{"id":"b46a3020-19a7-47ad-b465-62222e0af1f0","arxiv_id":"2507.23272","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"SAM2 can segment breast tumors in 3D MRI with a single bounding-box prompt, and center-outward propagation yields the best volumetric Dice.","lead":"This paper tests whether SAM2, a general AI image segmentation model, can outline breast tumors in 3D MRI scans using only one box drawn on a single slice. The authors find that starting from the middle of the tumor and propagating outward gives the most reliable outlines, and that providing a full mask instead of a box improves accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The mask-based Dice of 0.71 may be inflated because the ground-truth central-slice mask is used as the SAM2 prompt and then included in the volumetric Dice; the paper never excludes or discloses this.","rationale":"The reader's weakest assumption concerns alignment of MAMA-MIA masks with the extracted 2D slices, which is a valid reproducibility and correctness risk. However, a more direct threat to the reported numbers is the treatment of the prompted slice in the mask-based tracking experiment. If the ground-truth central-slice mask is included in the volumetric Dice, the 0.71 result is not a clean measure of SAM2's propagation ability, and the claimed benefit over bounding-box tracking is confounded. This is an internal protocol issue rather than an external data-alignment issue, and it can be settled by a straightforward recomputation without requiring new data. The bounding-box-based central claim, including the center-outward comparison, is not invalidated by this concern, so the existing conditional verdict remains appropriate, but the manuscript should clarify and, if necessary, revise the mask-based result. This is why the agreement with the reader is partial: we identify a distinct but related evaluation-validity concern.","tokens_in":7153,"tokens_out":14142,"duration_ms":172521,"concrete_test":"Recompute the mask-based tracking Dice after excluding the prompted central slice from both P and G, or equivalently report per-slice Dice with the initial slice removed. If the mean drops toward the 0.57 bounding-box level, the claimed improvement is an artifact of prompting with ground truth. The same exclusion should be applied to the bounding-box results so the two tracking modes are compared on an equal footing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3, 'Comparison to Mask-based Tracking', the authors report a mean volumetric Dice of 0.71 when SAM2 is given a segmentation mask on the central slice as input. The text does not explicitly state that this input mask is the MAMA-MIA ground-truth mask, but that is the only plausible source of a mask prompt in the described pipeline. Equation (1) defines volumetric Dice over the full volume, and the paper does not mention excluding the prompted slice from either P or G. If the central-slice mask is ground truth, that slice contributes a per-slice Dice of exactly 1.0 by construction, not by SAM2's segmentation ability. Because the central slice is chosen as the slice with the largest and clearest tumor, its contribution to the volume-level Dice is disproportionately large. The reported improvement from 0.57 (bounding box) to 0.71 (mask) may therefore be substantially an artifact of scoring the prompt itself rather than measuring propagation quality. The bounding-box results are not affected by this particular leakage, but the headline mask-based number and the claimed advantage of mask-based tracking over bounding-box tracking are directly dependent on this unresolved protocol detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using SAM2, a zero-shot vision foundation model for image and video segmentation, to segment breast tumors in 3D MRI by treating the volume as a sequence of 2D slices. Starting from a single bounding-box prompt on one slice, the authors compare three slice-wise propagation orders (bottom-to-top, top-to-bottom, center-outward) on 279 patients from the MAMA-MIA/Duke breast MRI dataset and report mean volumetric Dice scores of 0.57 for the best bounding-box strategy and 0.71 when an initial segmentation mask is provided on the central slice. They also analyze correlations between Dice and tumor size, volume, and slice count, and identify multi-focal small lesions as a common failure mode.","tokens_in":7379,"tokens_out":6226,"duration_ms":71786,"significance":"The study addresses an accessibility-oriented question with practical relevance: whether an open-source, training-free video segmentation model can provide useful interactive 3D tumor segmentation from minimal annotations. Its strengths include the use of a public benchmark with expert voxel-level masks, the relatively large 279-patient evaluation cohort, and the zero-shot setting with no training or fine-tuning, which makes the pipeline easy to reproduce conceptually. The failure-mode analysis (scattered small lesions) is clinically plausible and actionable. However, the quantitative conclusions are currently conditional on unresolved protocol details, most importantly the scoring of the mask-prompt experiment, and the lack of baselines and statistical tests makes the performance claims difficult to interpret. If the authors address these points, the result would be a solid empirical contribution to the workshop literature on foundational models in medical imaging.","major_comments":[{"comment":"The manuscript does not state the source of the segmentation mask used as the SAM2 prompt in the mask-based tracking experiment. If this mask is the MAMA-MIA ground-truth mask from the central slice, as the described pipeline strongly suggests, then Eq. (1) scores that prompted slice as part of the volume; because the prompt exactly equals ground truth, that slice contributes a per-slice Dice of 1.0 by construction. Since the central slice is selected as the slice with the largest and clearest tumor, this single slice can inflate the volumetric Dice substantially. The reported improvement from 0.57 to 0.71 is therefore not a clean measure of propagation quality. The authors must disclose the mask source and either exclude the prompted slice from both P and G in Eq. (1) or report Dice separately for propagated slices only; the abstract and conclusion should then be updated if the corrected number changes.","section":"Section 3, 'Comparison to Mask-based Tracking'; Eq. (1)"},{"comment":"The claim that center-outward propagation outperforms the other two strategies is based on histograms, boxplots, and per-patient winner counts, but no statistical test or confidence interval is provided. Given 279 paired observations, a paired test (e.g., Wilcoxon signed-rank) and an effect size should be reported. Moreover, the center slice is chosen as the slice with the largest and clearest tumor, so the comparison confounds propagation direction with prompt difficulty: initializing from an easier slice may explain the advantage even if propagation itself is not better. Please quantify the difference and discuss the selection protocol.","section":"Section 3, 'Quantitative Results' and Fig. 3"},{"comment":"All Dice scores depend on the exact alignment between the extracted 2D pre-contrast DCE-MRI slices and the MAMA-MIA expert voxel masks, but the paper does not validate this alignment; it references the extraction procedure of prior work [5,12,14,15]. If the masks were generated on a different contrast phase or at a different resolution/spacing, the volumetric Dice values would be systematically biased. The authors should either provide evidence of alignment (e.g., visual audit, registration metrics, or original spacing information) or treat this as a limitation and test the sensitivity of the results to the extraction parameters.","section":"Section 2.1"},{"comment":"The paper does not compare against any existing supervised or medical-image segmentation method (e.g., nnU-Net, UNETR, or a MedSAM variant) on the same 279-patient subset. Without such a baseline, the phrase 'strong segmentation performance' in the abstract and conclusion is not calibrated, and a reader cannot judge whether the proposed zero-shot workflow is competitive or merely convenient. Adding at least one reference baseline, or explicitly reframing the contribution as a zero-shot feasibility study without a performance claim, would make the claims precise.","section":"Section 3 (overall)"}],"minor_comments":[{"comment":"The rule for choosing the 'central slice' is not defined operationally; state whether it is the slice with maximum tumor area, the middle of the tumor-positive range, or a manual choice, and whether this uses ground-truth information.","section":"Section 2.3 and Fig. 2"},{"comment":"Provide the exact SAM2 checkpoint, whether the video memory/offline mode was used, and any prompt post-processing (e.g., thresholding of logits), so the experiments are reproducible.","section":"Section 2.2 and experimental setup"},{"comment":"State how the 279-patient subset was selected from the 922-patient Duke dataset and whether the exclusion affects the generalizability of the failure-mode analysis.","section":"Section 2.1"},{"comment":"R^2 values for linear fits are reported but no confidence intervals or correlation coefficients; a Spearman correlation with a confidence interval would be more informative than the linear-regression R^2 for these skewed distributions.","section":"Section 3, 'Factors Affecting Dice Score'"},{"comment":"The abstract says 'single bounding box annotation on one slice', but the mask-based experiment in Section 3 requires a full mask as input; please restate the abstract and conclusion to distinguish the bounding-box workflow from the mask-prompt analysis.","section":"Abstract and Section 3"},{"comment":"The supplementary material is referenced in Section 3 but is not included in the arXiv submission; please make it available or remove the reference.","section":"Section 3, qualitative results"}],"recommendation":"major_revision","confidential_remarks":"The mask-prompt scoring protocol is the key issue; the editor should ask for a rerun excluding the prompted slice and for baseline comparisons. The paper is otherwise within the workshop's scope and the zero-shot finding is of interest, but the current quantitative claims are not yet fully supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a clean, honest empirical study of using SAM2 to segment breast tumors in 3D MRI from a single bounding-box prompt. The central comparison—center-outward propagation beats bottom-to-top and top-to-bottom—is plausible and backed by distributions on 279 patients from the public MAMA-MIA dataset. That part is worth knowing if you work on applying video trackers to volumetric medical data.\n\nWhat's new: it's a direct application of SAM2, but the specific evaluation of propagation strategies and the failure-mode analysis (fragmented lesions) are not in the papers it cites. The paper builds entirely on open-source tools and a public benchmark, which fits its stated equity/accessibility goal.\n\nSoft spots. The biggest one is the mask-based tracking comparison. The paper reruns center-outward with 'a segmentation mask on a single central slice as input,' but never says where that mask comes from. The only plausible source is the MAMA-MIA ground truth. Equation (1) defines volumetric Dice over the full volume, and the paper doesn't mention excluding the prompted slice. If the input mask is ground truth and is included in the Dice, that slice contributes a Dice of exactly 1.0 by construction. Since the central slice is the largest and clearest tumor slice, the reported jump from 0.57 to 0.71 is substantially an artifact of scoring the prompt itself. This doesn't affect the bounding-box results or the propagation-strategy comparison, but it means the mask-based numbers should not be cited as-is. The stress-test note is correct on this.\n\nThere are also the usual application-paper gaps: no comparison to supervised baselines (e.g., nnU-Net, MedSAM2), no error bars or significance tests for the strategy differences, and the exact SAM2 checkpoint and slice-extraction pipeline are underspecified. The abstract's 'strong segmentation performance' oversells a mean Dice of 0.57.\n\nWho is this for? Groups considering zero-shot SAM2 for volumetric segmentation, especially in low-resource settings. The paper deserves a serious referee, but the mask-based section needs a rewrite: either exclude the prompt slice from the Dice computation or report the per-slice score on the prompt slice separately. I'd engage with it conditionally.","headline":"Useful empirical comparison of SAM2 propagation strategies for breast MRI, but the mask-based tracking result is likely inflated by including the ground-truth prompt slice in the Dice computation.","tokens_in":7899,"tokens_out":3185,"would_cite":false,"duration_ms":35859,"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":"A zero-shot video-segmentation model can turn a single annotated slice of a breast MRI into a full 3D tumor segmentation.","keywords":["SAM2","breast MRI","tumor segmentation","zero-shot segmentation","video object tracking","foundation models","volumetric Dice","equitable AI"],"falsifier":"Re-run the full pipeline on a subset of MAMA-MIA cases after independently verifying or re-registering the expert masks to the DCE-MRI slices; if mean volumetric Dice changes materially or the center-outward advantage over directional propagation disappears, the reported performance rests on an alignment artifact rather than on SAM2's tracking ability.","tokens_in":6986,"feed_emoji":"🩻","tokens_out":5469,"duration_ms":57758,"temperature":0.7,"pith_summary":"This paper asks whether a general-purpose image and video segmentation model, SAM2, can segment breast tumors across a full 3D MRI volume when given only one annotated slice. The authors report that it can: a single bounding box propagated slice-by-slice yields a mean volumetric Dice of 0.57, and using a segmentation mask as the prompt raises the mean to 0.71, with no training or fine-tuning. They compare three propagation directions and find that starting at the central tumor slice and moving outward outperforms bottom-to-top and top-to-bottom propagation. The practical stake is that an open-source, zero-shot model could lower the cost barrier for 3D breast MRI interpretation in settings where commercial medical AI is unaffordable. The paper also identifies fragmented multi-lesion tumors as the main failure mode.","feed_headline":"Zero-shot SAM2 segments 3D breast tumors from one box","feed_subtitle":"Center-outward propagation lifts mean Dice to 0.71 with mask prompts, no fine-tuning.","key_machinery":"The central mechanism is SAM2's tracking functionality, which takes the predicted mask from the previous frame as contextual guidance for the next frame. Applied slice-by-slice to axial MRI slices, this turns a single bounding-box or mask prompt into a chain of segmentations spanning the tumor volume. The paper's novel procedural choice is the center-outward traversal: initializing at the slice where the tumor is largest and clearest, then propagating both upward and downward, reduces long-range tracking errors compared with starting at an end slice.","core_discovery":"The central discovery is that SAM2, a vision foundation model trained for image and video segmentation rather than for volumetric medical data, can be reused as a 3D tumor segmenter by treating a breast MRI as a stack of 2D frames and propagating a single-slice prompt across adjacent slices. On the MAMA-MIA expert-segmented subset of the Duke Breast Cancer Dataset, the center-outward propagation strategy achieves mean volumetric Dice of 0.57 with bounding-box prompts and 0.71 with mask prompts, and it wins for the majority of the 279 patients against the two directional strategies. The authors additionally find that tumor size, volume, and slice count show no meaningful correlation with Dice score, while multiple small scattered lesions produce poor segmentations.","pith_inferences":["Because the method is domain-agnostic in principle, the same single-box tracking recipe could be tested on other volumetric modalities (CT, ultrasound sweeps) and on longitudinal MRI where the 'frames' are time points rather than slices.","A natural next step is to replace the human-supplied box with an automatic detector; if the detector finds the tumor-bearing slice and a box, the whole loop becomes annotation-free, though the paper does not evaluate this.","The center-outward advantage may reflect a more general principle: seeding propagation at the most informative frame improves error containment in any video-object-tracking setting, which is testable outside medicine.","The reported Dice values give a lower-bound style estimate for what a zero-shot tracker can do on unseen volumetric data; a supervised 3D U-Net baseline on the same 279 patients would make the cost-performance trade-off concrete."],"forward_implications":["If confirmed, a zero-shot SAM2 workflow can produce clinically plausible 3D tumor volumes from a single annotation, removing the need for large annotated training sets or task-specific fine-tuning.","Center-outward propagation should be the default traversal order for SAM2-based slice-wise medical tracking, since it maximizes per-patient Dice and consistency.","Using a segmentation mask rather than a bounding box as the initial prompt improves mean volumetric Dice from 0.57 to 0.71, so interactive annotators should prefer mask prompts whenever available.","The absence of correlation between Dice and tumor volume or slice count implies that failure is not simply a matter of tumor size; the identified multi-lesion failure mode suggests an explicit pre-selection or multi-prompt handling is needed for fragmented tumors.","Deployment cost in low-resource settings is lowered because the pipeline needs only one human annotation and open-source components."],"supporting_citations":[{"why":"Supplies the SAM2 model whose tracking functionality carries the slice-wise propagation.","marker":"[21]"},{"why":"Provides the expert voxel segmentations used as ground truth for the 279-patient evaluation.","marker":"[8]"},{"why":"Original Duke Breast Cancer Dataset with 3D bounding box annotations from which the slices and prompts are derived.","marker":"[22]"},{"why":"Supplies the slice-extraction procedure that determines which 2D MRI slices are fed to SAM2 and paired with masks.","marker":"[5,12,14,15]"},{"why":"Prior work adapting SAM2 to 3D medical data, used as context for the paper's zero-shot approach.","marker":"[17]"},{"why":"Prior demonstration of SAM2 on medical images as video, which the paper extends to breast MRI tumor segmentation.","marker":"[26]"}],"fun_headline_variants":["One box, zero training: SAM2 segments 3D breast tumors","Center-out propagation boosts SAM2 for 3D breast tumor segmentation","SAM2 zero-shot turns 3D breast MRI into stackable 2D slices","Affordable SAM2 maps 3D breast tumors from single click"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the MAMA-MIA expert voxel segmentations are correct and are aligned with the pre-contrast DCE-MRI slices extracted using the procedure of refs. [5,12,14,15], and the paper does not independently validate this alignment or provide the extraction pipeline.","fun_headline_variants_meta":{"raw":{"variants":["One box, zero training: SAM2 segments 3D breast tumors","Center-out propagation boosts SAM2 for 3D breast tumor segmentation","SAM2 zero-shot turns 3D breast MRI into stackable 2D slices","Affordable SAM2 maps 3D breast tumors from single click"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001121,"raw_usage":{"total_tokens":4651,"prompt_tokens":918,"completion_tokens":3733,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":3653}},"tokens_in":534,"tokens_out":3733,"duration_ms":29637,"temperature":1.0,"reasoning_tokens":3653,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:52:54.121721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full pipeline on a subset of MAMA-MIA cases after independently verifying or re-registering the expert masks to the DCE-MRI slices; if mean volumetric Dice changes materially or the center-outward advantage over directional propagation disappears, the reported performance rests on an alignment artifact rather than on SAM2's tracking ability.","supporting_citations":[{"cited_title":"Scientific data 12(1), 453 (2025) 10 Kang et al","cited_arxiv_id":null,"evidence_quote":"Provides the expert voxel segmentations used as ground truth for the 279-patient evaluation."},{"cited_title":"British journal of cancer 119(4), 508–516 (2018)","cited_arxiv_id":null,"evidence_quote":"Original Duke Breast Cancer Dataset with 3D bounding box annotations from which the slices and prompts are derived."}],"review_version":1}