{"id":"e56cbd18-a910-4d49-a5f0-f5743e2e82df","arxiv_id":"2502.00630","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Self-Prompt-SAM automatically generates point, box, and mask prompts for a fine-tuned SAM and reports state-of-the-art Dice scores on three medical segmentation benchmarks.","lead":"Self-Prompt-SAM adapts SAM, a general-purpose image segmentation model, to 3D medical scans by training a small module to generate its own point, box, and mask prompts, and by adding depth-aware adapters to the frozen backbone. The authors report Dice-score improvements over nnUNet of 2.3% on AMOS2022, 1.6% on ACDC, and 0.5% on Synapse.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'first SAM-based framework without prompts' claim is directly contradicted by the paper's own cited auto-prompt SAM methods, undermining a central contribution.","rationale":"The reader's weakest assumption focused on the transferability of the prompt-combination choice from ground-truth masks to automatically generated masks. That is a valid missing validation, but it concerns the optimality of a design choice rather than a direct contradiction. The 'first' claim is an explicit, verifiable assertion that is contradicted by the paper's own references, making it a more clear-cut load-bearing issue. If the 'first' claim is false, a central contribution collapses, and the SOTA claim is not positioned against the most relevant auto-prompt baselines. The reader's recommended CONDITIONAL verdict remains appropriate: the technical content may be salvageable, but the authors must correct the novelty claim, add comparisons with auto-prompt SAM methods, and address the prompt-transfer validation. I therefore leave the verdict unchanged while highlighting a different primary concern.","tokens_in":11441,"tokens_out":8955,"duration_ms":83676,"concrete_test":"Conduct a literature check: retrieve references [22] and [31]. If either describes a SAM-based medical segmentation method that generates prompts automatically (e.g., via a prompt generator or mask classification) without user-provided points or boxes, then the Section 1 'first' claim is false. To test the SOTA claim, re-run the comparison including MaskSAM and AutoSAM on ACDC, Synapse, and AMOS; if Self-Prompt-SAM does not outperform them, the headline 'state-of-the-art' claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 1 states: 'To the best of our knowledge, the proposed Self-Prompt-SAM is the first SAM-based image segmentation framework without any prompts provided.' Yet the same paper cites [22] 'Auto-prompting SAM for mobile friendly 3D medical image segmentation' and [31] 'MaskSAM: Towards auto-prompt SAM with mask classification for medical image segmentation,' the latter being the authors' own prior work. Both are SAM-based medical segmentation frameworks that generate prompts automatically and do not require manual prompts. This is an internal contradiction: either the cited works are not 'SAM-based image segmentation frameworks without prompts' (in which case they should not be cited as relevant related work), or the 'first' claim is false. No comparison with either method appears in any experiment table, so the claim to state-of-the-art performance is also not established against the most relevant auto-prompt competitors. This is not a stylistic issue; it directly undermines the paper's central novelty claim and renders the reported SOTA relative to a cherry-picked baseline set.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Self-Prompt-SAM, a framework that adapts the pre-trained 2D SAM to 3D medical image segmentation without manual prompts. It introduces a multi-scale prompt generator (MSPGenerator) that outputs auxiliary multi-class masks, from which box prompts, point prompts (selected by Euclidean distance transform), and mask prompts are derived. To handle 3D data, the authors inject a DFusedAdapter into each transformer block of the image encoder and mask decoder, add depth positional embeddings, an input modality adapter (MAdapter), and an MC-Adapter for multi-class output, while keeping the original SAM weights frozen. Experiments are reported on AMOS2022, Synapse, and ACDC with average Dice as the metric, and a nine-stage ablation (S1-S9) is performed on ACDC. The abstract and conclusion claim state-of-the-art results and a 2.3/0.5/1.6 point improvement over nnUNet on the three datasets.","tokens_in":11659,"tokens_out":8554,"duration_ms":76075,"significance":"The automatic-prompt pipeline is practically attractive, and the staged ablation in Table 4 is a genuine strength: it isolates the contribution of the prompt generator, the depth-fused adapter, depth positional embeddings, the modality adapter, and the multi-class adapter. If the experimental claims are verified, the paper would provide a useful recipe for prompt-free SAM deployment in volumetric medical imaging. The reported gains over strong baselines such as nnUNet are plausible in magnitude. However, the current version overstates its novelty and the AMOS evidence is confounded by the Table 1 row label; the central comparisons also lack statistical support and exclude the most relevant auto-prompt baselines.","major_comments":[{"comment":"The claim that Self-Prompt-SAM is \"the first SAM-based image segmentation framework without any prompts provided\" is internally inconsistent with the paper's own reference list. Reference [22] (Auto-prompting SAM for mobile friendly 3D medical image segmentation) and reference [31] (MaskSAM: Towards auto-prompt SAM with mask classification for medical image segmentation, by the same first author) are both SAM-based medical segmentation frameworks that generate prompts automatically and require no manual prompts. Neither method appears in any comparison table, and Appendix B's taxonomy does not place either method in its categories. The novelty claim should be narrowed to the specific mechanism (e.g., joint learning of the prompt generator with the mask decoder, or semantic multi-class output), and the state-of-the-art claim should be tested against these automatic-prompt baselines.","section":"Section 1, Contribution (i), and References [22], [31]"},{"comment":"The AMOS result row is labeled \"MaskSAM (Ours)\", not \"Self-Prompt-SAM (Ours)\", while the text and abstract attribute the 2.3% improvement over nnUNet to Self-Prompt-SAM. If the row is the authors' previously published MaskSAM model, then the AMOS claim does not concern the proposed method; if it is a labeling error, it must be corrected. The evaluation protocol also needs clarification: the table caption says \"AMOS testing dataset on the leaderboard\", whereas Section 3.1 says \"all results are based on 5-fold cross-validation without any ensembles\".","section":"Section 3.1, Table 1"},{"comment":"The optimal prompt combination (points + boxes + masks) and the distance-transform center-point rule are selected from experiments in which prompts are generated from ground-truth masks. In the deployed pipeline, prompts come from MSPGenerator auxiliary masks, which are predictions with errors. The paper does not report any experiment that compares prompt-type rankings under automatic prompts, nor does it quantify the gap between ground-truth prompts and automatic prompts. This is load-bearing because the prompt-design choice could change when the auxiliary masks are imperfect. Please add an ablation with MSPGenerator-generated prompts (e.g., points only, boxes only, masks only, and their combinations) and report auxiliary-mask accuracy.","section":"Section 2.4, Figure 1"},{"comment":"The comparisons report a single Dice value per method with no standard deviations, confidence intervals, or significance tests. The claimed Synapse advantage over nnUNet is 86.74 vs. 86.21 on a random 18/12 split, i.e., 0.53 points, which is within the range that could be explained by split variation. The claims of \"outperforming\" nnUNet and nnFormer are therefore not established at the reported precision. Provide multiple runs or repeated splits with mean±std, or significance tests, for at least the comparisons with margins below 1 point.","section":"Tables 2 and 3"}],"minor_comments":[{"comment":"The sentence \"our model surpasses SAMed, nnUNet, and nnFormer by 4.9%, 0.5%, 0.0%, and 0.7%\" gives four percentages for three methods; according to Table 2, the margins are 4.86, 0.53, and 0.17 points, so the sentence should be corrected.","section":"Section 3.1, Synapse paragraph"},{"comment":"The adapter name is rendered inconsistently as \"DFusedAdapter\", \"DfusedAdapter\", \"DFuserAdapter\", and \"vApdater\" (Table 4); please standardize.","section":"Sections 2.1 and 3.2"},{"comment":"The text refers to the Myo class \"in Figure 1(d)\", but Figure 1(d) shows depth incontinuity, not the central-point experiment; the reference should presumably be to Figure 1(b).","section":"Section 2.4, Figure 1"},{"comment":"The statement that for empty class channels \"we assign the values of both a box prompt and a point prompt to zero\" is unclear, since a zero-valued box or point is not a standard SAM prompt; clarify whether the class is skipped or prompted with a placeholder.","section":"Appendix A"},{"comment":"The ablation is performed only on ACDC; the conclusions about MSPGenerator and DFusedAdapter would be more convincing with at least one additional dataset, especially because the Synapse margins are small.","section":"Table 4 and Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The main editorial risk is the provenance of Table 1: if the \"MaskSAM (Ours)\" row is carried over from the authors' previous paper, the headline AMOS result is not for Self-Prompt-SAM. I would ask the authors to clarify this and to add comparisons with MaskSAM and Auto-prompting SAM before considering the paper. The \"first\" claim should also be checked against the cited auto-prompt literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper before anything else: the method itself is coherent and the ablation is informative, but the headline novelty claim is directly contradicted by the paper's own cited references, and the AMOS comparison table is labeled with a different method name. Those are not cosmetic issues.\n\nWhat is actually new: a multi-scale prompt generator (MSPGenerator) that turns SAM's internal features into auxiliary masks, from which boxes and distance-transform-selected points are derived; a depth-fused adapter (DFusedAdapter) that lets a 2D ViT see 3D context; and a small MC-Adapter for multi-class output. The ablation in Table 4 is the paper's strongest part—each added component gives a measurable step up on ACDC, which is real evidence that the design choices matter. The ACDC and Synapse results, if correct, are competitive: 93.26% on ACDC against nnUNet's 91.61%, and 86.74% vs 86.21% on Synapse. Those are modest but consistent gains.\n\nThe soft spots are in proportion to how much they undercut the central claims. First, the paper states it is \"the first SAM-based image segmentation framework without any prompts provided,\" yet cites [22] Auto-prompting SAM and [31] MaskSAM (the authors' own prior work) as automatic-prompt SAM methods for medical segmentation. That is an internal contradiction. No comparison with either method appears in any table, so the state-of-the-art claim is not established against the most relevant competitors. Second, Table 1 — the AMOS leaderboard result — labels the row \"MaskSAM (Ours)\" rather than Self-Prompt-SAM. That is either a typo or a conflation with prior work, and it makes the 2.3% gain over nnUNet hard to trust as a headline result for this paper. Third, there are no error bars or significance tests; differences of 0.5–1.6% could be noise. Fourth, the prompt combination (points+boxes+masks) and the center-point rule were chosen using ground-truth prompts, but the final pipeline uses automatic prompts from MSPGenerator, which contain errors. No experiment compares GT prompts with auto prompts in the full model, so the optimality of the rule under noisy masks is assumed, not shown.\n\nThis paper deserves a serious referee because the method is well-constructed and the ablation is solid. But it needs major revision before acceptance: rename or clarify Table 1, add quantitative comparisons with auto-prompt SAM baselines, address the \"first\" claim, and release code. I would not cite it in my own work yet.","headline":"A workable SAM adaptation with self-generated prompts, but the 'first' claim is contradicted by the paper's own citations and the AMOS table is mislabeled.","tokens_in":12201,"tokens_out":1669,"would_cite":false,"duration_ms":16597,"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":"The paper claims that a fully automatic prompt generator—multi-scale masks, boxes, and distance-transform-chosen center points—lets a frozen-weights SAM beat nnUNet on three public 3D medical segmentation benchmarks.","keywords":["medical image segmentation","Segment Anything Model","automatic prompt generation","parameter-efficient fine-tuning","3D adapters","multi-scale prompt generator","Euclidean distance transform"],"falsifier":"Compute Dice on ACDC with the automatic prompts replaced by ground-truth prompts and, separately, with point prompts selected randomly inside the predicted masks. If the gap between automatic and ground-truth prompts is small, or if the points+boxes+masks combination stops being optimal once masks contain errors, the central claim that the self-prompt design causes the gain collapses.","tokens_in":11221,"feed_emoji":"🩻","tokens_out":5587,"duration_ms":49272,"temperature":0.7,"pith_summary":"This paper tries to show that the Segment Anything Model (SAM), a prompt-driven segmentation foundation model trained on natural images, can be turned into a fully automatic 3D medical image segmenter without any human-supplied prompts. The authors build a self-prompt loop: a multi-scale prompt generator attached to SAM's image encoder produces auxiliary masks, which are then converted into box prompts, mask prompts, and point prompts placed at the mask's most central pixel by a Euclidean distance transform. To make the 2D model work on 3D volumes, they freeze all original SAM weights and insert lightweight depth-fused adapters, depth positional embeddings, and a multi-class adapter. They report that the resulting Self-Prompt-SAM surpasses nnUNet by 2.3% on AMOS2022, 1.6% on ACDC, and 0.5% on Synapse, which is the evidence for the claim that prompt-free SAM adaptation is viable and state of the art.","feed_headline":"Self-prompted SAM beats nnUNet on 3 medical benchmarks","feed_subtitle":"The model generates its own points, boxes, and masks from the image, and reports Dice gains of up to 2.3% over nnUNet.","key_machinery":"The load-bearing mechanism is the self-prompt loop formed by the MSPGenerator and the prompt encoder. The MSPGenerator is a hierarchical convolutional-and-transpose-convolutional decoder that takes five feature maps of different scales from SAM's image encoder and outputs auxiliary multi-class masks with deep supervision. Those masks are turned into three prompt types: one-hot mask prompts, bounding-box prompts, and a single point prompt selected by Euclidean distance transform as the pixel farthest from the mask boundary, i.e. the most central point. The other central block is the DFusedAdapter, an invert-bottleneck adapter with two extra fully connected layers processing the depth dimension and a skip connection, inserted after multi-head self-attention and in parallel with the MLP in every transformer block; it is what lets the frozen 2D SAM read 3D context.","core_discovery":"The central discovery claimed is that SAM's own internal features are enough to generate the prompts SAM needs, so the human can be removed from the loop entirely. Starting from multi-scale feature maps of the frozen image encoder, the MSPGenerator predicts auxiliary multi-class masks; those masks are encoded as one-hot semantic prompts, are used to compute bounding-box prompts, and are processed by an Euclidean distance transform to pick the point prompt farthest from any boundary. The authors further claim that among prompt combinations tried on ground truth, points plus boxes plus masks is the most robust, and that the resulting pipeline—with the DFusedAdapter injecting 3D depth information into every transformer block and the MC-Adapter mapping binary outputs to semantic classes—achieves state-of-the-art Dice on AMOS2022, ACDC, and Synapse while keeping every original SAM weight frozen.","pith_inferences":["A natural test the paper does not run is to feed ground-truth prompts into the same frozen-SAM pipeline and compare with the automatic prompts; the size of that gap would separate prompt-generation quality from SAM's own segmentation ability.","The distance-transform 'most central point' rule is a general recipe for point-prompt selection that could transfer to other promptable foundation models beyond SAM, wherever a coarse mask is available.","The same self-prompt architecture could be evaluated on non-CT modalities such as MRI and ultrasound, where the MAdapter's modality adaptation and the depth-fused 3D reasoning would face different intensity and anisotropy statistics.","If automatic prompts degrade on small or low-contrast organs, a promising extension is to let the mask decoder refine the auxiliary mask and re-generate prompts iteratively, something the current single-pass design does not do."],"forward_implications":["If the claims hold, clinicians never need to specify points or boxes: the model generates its own prompts from the image, which is the difference between a research demo and a usable tool.","Freezing all original SAM weights means the pre-trained zero-shot knowledge is retained, and only small adapter modules are trained, so adapting to a new medical dataset is parameter-efficient.","The reported margins over nnUNet—2.3% on AMOS2022, 1.6% on ACDC, 0.5% on Synapse—imply that prompt-free SAM adaptation can beat dedicated full-training segmentation pipelines on public benchmarks.","The same auxiliary masks carry semantic labels through one-hot encoding, giving SAM the multi-class output that binary-mask SAM natively lacks."],"supporting_citations":[{"why":"Supplies the base SAM model that the framework freezes and adapts.","marker":"[20]"},{"why":"The principal baseline that the reported gains (2.3%, 1.6%, 0.5%) are measured against.","marker":"[16]"},{"why":"One of the three benchmark datasets used for the leaderboard comparison.","marker":"[19]"},{"why":"The cardiac MRI benchmark dataset used for evaluation and ablation.","marker":"[2]"},{"why":"The third benchmark; the paper follows its 18/12 split from [6].","marker":"[21]"},{"why":"A leading SAM-based medical adaptation baseline that Self-Prompt-SAM claims to surpass (by 4.9% on Synapse).","marker":"[34]"},{"why":"The SAM-based 3D baseline compared on Synapse and ACDC.","marker":"[4]"},{"why":"Supplies the parameter-efficient fine-tuning method (adapter insertion) that the DFusedAdapter extends.","marker":"[13]"},{"why":"Supplies the adapter placement (after MSA, parallel to MLP) that DFusedAdapter follows.","marker":"[33]"}],"fun_headline_variants":["SAM learns to prompt itself, beats nnUNet on 3 benchmarks","Self-prompting SAM outdoes nnUNet on AMOS, ACDC, Synapse","No human prompts needed: SAM self-prompts and beats nnUNet","SAM generates its own prompts, then tops nnUNet on 3 sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The prompt recipe—boxes plus the most central point plus masks—was chosen using perfect ground-truth masks, and the whole method assumes that same recipe stays best when the auxiliary masks contain errors.","fun_headline_variants_meta":{"raw":{"variants":["SAM learns to prompt itself, beats nnUNet on 3 benchmarks","Self-prompting SAM outdoes nnUNet on AMOS, ACDC, Synapse","No human prompts needed: SAM self-prompts and beats nnUNet","SAM generates its own prompts, then tops nnUNet on 3 sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000603,"raw_usage":{"total_tokens":2821,"prompt_tokens":961,"completion_tokens":1860,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":1775}},"tokens_in":577,"tokens_out":1860,"duration_ms":12317,"temperature":1.0,"reasoning_tokens":1775,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T18:15:27.164212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute Dice on ACDC with the automatic prompts replaced by ground-truth prompts and, separately, with point prompts selected randomly inside the predicted masks. If the gap between automatic and ground-truth prompts is small, or if the points+boxes+masks combination stops being optimal once masks contain errors, the central claim that the self-prompt design causes the gain collapses.","supporting_citations":[{"cited_title":"NeurIPS (2022) 1, 4, 6","cited_arxiv_id":null,"evidence_quote":"One of the three benchmark datasets used for the leaderboard comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The cardiac MRI benchmark dataset used for evaluation and ablation."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"The third benchmark; the paper follows its 18/12 split from [6]."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Supplies the parameter-efficient fine-tuning method (adapter insertion) that the DFusedAdapter extends."}],"review_version":1}