper has several strengths: it clearly identifies a boundary-quality failure mode that Dice misses; it keeps the image and prompt encoders frozen, limiting trainable parameters; it evaluates on an external dataset; and it reports code availability. However, the central contribution as currently stated is not supported because the test-time prompt source is never described and the published protocol allows prompts to be derived from ground-truth masks. The PFUS1 external validation also lacks a fixed-prompt baseline, so the adaptive mechanism is not isolated. The ideas are promising, but the main evaluation needs substantial rework before the claims can be accepted.","major_comments":[{"comment":"The manuscript never specifies how prompts are generated for the held-out test sets. Section IV-B-2 states that ground-truth masks are used only to construct prompt coordinates (Eqs. 5 and 6), and no inference-time prompt generator is described. If the same protocol was applied to the test slices in Table III, the reported Dice/HD95/MSD are produced with oracle prompts derived from ground truth. In that case BAP-MOS is not a standalone segmentation method, and the comparison with U-Net and nnU-Net, which receive no prompts, is not a valid automatic-segmentation comparison. The authors must either describe and evaluate an image-based prompt generator at test time, or explicitly reframe the method as interactive promptable segmentation and compare with fixed-prompt MedSAM/SAM under the same user-provided prompt protocol.","section":"IV-B-2, V-B"},{"comment":"The external PFUS1 row reports BAP-MOS (MedSAM) with Dice 0.849, HD95 10.062, and MSD 5.034 against an FPN baseline that reports Dice only (0.710). No fixed-prompt MedSAM or fixed-prompt SAM result on PFUS1 is provided, so the claim that adaptive prompt allocation transfers to the external domain is not isolated; the gain could come from decoder fine-tuning or from the oracle-prompt protocol rather than from adaptive selection. Add fixed-prompt baselines on PFUS1 with the same prompt source as BAP-MOS.","section":"Table III"},{"comment":"The outer TPE objective (Eq. 2) minimizes validation MSD and the inner reward (Eq. 8) is a weighted combination of Dice, MSD, and HD95, which are also the final evaluation metrics in Table III. Selecting hyperparameters and prompt-selection policies on validation versions of the same metrics can inflate the reported improvements relative to baselines whose hyperparameters are not tuned on those metrics. The authors should report the sensitivity of the results across independent validation splits and state the hyperparameter-tuning protocol used for U-Net, nnU-Net, SAM, and MedSAM baselines.","section":"IV-A, Eq. (8)"}],"minor_comments":[{"comment":"The sentence 'The parameter W specifies the sliding reward-memory length to track non-stationary distributions to track non-stationary distributions' repeats the phrase 'to track non-stationary distributions' and should be corrected.","section":"IV-A"},{"comment":"There is a typo in 'V alid measurements' near Eq. (8); it should read 'Valid measurements'.","section":"IV-B-2"},{"comment":"The plot legend uses 'BAPMOS_SAM' and 'BAPMOS_MEDSAM' without hyphens; these should be made consistent with the notation 'BAP-MOS (SAM)' used in the text and tables.","section":"Fig. 3"},{"comment":"The rows 'SAM' and 'MedSAM' in Table III should be explicitly labeled as fixed-prompt baselines, matching the abstract and the experimental design, so the reader can identify the exact comparison protocol.","section":"Table III"}],"recommendation":"major_revision","confidential_remarks":"The oracle-prompt concern is the central issue: as written, the paper does not establish that BAP-MOS is a deployable automatic segmentation method. The authors should consider reframing the contribution as interactive promptable segmentation and adjusting the baselines accordingly, or adding an inference-time prompt generator. The PFUS1 evaluation is also currently incomplete as a test of adaptive prompt selection because the fixed-prompt control is missing. The metric-selection concern, while less severe, should be addressed with additional ablation or a clearer statement of baseline tuning. The paper's core idea is defensible, but the evaluation protocol must be substantially revised."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The adaptive-prompting idea here is worth a look, and the paper does some things well. The empirical observation in Table II—that prompt schedule moves HD95 and MSD by roughly 95–98% while Dice stays flat—is a genuinely useful point for anyone working with promptable segmenters. The per-organ bandit formulation with UCB-Tuned, block-wise decisions, and the organ-scaled negative ring is a new combination, and the paper documents the search cost and learned preferences rather than just reporting one number. Keeping the encoders frozen and only updating the decoder is also a sensible design choice. But the load-bearing flaw is real. Section IV-B-2 says ground-truth masks are used to construct prompt coordinates, and Section V-B reports test Dice/HD95/MSD without ever describing an inference-time prompt source. If test prompts come from GT, the comparison against U-Net and nnU-Net is apples-to-oranges, and the headline 48% HD95 / 45% MSD reductions are not evidence for a standalone segmentation method. The PFUS1 row doesn't repair this: it lacks a fixed-prompt MedSAM baseline, so you can't isolate the adaptive mechanism there either. There's also an unexplained inconsistency between Tables I and II—Box multi-organ HD95 is 4.21 in one and roughly 0.5 in the other—that the text never addresses. The outer TPE loop optimizing validation MSD and the inner reward using MSD/HD95 is circular to some degree, though separate held-out test sets soften that; it means the gains are partly the expected result of tuning to the metric being reported. This paper is for people working on promptable segmentation or ultrasound evaluation. It deserves a serious referee because the mechanism is plausible and the motivation is sound, but as written the central claim doesn't hold up. I'd recommend reject or major revision: the authors need to either add an image-based test-time prompt generator or explicitly reframe as an interactive/oracle-prompt study, and add a fixed-prompt MedSAM baseline on PFUS1. If they fix that, the idea may be worth publishing; right now it's a cautionary example of evaluation gaps.","headline":"The adaptive-prompting idea is interesting and the motivation is real, but the reported gains are not evidence for a deployable method: test-time prompts appear to come from ground-truth masks, and no inference-time prompt source is described.","tokens_in":693,"tokens_out":788,"would_cite":false,"duration_ms":39089,"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":"BAP-MOS claims that choosing prompts per organ with a bandit during decoder-only fine-tuning cuts boundary errors by roughly 48% in multi-organ ultrasound segmentation.","keywords":["multi-organ segmentation","ultrasound","adaptive prompting","multi-armed bandit","boundary-aware reward","promptable foundation models"],"falsifier":"Run the same benchmark with test prompts generated by a mask-free automatic prompt generator (e.g., a detector or saliency method) instead of ground-truth masks, and check whether the HD95 and MSD gains over fixed-prompt baselines persist; if they vanish, the reported improvement depends on oracle masking at test time.","tokens_in":12785,"feed_emoji":"🩺","tokens_out":6156,"duration_ms":52507,"temperature":0.7,"pith_summary":"The paper claims that the choice of prompt type—box, point, or a combination—is a major and under-used lever for multi-organ ultrasound segmentation, and that choosing prompts per organ with a bandit algorithm during decoder-only fine-tuning sharply improves boundary accuracy. On a pooled prostate-region TRUS benchmark, the method reports Dice 0.982, HD95 0.482 mm, and MSD 0.204 mm, roughly 48% lower HD95 and 45% lower MSD than the strongest conventional baseline. The authors argue this shows adaptive prompt allocation can improve boundary-sensitive segmentation without modifying the underlying foundation model.","feed_headline":"Adaptive prompting cuts boundary error 48% in multi-organ ultrasound","feed_subtitle":"Per-organ bandit selects box, point, or combined prompts during decoder fine-tuning, improving HD95 and MSD.","key_machinery":"The central object is the per-organ UCB-Tuned prompt selector running on a decision-block schedule, with an organ-scaled negative-prompt ring. The ring radius $r_o = \\max\\left(r_{\\min}, \\left\\lfloor \\alpha \\sqrt{A_o} \\right\\rfloor\\right)$ places negative points just outside each organ, and the bounded reward $R_o = 1 - \\left[0.3(1-\\mathrm{Dice}_o) + 0.5 \\frac{\\min(\\mathrm{MSD}_o, \\tau)}{\\tau} + 0.2 \\frac{\\min(\\mathrm{HD95}_o, 3\\tau)}{3\\tau}\\right]$ clips boundary distances so the bandit's variance cap applies. This machinery lets the model adapt which prompt guides each organ during fine-tuning while the image and prompt encoders stay frozen.","core_discovery":"BAP-MOS treats prompt selection as an organ-specific sequential decision problem. An inner UCB-Tuned loop chooses among box, point, and combined prompts for each organ, holding the choice fixed for $K$ training batches and updating reward statistics from a bounded Dice–MSD–HD95 validation probe; an outer Tree-structured Parzen Estimator searches over prompt-geometry and reward parameters. The key reported findings are that fixed prompt schedules produce up to 98% variation in HD95 while Dice moves only 2.5%, and that the adaptive selector converges to a stable preference for box and combined prompts while improving boundary metrics over all fixed baselines.","pith_inferences":["Because the paper never specifies an inference-time prompt source, a natural extension is to pair BAP-MOS with an automatic prompt predictor; the reported gains may shrink if prompt coordinates are noisier than ground truth.","A fair evaluation would include an oracle-prompt upper bound and a mask-free-prompt lower bound to separate the benefit of adaptation from the benefit of knowing the mask.","The bandit-over-prompt structure could transfer to other promptable foundation models and modalities where boundary metrics matter, such as CT or MRI multi-organ segmentation, with a reward re-weighted toward the clinically relevant distance."],"forward_implications":["Decoder-only fine-tuning plus bandit prompt choice meets or beats full-model fine-tuning at a fraction of the parameter cost.","HD95 and MSD can vary by roughly 95–98% across fixed prompt schedules even when Dice changes by only 2.5%, so overlap-only evaluation is insufficient for promptable ultrasound segmentation.","Per-organ prompt preferences are learnable and stable: the selector converges to box/combined dominance, suggesting that a well-chosen fixed schedule may recover much of the gain.","The framework is backbone-agnostic: adapting MedSAM with BAP-MOS closes most of the gap to SAM on the TRUS benchmark, and the same approach shows improvement on an external pelvic-floor ultrasound corpus."],"supporting_citations":[{"why":"Supplies the promptable SAM backbone and the box/point prompt interface BAP-MOS adapts.","marker":"[13]"},{"why":"Provides the MedSAM medical foundation model used as an alternative backbone and as the external PFUS1 adaptation target.","marker":"[17]"},{"why":"Defines the UCB-Tuned bandit algorithm that runs the inner per-organ prompt selection loop.","marker":"[16]"},{"why":"Provides the TPE hyperparameter optimizer used in the outer loop to search the prompt-selection parameter vector.","marker":"[38]"},{"why":"Supplies the U-Net baseline, the strongest conventional comparison on the TRUS benchmark.","marker":"[8]"},{"why":"Supplies the nnU-Net self-configuring pipeline baseline that BAP-MOS must beat.","marker":"[10]"},{"why":"Provides the external PFUS1 pelvic-floor ultrasound dataset used to test generalization.","marker":"[4]"},{"why":"Defines the MSD/HD95 boundary metrics that form the core evaluation and the reward signal.","marker":"[5]"}],"fun_headline_variants":["Bandit-based prompts cut boundary error 48% in ultrasound","Adaptive prompt selection boosts multi-organ boundary accuracy","Closed-loop prompting reduces HD95 by 48% for multi-organ ultrasound","Per-organ bandit chooses prompts, slicing boundary error nearly in half","Adaptive prompt bandit improves boundary metrics in multi-organ ultrasound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that ground-truth masks are available to construct the box, point, and combined prompt coordinates used during evaluation; the paper states this for training but never specifies an inference-time prompt source.","fun_headline_variants_meta":{"raw":{"variants":["Bandit-based prompts cut boundary error 48% in ultrasound","Adaptive prompt selection boosts multi-organ boundary accuracy","Closed-loop prompting reduces HD95 by 48% for multi-organ ultrasound","Per-organ bandit chooses prompts, slicing boundary error nearly in half","Adaptive prompt bandit improves boundary metrics in multi-organ ultrasound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000926,"raw_usage":{"total_tokens":3997,"prompt_tokens":1005,"completion_tokens":2992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":2906}},"tokens_in":621,"tokens_out":2992,"duration_ms":21508,"temperature":1.0,"reasoning_tokens":2906,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:17:31.054132+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same benchmark with test prompts generated by a mask-free automatic prompt generator (e.g., a detector or saliency method) instead of ground-truth masks, and check whether the HD95 and MSD gains over fixed-prompt baselines persist; if they vanish, the reported improvement depends on oracle masking at test time.","supporting_citations":[{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Supplies the promptable SAM backbone and the box/point prompt interface BAP-MOS adapts."},{"cited_title":"Segment anything in medical images,","cited_arxiv_id":null,"evidence_quote":"Provides the MedSAM medical foundation model used as an alternative backbone and as the external PFUS1 adaptation target."},{"cited_title":"Finite-time analysis of the multiarmed bandit problem,","cited_arxiv_id":null,"evidence_quote":"Defines the UCB-Tuned bandit algorithm that runs the inner per-organ prompt selection loop."},{"cited_title":"Algorithms for hyper-parameter optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the TPE hyperparameter optimizer used in the outer loop to search the prompt-selection parameter vector."},{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net baseline, the strongest conventional comparison on the TRUS benchmark."},{"cited_title":"nnu-net,","cited_arxiv_id":null,"evidence_quote":"Supplies the nnU-Net self-configuring pipeline baseline that BAP-MOS must beat."},{"cited_title":"Pfus1: Premier pelvic floor ultrasound segmenta- tion dataset. a resource for advancing research","cited_arxiv_id":null,"evidence_quote":"Provides the external PFUS1 pelvic-floor ultrasound dataset used to test generalization."},{"cited_title":"Metrics for evaluating 3d medi cal image segmentation: analysis, selection, and tool,","cited_arxiv_id":null,"evidence_quote":"Defines the MSD/HD95 boundary metrics that form the core evaluation and the reward signal."}],"review_version":1}