{"id":"6fcbc225-549e-4e6b-898e-fa499b709c06","arxiv_id":"2412.16381","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"VerSe unifies automatic and interactive cardiac MRI segmentation by prompting one shared transformer backbone with both object queries and click queries.","lead":"VerSe is a single deep learning model that performs both automatic cardiac MRI segmentation and click-based correction by doctors. If its reported accuracy holds, it could cut manual annotation time in heart imaging workflows.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Interactive efficiency claim is protocol-dependent: all NoC/Dice(n) numbers come from a single simulated click policy that matches VerSe's training, with no evidence it transfers to other policies or real expert clicks.","rationale":"The reader's weakest assumption identifies the simulated click strategy as the point where the central claim is least secure, and I agree. The paper's contribution is explicitly framed around interaction efficiency, and every quantitative statement about that efficiency is generated by the Sec. 2.5 protocol. Because VerSe is trained with this exact protocol, the evaluation is not independent evidence for a general interaction-efficiency advantage; it is evidence about performance under one specific, machine-generated click distribution. The paper provides no human-click study, no alternative-simulator ablation, and no analysis of how baselines' training-time click distributions differ. This is a load-bearing concern because the clinical motivation of the paper is human-in-the-loop refinement, and the reported NoC values are the primary evidence for the headline claim. The Mode-2 vs. Mode-3 comparison in Table 2 raises an additional fairness issue, but the click-protocol concern is more fundamental: it affects even the Mode-3 numbers that the paper presents as the fair comparison. I therefore do not change the reader's CONDITIONAL verdict; the paper needs additional evidence that the efficiency gains are robust to the click protocol. The available code is a positive, since it makes the proposed re-evaluation feasible, but it does not by itself resolve the protocol-dependence concern.","tokens_in":12047,"tokens_out":9259,"duration_ms":83710,"concrete_test":"Re-run the Table 2 comparison on ACDC and M&Ms for VerSe Mode-3, SimpleClick, and SegNext under at least two alternative click policies: (i) sample the next click uniformly from the set of misclassified pixels (the RITM-style policy), and (ii) place the next click at the boundary point of the largest false-negative region nearest to the current mask edge. Report NoC85/NoC90/NoC95 and Dice(1)/Dice(20) for each policy. If VerSe's NoC advantage over SimpleClick/SegNext disappears or reverses under either alternative policy, the headline efficiency claim is protocol-dependent rather than a robust property of the model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim is conditional on the click-simulation protocol of Sec. 2.5. VerSe is trained with simulated clicks placed at the center of the largest connected component of the misclassified region, and the same protocol generates every Dice(n) and NoC number in Tables 2 and 3. The paper does not establish that this protocol is representative of how clinicians correct cardiac MRI masks, nor does it show that the ranking is stable under the click distributions used to train the baselines (e.g., RITM-style sampling from error regions). Because the model's training distribution is exactly the evaluation protocol, the reported NoC gains could reflect overfitting to a convenient simulator rather than a genuine interaction-efficiency advantage. For example, placing clicks at the center of a large error blob is arguably easier for a model that has only seen such clicks; human experts often click near boundaries or on scattered false-positive regions. If the simulator is changed, VerSe's NoC improvements may shrink or reverse. Mode-2 numbers in Table 2 are additionally conflated with automatic initialization (NoC85=0.431 on ACDC), but even the fair Mode-3 comparison uses the same protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VerSe, a unified cardiac MRI segmentation framework that jointly learns object queries and click queries as prompts for a shared transformer decoder, supporting fully automatic segmentation (Mode-1), automatic segmentation followed by interactive click refinement (Mode-2), and pure click-driven interactive segmentation (Mode-3). The architecture introduces semantic feature queries, foreground-background masked attention with separate positive and negative click branches, and residual multi-scale connections. The authors evaluate on seven cardiac MRI datasets and two out-of-distribution datasets, reporting Dice and Number-of-Clicks metrics, and claim state-of-the-art accuracy and interaction efficiency relative to RITM, iSegformer, SimpleClick, and SegNext.","tokens_in":12313,"tokens_out":4706,"duration_ms":41344,"significance":"If the reported results hold under fair evaluation, the multi-query integration is a useful architectural direction: it allows one model to serve both automatic and interactive workflows, and the evaluation span of nine datasets with code release is a practical strength. The consistently strong Mode-3 results on most datasets, especially at Dice(20) and NoC95 on LGE and bSSFP cardiac data, suggest the model is a competitive interactive segmenter. However, the central efficiency claim is currently supported only under a click-simulation protocol that matches VerSe's own training, and the Mode-2 comparison against click-only baselines is not apples-to-apples; these issues must be addressed before the SOTA claims can be accepted.","major_comments":[{"comment":"The Mode-2 results are not comparable with the interactive baselines because Mode-2 receives an automatic initial mask from object queries, while the baselines start from empty masks. For example, NoC85=0.431 on ACDC indicates that the automatic initialization already meets 85% Dice for most slices, so the reported click counts reflect automatic accuracy rather than interactive efficiency. Please either remove Mode-2 from the comparison against click-only methods, or give the baselines the same automatic initialization (e.g., an automatic model followed by click-based refinement) and report the total interaction cost including initialization.","section":"§3.1, Table 2"},{"comment":"The click simulation rule—placing each new click at the center of the largest connected component of the misclassified region—is used both to train VerSe and to generate every Dice(n) and NoC number in the evaluation. The paper does not establish that this rule matches real expert correction behavior, nor that the ranking of methods is stable under other click policies. If the same rule was applied to the baselines, they are at a disadvantage because their training used different click distributions (e.g., RITM-style sampling); if not, the comparison is inconsistent. Please evaluate with at least one additional click policy (e.g., RITM-style randomized sampling from error regions, boundary-biased clicks, or scattered misclassified points) and report the ranking under each policy, and temper the efficiency claims accordingly.","section":"§2.5, Tables 2 and 3"},{"comment":"The text states that on the M&Ms dataset VerSe achieves a Dice(1) score of 89.757%, surpassing SimpleClick (85.335%) and SegNext (85.197%), but Table 2 lists 87.460 for VerSe (Mode-3) on M&Ms; the 89.757 value is the ACDC Dice(1). This misreported number appears in a central comparison passage and should be corrected, and the surrounding discussion should be checked for similar transcription errors.","section":"§3.1, Table 2"},{"comment":"The claim that VerSe (Mode-3) 'consistently achieves the best Dice scores and lower interaction costs among six out of seven datasets' is not fully supported by the per-metric results. On MyoPS++ (T2), SimpleClick has higher Dice(1) (79.231 vs. 74.096) and lower NoC90 (6.253 vs. 7.772), even though VerSe has higher Dice(20) and lower NoC95. Please report which specific metrics are best per dataset or use a composite metric that accounts for trade-offs among Dice(1), Dice(20), NoC85, NoC90, and NoC95.","section":"§3.1, Table 2"}],"minor_comments":[{"comment":"There are typos: 'mutiple queries' should be 'multiple queries', and the keyword 'Mutiple Prompts' should be 'Multiple Prompts'.","section":"Abstract and Keywords"},{"comment":"The text says that for M&Ms-2 only long-axis (LA) cine images were used, but Table 1 lists M&Ms-2 as bSSFP without noting the LA/SA distinction; please clarify in the table or caption.","section":"Table 1"},{"comment":"References [3] and [4] appear to be the same paper (Bernard et al., IEEE TMI 2018); please merge or remove the duplicate.","section":"References"},{"comment":"No standard deviations or significance tests are reported for the main Dice and NoC comparisons; adding results over multiple runs (or at least a reproducibility statement) would strengthen the SOTA claims.","section":"§3, Evaluation Metrics"},{"comment":"The label 'Sematic feature query encoder' contains a typo; it should read 'Semantic'.","section":"Figure 3(a)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable systems contribution with a broad evaluation, but the main efficiency claims need to be re-framed or re-verified. The Mode-2 comparison issue is straightforward to fix by giving baselines the same initialization, while the click-policy dependence is more consequential; if additional policies reverse the ranking, the authors should soften the SOTA wording. I do not see signs of circularity or fabricated evaluation; the numbers are measured on held-out test sets."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The VerSe paper is a clean, well-engineered attempt to unify automatic and interactive segmentation in one cardiac MRI model. The real novelty is joint learning of learnable object queries (Mask2Former-style) and click queries (SAM-style) in a shared transformer decoder, plus a semantic feature query and foreground/background masked attention that separates positive and negative click branches. Mode-3 interactive results are consistently strong: on ACDC, M&Ms, and LASCarQS++ it beats SimpleClick and SegNext by a clear margin in Dice(20) and NoC. Ablations support the design choices, and code is provided.\n\nThe soft spots are real but not fatal. The Mode-2 comparison in Table 2 is apples-to-oranges: starting from an automatic mask gives a free head start, so NoC85=0.431 on ACDC is not comparable to click-only baselines. That comparison should be reframed or re-run. More importantly, all interactive numbers come from a simulated click policy (center of largest misclassified component) that matches how VerSe was trained, with no evidence it transfers to other policies or real experts. If the policy changes, efficiency gains could shrink. The OOD results are mixed: on OAIZIB, VerSe actually has lower Dice than SimpleClick, though lower NoC80; that undercuts the blanket claim of significant improvement. And there are no error bars or significance tests anywhere, which is thin for a paper claiming a new benchmark.\n\nThat said, the central idea is sound, the Mode-3 comparisons are mostly fair (all models evaluated under the same protocol, with the caveat above), and the gaps are large enough that the main result is unlikely to be pure protocol artifact. The paper deserves serious peer review, where a good referee can push on simulation robustness and the Mode-2 framing. I'd cite it if I were working in interactive medical segmentation, and I'd bring it to reading group as a useful example of multi-query design.","headline":"A genuinely useful unification of automatic and interactive cardiac MRI segmentation with strong Mode-3 results, though the efficiency claim leans on a favorable click-simulation protocol.","tokens_in":12836,"tokens_out":2172,"would_cite":true,"duration_ms":18327,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"VerSe claims that a single segmentation model can serve both fully automatic and interactive cardiac MRI segmentation by prompting one shared backbone with object queries and click queries, and that this unified design sets new best…","keywords":["cardiac MRI segmentation","interactive segmentation","multi-query prompting","object queries","click queries","transformer decoder","medical image segmentation"],"falsifier":"A reader study in which radiologists correct VerSe masks on the same seven datasets, with their real clicks logged; if NoC90 under real clicks moves toward SimpleClick's level or beyond, the simulated-click evaluation overstates VerSe's interaction efficiency.","tokens_in":11854,"feed_emoji":"🫀","tokens_out":7666,"duration_ms":61288,"temperature":0.7,"pith_summary":"Cardiac MRI segmentation typically splits into two unsatisfying modes: automatic models that need manual correction on basal and apical slices, and interactive click-based models that start from scratch and burn many clicks on 3D volumes. VerSe argues that one model can do both, and do each better, if the same backbone is prompted by two kinds of queries at once: a learned object query that names the target structure for automatic segmentation, and click queries that mark where a clinician would correct the current mask. The paper reports that this design outperforms specialized automatic and interactive baselines on Dice accuracy and on clicks-to-target across seven cardiac MRI datasets, and that the interactive mode generalizes to brain and knee MRI without retraining. The point matters because it suggests clinical tools could stop alternating between separate automatic and interactive systems and instead interactively refine one trustworthy initial mask.","feed_headline":"VerSe needs half the clicks for 90% Dice on cardiac MRI","feed_subtitle":"The same transformer-decoder weights handle object-query autopilot and click-based refinement.","key_machinery":"The load-bearing component is multi-query integration: a small set of learnable object-query vectors per target and a click-query stream that combines a sparse positional encoding of each click with a semantic feature query obtained by average-pooling a local feature patch around the click at three scales and projecting it through an MLP. These prompts are processed by a shared transformer decoder whose foreground-background masked attention forces positive clicks to attend to the current-mask foreground and negative clicks to the background, after which the updated prompts are concatenated and used as keys and values to update the image features. Multi-scale residual connections resample earlier decoder features and add them to the next scale, letting the prompts interact with features at 1/8, 1/4, and 1/2 resolutions. The whole design is what lets one forward pass select an automatic object, refine it from clicks, or start interactively from an empty mask.","core_discovery":"On the paper's own terms, the discovery is a mode-unifying prompting mechanism: VerSe trains object queries (small groups of learnable vectors per anatomical target) and click queries (sparse positional encodings plus 'semantic feature queries' read from local image feature patches around each click) to jointly drive a shared UTNet encoder and a hybrid-attention transformer decoder. The model's three operating modes—automatic, automatic-then-refine, and click-only—emerge from which prompts are active, with no separate heads or weights per mode. Across seven cardiac MRI datasets, the paper reports the highest Dice scores and lowest click counts among compared methods, with Mode-2 reaching NoC90 in the 0.9-to-2.2 click range on the four bSSFP cine datasets, and Mode-3 exceeding 94% Dice at 20 clicks even on the hardest scar and edema datasets. On out-of-distribution brain and knee MRI, the click-only mode keeps an edge over baselines, supporting the paper's claim that the query-integration design generalizes beyond cardiac images.","pith_inferences":["Extension: the semantic feature query likely works because it gives the decoder local appearance context around a click, not just a coordinate; a testable corollary is that larger pooling windows or a different multi-scale fusion would help most on low-contrast scar targets.","Extension: the decoder's interchangeability of query types suggests that box, text, or scribble prompts could be added as one more query type without rearchitecting the model.","Extension: the reported click-efficiency numbers depend on the simulated user placing clicks at the center of the largest error region; real clinicians may click on boundaries or scattered regions, so the strongest validation would be a reader study measuring NoC90 with actual radiologist clicks."],"forward_implications":["A clinician can run automatic segmentation first and correct only the slices that fail, with Mode-2 reaching NoC85 of 0.43 on ACDC and 0.88 on M&Ms-2.","On the hardest scar and edema targets, pure interactive Mode-3 is the better fallback: despite starting from an empty mask, it ends at 94.2–95.0% Dice after 20 clicks, far above all compared baselines.","The same weights transfer across MRI contrasts and body regions: training on cardiac MRI and testing on brain tumor MRI gives 94.5% Dice in 10 clicks with VerSe, roughly 2 points above SimpleClick.","Because all three modes share one encoder and decoder, deployment is one model instead of two or more specialized networks."],"supporting_citations":[{"why":"UTNet is adopted as the image encoder backbone, supplying the cardiac-MRI-specific multi-scale features that the shared decoder refines.","marker":"[16]"},{"why":"SAM is the source of the sparse positional click-query encoding used to turn click points into prompts.","marker":"[19]"},{"why":"RITM provides the iterative click-simulation training scheme and mask-guidance baseline that VerSe extends by placing new clicks at the center of the largest misclassified region.","marker":"[35]"},{"why":"SimpleClick is the state-of-the-art interactive segmentation baseline whose Dice and NoC numbers are the main comparison in Tables 2 and 3.","marker":"[28]"},{"why":"Mask2Former's masked-attention transformer is the basis of the foreground-background masked attention that routes positive and negative click queries to target and context regions.","marker":"[9]"},{"why":"The separate foreground and background attention branches extend this video object segmentation work's explicit object/background separation.","marker":"[10]"}],"fun_headline_variants":["VerSe merges object and click prompts to cut clicks and lift Dice","VerSe unifies auto and interactive cardiac MRI segmentation with shared queries","VerSe trains object and click queries together for one versatile segmenter","VerSe: one prompt set for auto and click-guided cardiac MRI segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All measured click efficiency rests on the assumption that placing each new click at the center of the largest connected misclassified region matches how a human expert would correct the model.","fun_headline_variants_meta":{"raw":{"variants":["VerSe merges object and click prompts to cut clicks and lift Dice","VerSe unifies auto and interactive cardiac MRI segmentation with shared queries","VerSe trains object and click queries together for one versatile segmenter","VerSe: one prompt set for auto and click-guided cardiac MRI segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00097,"raw_usage":{"total_tokens":4138,"prompt_tokens":970,"completion_tokens":3168,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":3091}},"tokens_in":586,"tokens_out":3168,"duration_ms":20664,"temperature":1.0,"reasoning_tokens":3091,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:37:24.309638+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader study in which radiologists correct VerSe masks on the same seven datasets, with their real clicks logged; if NoC90 under real clicks moves toward SimpleClick's level or beyond, the simulated-click evaluation overstates VerSe's interaction efficiency.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"UTNet is adopted as the image encoder backbone, supplying the cardiac-MRI-specific multi-scale features that the shared decoder refines."},{"cited_title":"In: 2022 IEEE International Conference on Im- age Processing (ICIP)","cited_arxiv_id":null,"evidence_quote":"RITM provides the iterative click-simulation training scheme and mask-guidance baseline that VerSe extends by placing new clicks at the center of the largest misclassified region."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"SimpleClick is the state-of-the-art interactive segmentation baseline whose Dice and NoC numbers are the main comparison in Tables 2 and 3."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"The separate foreground and background attention branches extend this video object segmentation work's explicit object/background separation."}],"review_version":1}