{"id":"7bb696c0-f2a9-4d12-bd99-d4b2c152766e","arxiv_id":"2411.16123","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A one-shot SAM prompting framework that uses test-time image warping to generate mask, point, and box prompts, achieving the highest reported DICE scores among the compared baselines across five medical datasets.","lead":"Med-PerSAM adapts the Segment Anything Model (SAM) to medical images using a single labeled reference image plus a warping model that transfers the reference mask to each unlabeled test image, then feeds the resulting mask, point, and box prompts into SAM. It reports large DICE gains over one-shot SAM baselines on five medical datasets, but it achieves this by training the warping model on the unlabeled test set, a transductive step the 'one-shot' label undersells.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported margin is not yet supported as one-shot: Med-PerSAM trains a 19M-parameter warping model on the unlabeled test set and tunes point counts by test DICE, while PerSAM/Matcher are inference-only; only a held-out or controlled comparison can verify the claim.","rationale":"The claim under test is the empirical performance claim in Table 1: DICE scores of 92.0, 87.8, 74.2, 91.3, and 72.9 demonstrate a large, reliable gain from one-shot visual prompt engineering, with no training on evaluation data. For this claim to hold, the evaluation protocol must isolate the contribution of prompt engineering and must not let the proposed method see the test set differently from its baselines. The most load-bearing flaw is not that warping can fail on extreme cases (the paper concedes this in the Conclusion), but that every reported number is generated under a protocol that is asymmetric between Med-PerSAM and PerSAM/Matcher. The warping model is trained on the unlabeled test set (Sec. 3.1), and the point-prompt count is chosen by test DICE (Appendix C.2). This is a direct threat to the central empirical claim because the comparison is transductive versus inductive. A held-out evaluation is a simple, decisive check. If Med-PerSAM retains its margin under a fair split, the conditional acceptance is confirmed; if not, the headline overclaims. I agree with much of the reader's analysis, but I would frame the load-bearing assumption as evaluation fairness rather than flow accuracy: the warping caveat only limits scope, whereas the transductive protocol potentially inflates all five reported gains. The paper does have genuine strengths: extensive ablations, averaging over ten reference samples, cross-dataset robustness checks, and perturbation tests. These indicate the method works, but they do not quantify how much of the margin comes from test-time training and test-set hyperparameter selection. The reader's conditional verdict is appropriate, and my concern reinforces that condition rather than moving the verdict in a new direction.","tokens_in":21303,"tokens_out":4802,"duration_ms":45181,"concrete_test":"For each of the five datasets, randomly split patients into an adaptation pool A and a held-out evaluation pool E (e.g., 70/30, stratified by anatomy), with the reference image drawn only from A. Run all hyperparameter searches (point count K, loss coefficients, kernels, epochs) and all warping-model training/retraining using A alone, then compute DICE on E once. Repeat over 5 splits. If Med-PerSAM's held-out DICE drops below the PerSAM/Matcher baselines or falls by more than roughly 5 points relative to Table 1, the reported margin is substantially an artifact of transductive training and test-set tuning; if it holds, the one-shot claim is strengthened. As a secondary control, allow PerSAM and Matcher the same unlabeled test-image access (e.g., self-training) to isolate the prompt-engineering contribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim that Med-PerSAM achieves large DICE gains as a one-shot, training-free SAM adaptation (Sec. 4.4), the evaluation must be fair and must not use the evaluation data. Two linked practices undermine this. First, the warping model fθ (NICE-Trans-based, ~19M parameters) is trained on the unlabeled test images themselves before SAM inference: Sec. 3.1 states that \"training takes place between the reference sample and the test set, similar to the setting of test-time training/adaptation,\" and Eq. 3 minimizes Ltrain over the test images I_i. Med-PerSAM is therefore transductive: it has seen every test image as a training example. PerSAM and Matcher are pure inference baselines and never train on the test set, so the reported margins (up to 65% over Matcher) conflate prompt engineering with test-time adaptation. Second, hyperparameters are tuned on the same test set: Appendix C.2 states that the number of point prompts was selected by varying K and choosing the setting that maximizes DICE on test images (Fig. 7), and point counts differ by dataset in Sec. 4.3; loss coefficients (Table 12), augmentation ranges (Table 13), and kernel sizes (C.1) are also dataset-specific. On small cohorts (CAMUS, JSRT, BUU), even a few tuned choices can inflate scores. The conclusion's caveat about pose/scan-range variation limits scope but does not fix the comparison asymmetry. Thus the headline \"superior performance without additional training\" is not currently supported; a fair evaluation must either hold out the test split or give baselines the same transductive access.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Med-PerSAM, a one-shot visual prompt tuning framework for SAM in medical image segmentation. Given a single reference image and its mask plus an unlabeled test set, a lightweight warping model (about 19M parameters, based on NICE-Trans) is trained to align the reference to each test image, producing a warped mask that serves as a mask prompt; point and box prompts are then extracted via morphological operations and similarity-based subregion selection, and SAM's iterative predictions are used to refine prompts and to retrain the warping model with pseudolabels. Experiments on five 2D medical datasets report large DICE gains over PerSAM, Matcher, and other foundation-model baselines, with additional ablations, cross-dataset tests, and perturbation robustness studies. The central claim is that Med-PerSAM achieves superior performance 'without additional fine-tuning of SAM or manual visual prompts' in a one-shot setting.","tokens_in":21508,"tokens_out":3824,"duration_ms":37831,"significance":"If the reported gains were obtained under the claimed protocol, the paper would make a useful contribution: automated visual-prompt generation for SAM in medical imaging, a lightweight prompt-tuning mechanism, and an iterative refinement loop that appears to improve robustness. The manuscript is also commendably thorough in its ablations, including multiple reference samples (Table 2), cross-dataset reference swaps (Table 5), perturbation studies (Tables 6 and 15), and mask-prompt perturbation experiments (Fig. 6). However, the headline claim is not currently supported because the method trains on the unlabeled test set (transductive) and tunes key hyperparameters on test DICE, whereas the main baselines are inference-only. The comparison in Table 1 therefore conflates test-time adaptation and test-set selection with one-shot, training-free adaptation. With a corrected evaluation protocol, the method could still be valuable, but it would more accurately be characterized as a transductive/test-time-adaptation approach rather than a standard one-shot inference method.","major_comments":[{"comment":"The warping model fθ is trained on the unlabeled test images before SAM inference. Section 3.1 states that 'training takes place between the reference sample and the test set, similar to the setting of test-time training/adaptation', and Eq. (3) minimizes Ltrain over the test images I_i. Consequently, every test image is seen as a training example, whereas PerSAM and Matcher are pure inference baselines that never train on the test set. The large margins in Table 1 therefore reflect an asymmetric comparison: Med-PerSAM receives test-time adaptation of a 19M-parameter model, while the baselines do not. To support the claim of one-shot, training-free superiority, the authors should either evaluate on a held-out split with a separate validation set, compare against transductive/test-time-adaptation baselines under the same protocol, or explicitly re-frame the method and claims as test-time adaptation rather than inference-only prompting.","section":"Sec. 3.1, Eq. (3); Sec. 4.3; Sec. 4.4"},{"comment":"Several hyperparameters are selected using the test set itself. Appendix C.2 says the number of point prompts K was chosen by varying K and measuring DICE on test images (Fig. 7), and Section 4.3 reports different K values per dataset. In addition, loss coefficients (Table 12), augmentation ranges (Table 13), and morphological kernel sizes (Appendix C.1) are dataset-specific. On small test cohorts such as CAMUS, JSRT, and BUU, even a small number of such choices can inflate reported scores. The authors should demonstrate that the reported performance is stable under a fixed hyperparameter configuration, or use a held-out validation set for selection, clearly separated from the final test evaluation.","section":"Appendix C.2, Fig. 7; Sec. 4.3; Appendix B.1, Tables 12-13; Appendix C.1"},{"comment":"The iterative retraining loop uses SAM's own predictions as pseudolabels to retrain the warping model, after which those same SAM predictions are refined again. Since the final metrics are computed against external ground truth, this is not a logical circularity, but it creates a risk that the loop reinforces SAM's inductive biases rather than learning genuinely better prompts. The paper should report results with the retraining component disabled (i.e., using only the initially trained warping model and prompt generation, with and without prompt refinement) to quantify how much of the final gain is attributable to the pseudolabel loop, and ideally verify on a held-out set that the improvements from retraining generalize rather than reflecting adaptation to the specific test images.","section":"Sec. 3.3; Sec. 4.7; Fig. 6"}],"minor_comments":[{"comment":"There are repeated typos: 'signficantly' should be 'significantly', and the sentence 'our model model consistently achieves reliable results' contains a duplicated word.","section":"Sec. 4.4"},{"comment":"The heading 'Hopkin's statistics' should be 'Hopkins' statistics'. Also, the exclusion of Matcher from Table 4 is justified by different point numbers, but the comparison of clustering tendency would be more informative if Matcher were evaluated with the same number of points as the other methods.","section":"Sec. 4.5, Table 4"},{"comment":"The x-axis label 'N umber of Points' contains an awkward space; it should read 'Number of Points'.","section":"Fig. 7"},{"comment":"The phrase 'without additional fine-tuning' is ambiguous: SAM itself is indeed frozen, but the 19M-parameter warping model is trained on the test set. The wording should be clarified in the abstract and throughout, e.g., 'without fine-tuning SAM' rather than 'without additional training of the pretrained SAM or human intervention', which suggests no training occurs at all.","section":"Abstract; Sec. 4.3"},{"comment":"The statement that replacing SAM with SAM2 yields 'similar performance' is supported by Table 14, but the discussion would benefit from reporting whether the differences are statistically significant given the small per-dataset differences.","section":"Sec. 4.3; Appendix B.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically solid and the ablations are unusually thorough, but the evaluation protocol is the main obstacle. The authors appear to be aware of the transductive nature of their method ('similar to the setting of test-time training/adaptation'), yet the main framing and the comparison against inference-only baselines do not account for it. If the authors can provide a held-out evaluation with fixed hyperparameters, or a fair transductive comparison, the contribution would likely be publishable as a test-time-adaptation approach. In its current form, the headline claim of one-shot, training-free superiority is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi colleague,\n\nHere's my take on Med-PerSAM. The core idea is a solid step: instead of using pixel-similarity to place SAM prompts like PerSAM and Matcher, they train a lightweight deformable registration network to warp the single reference mask onto every test image, feed the warped mask plus point and box prompts into SAM, then iteratively retrain the warper using SAM outputs as pseudolabels. That is genuinely new relative to the PerSAM/Matcher line, and the paper backs it up with unusually thorough ablations: prompt-type combinations, point-prompt placement via subregions, augmentation loss, affine vs deformable, robustness to input perturbations, and cross-dataset reference changes. If the reported numbers are trustworthy, the method itself seems to work.\n\nBut the evaluation has a load-bearing asymmetry. The warping model is trained on the unlabeled test set before inference (Sec 3.1: 'training takes place between the reference sample and the test set'); it's transductive. PerSAM and Matcher are pure inference, so this is not an apples-to-apples comparison. On top of that, the number of point prompts is explicitly selected by maximizing DICE on the test set (Appendix C.2, Fig 7), and other hyperparameters like loss coefficients and augmentation ranges are dataset-specific. On small cohorts like CAMUS, JSRT, and BUU, that kind of tuning can inflate gains. The conclusion's caveat about extreme heterogeneity is a real scope limit, but it doesn't fix the comparison asymmetry.\n\nThe paper's own language is also sloppy: 'training-free SAM scheme' is true only for SAM itself, but a 19M-parameter warping model is trained on the test set. That is not 'one-shot' in the usual sense, and the abstract's 'eliminates the need for additional training' is misleading.\n\nThat said, the method is useful for what it is: a transductive, test-time-adaptation approach that needs only one labeled scan per dataset. The fix is straightforward - hold out a test split for warping-model training or give the baselines the same transductive access - and I'd like to see that revision. The paper deserves a serious referee; the core idea and ablation depth are worth the community's time. It just shouldn't be published with the current claim.","headline":"The method is promising and well-ablated, but the headline 'one-shot, training-free' claim is not supported because the warping model trains on the full unlabeled test set and some hyperparameters are tuned on test DICE, making the comparison with inference-only baselines unfair.","tokens_in":22208,"tokens_out":4205,"would_cite":false,"duration_ms":33183,"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":"With a single labeled medical image and a set of unlabeled test images, Med-PerSAM obtains one-shot segmentation from a frozen Segment Anything Model by warping the reference mask into a dense visual prompt and iteratively refining it…","keywords":["one-shot segmentation","Segment Anything Model","visual prompt tuning","medical image segmentation","image warping","deformable registration","test-time adaptation","prompt refinement"],"falsifier":"Run Med-PerSAM on a medical dataset with deliberately heterogeneous capture poses or scan ranges, such as spinal CT slices spanning different vertebral levels or chest X-rays with strong rotation and zoom beyond the augmentation range; if the optical flow cannot align the reference mask to the test image, the warped mask prompt will mislead SAM and the reported DICE margin over PerSAM and Matcher should collapse or reverse.","tokens_in":20950,"feed_emoji":"🩻","tokens_out":6617,"duration_ms":60361,"temperature":0.7,"pith_summary":"The paper tries to show that the Segment Anything Model can be made to segment medical images from a single labeled example without fine-tuning SAM and without manual prompts. The core move is to train a lightweight warping model, using only the reference image, its mask, and the unlabeled test images, that warps the reference mask onto each test image; that warped mask becomes a dense visual prompt from which point and box prompts are derived. SAM's outputs are then used as pseudolabels to retrain the warping model, creating an iterative loop. On five 2D medical datasets this yields DICE scores from 72.9 to 92.0, outperforming both prior one-shot SAM methods and generalist segmentation foundation models. If the claim holds, practical one-shot medical segmentation becomes available to users who cannot craft anatomical prompts.","feed_headline":"One-shot SAM hits DICE 92 with no fine-tuning","feed_subtitle":"A warping model turns one labeled medical image into prompts that beat PerSAM and Matcher on five datasets.","key_machinery":"The load-bearing object is the warping model $f_\\theta$, a lightweight affine-plus-deformable registration network based on NICE-Trans and adapted from 3D to 2D. It is trained with a warping loss $L_{\\text{warp}} = L_{\\text{img}}(\\hat{I}_{\\text{ref}\\to i}, I_i) + L_{\\text{reg}}(\\phi_{\\text{ref}\\to i})$ using SSIM or NCC similarity plus flow smoothness regularization, and an augmentation loss $L_{\\text{aug}}$ that adds a DiceCE segmentation term between the warped augmented mask and the augmented mask. The trained model produces the optical flow $\\phi_{\\text{ref}\\to i}$ and the warped mask $\\hat{M}_{\\text{ref}\\to i} = M_{\\text{ref}} \\circ \\phi_{\\text{ref}\\to i}$, which serves as the mask prompt for SAM. Erosion and dilation of that mask define candidate regions; a class-prototype cosine-similarity map $S_i$ selects $K$ positive points from eroded subregions and $K$ negative points from the dilated-minus-eroded ring, and $g_{\\text{box}}$ returns the minimal enclosing rectangle. SAM receives point, box, and mask prompts together, its output becomes a new mask prompt, and after inference the SAM prediction is used as a pseudolabel to retrain the warping model through $L_{\\text{retrain}} = L_{\\text{train}} + L_{\\text{seg}}(\\hat{M}_{\\text{ref}\\to i}, \\hat{M}_i)$. Iterating this loop transfers SAM's visual acuity back into the warping model and progressively improves the prompts.","core_discovery":"The central claim is that the poor performance of one-shot SAM in medical imaging comes from placing point prompts through pixel- or feature-similarity, which misplaces points in grayscale images where organs share intensities and clusters them around confident regions. Med-PerSAM claims to remove this bottleneck by learning the optical flow $f_\\theta$ that registers the single reference image to each test image, warping the reference mask accordingly to produce a dense mask prompt, and extracting positive and negative point prompts from eroded and dilated subregions of that mask plus a box prompt. Using these automatically generated visual prompts, the frozen SAM reaches DICE scores of 92.0, 87.8, 74.2, 91.3, and 72.9 on Shenzhen, OdontoAI, CAMUS, JSRT, and BUU, outperforming PerSAM, Matcher, and every other compared foundation model on all five datasets without any fine-tuning of SAM or human intervention.","pith_inferences":["Beyond the paper: because the warping model is trained on the unlabeled test set, this is a test-time adaptation scheme rather than a zero-shot predictor; deploying it on a single arriving image without access to the rest of the test set would require a different training procedure.","Beyond the paper: the registration bottleneck suggests a concrete extension where an explicit spatial transformer or a small atlas of multiple reference samples is used to rescue the mask prompt when pose or scan-range variation is extreme, a variant that could be tested directly on the failure cases the paper acknowledges.","Beyond the paper: if the warping-to-prompt loop is the true source of the gain, swapping the 2D warper for a 3D one should extend the same mechanism to volumetric medical data, an open direction the paper lists as future work.","Beyond the paper: the comparison with PerSAM and Matcher depends on the reference-sample selection protocol, so an independent re-run across many random reference samples, rather than the ten-sample average reported, would clarify how much of the advantage is robust to reference choice."],"forward_implications":["Frozen SAM can be personalized for a medical dataset from one labeled image plus unlabeled images, so no task-specific fine-tuning of the 641M-parameter model is needed.","Dense mask prompts derived from registration are more reliable than point prompts placed by pixel similarity; the subregion-based point extraction removes the clustering and misplacement problems that hurt PerSAM and Matcher.","Iterative retraining with SAM outputs as pseudolabels gives the framework a self-correction capacity; when the initial mask prompt is deliberately perturbed by a factor of 0.2, repeated retraining largely closes the performance gap.","The method is relatively robust to domain shift: changing the reference sample to one from a different chest X-ray dataset barely changes DICE, and test-sample perturbations degrade Med-PerSAM less than the strongest baselines.","Extension to multi-class segmentation is direct via one-hot mask prompts per foreground class, so the same warping-plus-prompt loop applies beyond binary organs."],"supporting_citations":[{"why":"Supplies the frozen Segment Anything Model whose visual prompts are engineered in this paper.","marker":"[27]"},{"why":"The main one-shot SAM baseline whose pixel-similarity point-prompt strategy Med-PerSAM replaces.","marker":"[68]"},{"why":"The second one-shot SAM baseline, using DINOv2 feature matching, which Med-PerSAM outperforms.","marker":"[31]"},{"why":"Provides the NICE-Trans architecture that the warping model is based on.","marker":"[35]"},{"why":"Establishes deformable inter-patient registration as a feasible approach, the premise underlying the warping of reference to test images.","marker":"[4]"},{"why":"Supplies the Shenzhen chest X-ray dataset used for lung segmentation experiments.","marker":"[22]"},{"why":"Supplies the CAMUS cardiac ultrasound dataset used for heart segmentation experiments.","marker":"[30]"},{"why":"Supplies the OdontoAI teeth segmentation dataset used in the experiments.","marker":"[54]"},{"why":"Supplies the JSRT chest X-ray dataset used for multi-class lung and heart segmentation.","marker":"[14]"},{"why":"Supplies the BUU lumbar spine segmentation dataset used in the experiments.","marker":"[28]"}],"fun_headline_variants":["Warped prompts boost one-shot SAM in medical imaging","No fine-tuning: Med-PerSAM beats SAM baselines on 5 datasets","One label, warped mask: Med-PerSAM lifts SAM's DICE to 92"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that one reference image can be warped, pixel by pixel, into each unlabeled test image so that the reference mask rides along and lands on the target organ; when capture pose or scan range varies too much for such warping, the prompt misaligns and the whole chain breaks.","fun_headline_variants_meta":{"raw":{"variants":["Warped prompts boost one-shot SAM in medical imaging","No fine-tuning: Med-PerSAM beats SAM baselines on 5 datasets","One label, warped mask: Med-PerSAM lifts SAM's DICE to 92"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000576,"raw_usage":{"total_tokens":2732,"prompt_tokens":974,"completion_tokens":1758,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":1694}},"tokens_in":590,"tokens_out":1758,"duration_ms":12206,"temperature":1.0,"reasoning_tokens":1694,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:32:54.315130+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Med-PerSAM on a medical dataset with deliberately heterogeneous capture poses or scan ranges, such as spinal CT slices spanning different vertebral levels or chest X-rays with strong rotation and zoom beyond the augmentation range; if the optical flow cannot align the reference mask to the test image, the warped mask prompt will mislead SAM and the reported DICE margin over PerSAM and Matcher should collapse or reverse.","supporting_citations":[{"cited_title":"Non-iterative coarse-to-fine transformer net- works for joint affine and deformable image registration","cited_arxiv_id":null,"evidence_quote":"Provides the NICE-Trans architecture that the warping model is based on."},{"cited_title":"V oxelmorph: a learning framework for deformable medical image registration","cited_arxiv_id":null,"evidence_quote":"Establishes deformable inter-patient registration as a feasible approach, the premise underlying the warping of reference to test images."},{"cited_title":"Two public chest x-ray datasets for computer-aided screening of pulmonary diseases","cited_arxiv_id":null,"evidence_quote":"Supplies the Shenzhen chest X-ray dataset used for lung segmentation experiments."},{"cited_title":"Deep learning for segmentation using an open large-scale dataset in 2d echocardiography","cited_arxiv_id":null,"evidence_quote":"Supplies the CAMUS cardiac ultrasound dataset used for heart segmentation experiments."},{"cited_title":"Improving anatomical plausibility in medical image segmentation via hybrid graph neural networks: applications to chest x-ray analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the JSRT chest X-ray dataset used for multi-class lung and heart segmentation."},{"cited_title":"Buu-lspine: A thai open lumbar spine dataset for spondylolisthesis detection","cited_arxiv_id":null,"evidence_quote":"Supplies the BUU lumbar spine segmentation dataset used in the experiments."}],"review_version":1}