{"id":"ee286b35-4add-4819-b2b3-cc5d735794f1","arxiv_id":"2412.13299","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A cascade that reuses each predicted slice as a new support example improves few-shot sequential medical image segmentation over plain UniverSeg on 5 of 8 cardiac structures.","lead":"This paper proposes a way to segment a stack of medical images using only a few labeled slices, by feeding each predicted slice back as context for the next. The method, tested on heart MRI scans, improves consistency in several anatomical regions over the base few-shot model, though not in all regions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing risk is that HVSMR may be in UniverSeg's pretraining set; Section 6 asserts otherwise without evidence, so the reported DSC gains may reflect memorization rather than few-shot generalization.","rationale":"The reader's weakest assumption correctly identifies the single most load-bearing issue: the evaluation depends on HVSMR not being in UniverSeg's pretraining data, and the paper provides no evidence for this exclusion. This is not a peripheral concern; it determines whether the reported DSC improvements measure generalization to unseen anatomy or familiarity with training data. The paper itself flags the assumption in Section 6, which is honest but does not resolve the problem. If HVSMR is in UniverSeg's pretraining, the baseline and ICS comparisons are confounded, and the five-region improvement cannot be interpreted as evidence for the method's general utility. I agree with the reader's conditional verdict: the claim is plausible but not yet established. The proposed check is concrete and feasible: the UniverSeg codebase and MegaMedical dataset list are public, and a simple registry lookup settles whether HVSMR was used. If the check finds HVSMR, the verdict should move to REJECT; if it finds HVSMR absent, the central empirical claim becomes much more credible. I did not identify a more fundamental internal inconsistency in the method or the writing. The off-by-one behavior in Algorithm 1 (querying initial slices) and the absence of a comparison to 4S are notable, but they are secondary to the data-contamination risk because they would not change the qualitative conclusion if the evaluation set were clean.","tokens_in":9949,"tokens_out":5522,"duration_ms":58458,"concrete_test":"Check the official UniverSeg/MegaMedical dataset registry (in the released code or paper supplement) for 'HVSMR' and for the 2016 HVSMR challenge dataset. If HVSMR appears in the list of 53 training datasets, re-run Table 1 with a held-out dataset verifiably absent from UniverSeg pretraining; if the significant gains in LA/RA/AO/PA/SVC do not replicate, the central claim is refuted. If HVSMR is absent, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that ICS significantly outperforms independent UniverSeg in LA, RA, AO, PA, and SVC—rests entirely on HVSMR serving as an unseen testbed. Section 6 explicitly frames this as an assumption: 'Because UniverSeg utilizes a large portion of publicly available datasets, it was necessary to rely on the limited open datasets that were not used for pre-training.' But no citation, dataset-list inspection, or control experiment is provided. UniverSeg was pretrained on 53 public datasets, and HVSMR is a public cardiac MRI dataset with a long-standing challenge version; overlap is a concrete possibility. If HVSMR or a predecessor appears in UniverSeg's training mix, the model has seen these patients and anatomies, so both baseline and ICS numbers reflect in-distribution performance or memorization, not few-shot generalization. Because the paper's stated reason for choosing HVSMR is exactly its exclusion from pretraining, this unverified assumption is load-bearing: a positive result on a training-set-similar test does not establish the method's advantage. If the assumption fails, the primary quantitative comparison is invalidated, and the paper's broader claim of reducing annotation burden in unseen sequential medical volumes is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes In-context Cascade Segmentation (ICS), a training-free extension of the UniverSeg few-shot segmentation framework for sequential medical volumes. ICS initializes a support set from a few labeled slices, then performs forward and backward inference across the volume, iteratively appending predicted masks to the support set (with a sliding window of size m) to propagate anatomical context. The method is evaluated on the HVSMR dataset for eight cardiac structures, reporting Dice similarity coefficients against the baseline UniverSeg applied independently per slice. The paper also studies the effect of the number of initial support slices and their positions. The main claimed finding is that ICS significantly improves DSC in five of eight regions (LA, RA, AO, PA, SVC) at p<0.05, with qualitative illustrations of improved inter-slice consistency.","tokens_in":10279,"tokens_out":5216,"duration_ms":44544,"significance":"If the empirical claims survive scrutiny, ICS offers a simple, practical way to reduce annotation burden for sequential medical images by exploiting an existing pretrained few-shot model without any training or fine-tuning. The idea of updating the support set with predicted masks resembles self-training/pseudo-label propagation, but applying it within UniverSeg's in-context learning framework and evaluating on cardiac MRI is a useful contribution. The paper is transparent about its limitations, which is commendable. However, the significance is contingent on two load-bearing validity conditions: (i) that HVSMR was not part of UniverSeg's pretraining, and (ii) that the statistical analysis supports the claimed significant improvements. Both are currently unverified. The ablation of support slice number and position is informative but descriptive, lacking statistical tests.","major_comments":[{"comment":"The assumption that HVSMR was not used in UniverSeg's pretraining is unverified. The paper states in Section 6: \"Because UniverSeg utilizes a large portion of publicly available datasets, it was necessary to rely on the limited open datasets that were not used for pre-training,\" and then restricts evaluation to HVSMR without citing the MegaMedical dataset composition or otherwise demonstrating non-overlap. Since HVSMR is a public cardiac MRI dataset that predates UniverSeg, and UniverSeg was trained on 53 public datasets (Section 3.2), this assumption is not safe. Please verify by listing the MegaMedical datasets (or documenting HVSMR's absence), or by running a control experiment on a dataset known to be excluded. If HVSMR is in the training set, the reported DSC gains reflect in-distribution performance rather than few-shot generalization, and the central claim of reducing annotation burden on unseen volumes is unsupported.","section":"Section 6 / Section 4.1"},{"comment":"The statistical comparison is under-specified and the significance claim is overstated. No test name is given (paired t-test, Wilcoxon signed-rank, or other), and no multiple-comparison correction is reported for the eight regions. With a Bonferroni threshold of 0.00625, only LA (p=0.0007), PA (p<0.0001), and SVC (p<0.0001) would remain significant; RA (p=0.0214) and AO (p=0.0206) would not. The abstract's claim that ICS \"significantly outperforms baseline methods\" should be revised to reflect a corrected analysis, and the manuscript should name the test, justify its assumptions, and report adjusted p-values or confidence intervals.","section":"Section 5 / Table 1"},{"comment":"Excluding slices without the target structure before evaluation biases the DSC comparison. The paper states: \"As a preprocessing step, we removed any slices that did not contain the region of interest (i.e., no annotated labels).\" In sequential medical volumes, empty slices are anatomically meaningful; a method that correctly predicts nothing in those slices should be rewarded, and a method that spuriously predicts there should be penalized. Removing these slices inflates DSC and may differentially affect baseline and ICS (e.g., ICS's tendency to over-segment, noted in Section 6). Please report results on the complete volumes and, if the filtered analysis is retained, discuss the direction of bias.","section":"Section 4.1 / Evaluation"}],"minor_comments":[{"comment":"The abstract says ICS \"significantly outperforms baseline methods\" (plural), but only one baseline (UniverSeg) is compared; this should be reworded to match the experimental scope.","section":"Abstract"},{"comment":"The symbol m is used both for the initial number of labeled slices and for the maximum support set size; clarify the relationship, and in the \"varying m\" experiment, specify whether the initial support set size and the sliding window size change together.","section":"Algorithm 1 / Section 4.4"},{"comment":"The initial support slice position experiment lacks details on how many positions were tested, which slice indices were used, and whether any statistical analysis was performed; the claim that \"positions strongly influence segmentation accuracy\" is currently supported only by visual inspection.","section":"Section 4.2 / Figure 8"},{"comment":"The box plots would benefit from paired significance markers and a note on the test used, consistent with Table 1.","section":"Figure 3"},{"comment":"The caption contains a typo: \"in-contest cascade segmentation\" should be \"in-context cascade segmentation.\"","section":"Figure 2"},{"comment":"Reference [26] contains a typo: \"Iternational\" should be \"International.\"","section":"References"},{"comment":"The first limitation is worded confusingly: \"the assumption that all volumes, including the boundary slices, are labeled\" is unclear because the method itself operates on unlabeled slices; the authors likely mean that the evaluation assumes ground-truth masks exist for every slice, which is the same issue as the empty-slice removal discussed in the major comments.","section":"Section 6 / Limitations"},{"comment":"The description of data augmentation (rotations by 90 degrees, 180 degrees, 270 degrees) applied to the support set is ambiguous: does this multiply the support set size by four at every inference step, and does the baseline receive the same augmentation? Clarify the procedure.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the HVSMR/UniverSeg overlap. I would encourage the editor to ask the authors to confirm this explicitly; if HVSMR is in fact part of MegaMedical, the paper's core result collapses and would need to be re-run on a verified out-of-training dataset. Additionally, the paper's relationship to the authors' prior 4S method [40] could be discussed more clearly to position the novelty of ICS relative to existing pseudo-label propagation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest but real extension of UniverSeg — feed predicted slices back into the support set and propagate forward/backward. It's clean and honest, but the only test dataset (HVSMR) may not actually be out-of-distribution for UniverSeg, and the paper itself flags that as an assumption without evidence. That's the load-bearing risk.\n\nWhat's new: the specific cascade formulation and the ablations on support count/position are not in the cited literature. The closest thing, 4S, requires retraining; ICS doesn't. That's a genuine difference. The paper also does some things well: it reports non-significant results, shows cases where baseline beats ICS, and discusses over-segmentation.\n\nSoft spots, in order of severity. First, the UniverSeg pretraining overlap. UniverSeg was trained on 53 public datasets; HVSMR is public and cardiac. The paper says in Section 6 it relied on datasets 'not used for pre-training' but gives no citation, no list, no control. If HVSMR is in the training mix, both baseline and ICS scores are in-distribution and the whole comparison collapses. This is fixable but essential: check the MegaMedical dataset list or test on an in-house/held-out volume. Second, evaluation choices inflate DSC: empty slices are removed before scoring, and no statistical test is named, with no multiple-comparison correction across 8 regions. The p-values are plausible but underreported. Third, no comparison to 4S, the most related method, even though it's cited and one author is shared. Even a brief discussion of why retraining-free matters would help. Fourth, no code or data, which limits reproducibility for a method whose whole appeal is being training-free.\n\nThe stress-test note is right to worry about the pretraining overlap. I'd go further: even if HVSMR is outside UniverSeg, the small gains (DSC 0.38-0.57) and single dataset mean the headline claim of 'reducing annotation burden' is only weakly supported. But the paper doesn't oversell too badly; the conclusion is appropriately cautious.\n\nWho is this for? People working on few-shot medical segmentation and annotation-efficient workflows. It's a workshop-level contribution with a correctable flaw. A serious referee should see it, mainly to force the authors to resolve the pretraining question and tighten the stats.\n\nRecommendation: send to peer review, but with major revision — the pretraining overlap must be addressed, and the abstract must stop saying 'baseline methods' when there's one baseline.","headline":"A modest but genuinely training-free cascade over UniverSeg; the paper is honest but must resolve whether HVSMR is actually outside UniverSeg's pretraining set.","tokens_in":10717,"tokens_out":2056,"would_cite":false,"duration_ms":19369,"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":"By feeding each slice's predicted mask back into the support set, ICS segmentation improves inter-slice consistency and Dice score on five of eight HVSMR cardiac regions.","keywords":["in-context learning","medical image segmentation","few-shot segmentation","inter-slice consistency","support set","UniverSeg","HVSMR","semi-supervised learning"],"falsifier":"Check whether HVSMR-2.0 appears in UniverSeg's MegaMedical pretraining corpus; if it does, rerun the comparison on a held-out in-house volume or a dataset released after UniverSeg's training cutoff and see whether the significant DSC gains in LA, RA, AO, PA, and SVC persist.","tokens_in":9697,"feed_emoji":"🩺","tokens_out":3932,"duration_ms":35009,"temperature":0.7,"pith_summary":"In-context Cascade Segmentation (ICS) adds no new training. It takes the pretrained few-shot segmenter UniverSeg and, after predicting a mask for one slice of a volume, appends that mask to the support set used for the next slice, running forward and backward from a few labeled slices. The paper claims this simple feedback loop gives significantly higher Dice scores than independent per-slice inference in the left atrium, right atrium, ascending aorta, pulmonary artery, and superior vena cava (p<0.05), while matching baseline performance in the other three regions. The motivation is practical: a radiologist labels only a handful of slices, and the model then propagates the labels through the volume with stable boundaries.","feed_headline":"Prediction feedback lifts cardiac segmentation in 5 of 8 regions","feed_subtitle":"In-context cascade beats per-slice baseline on HVSMR, cutting annotation load.","key_machinery":"The central mechanism is a sliding support set under a fixed pretrained model. Starting from m labeled slices, UniverSeg predicts slice i+1, that prediction is appended to the support set, and the oldest entry is dropped once the set exceeds m; the same procedure runs in reverse from the initial slice. Because UniverSeg's CrossBlock modules exchange features between query and support images, each new support mask injects nearby anatomical context into the next query, which is what enforces slice-to-slice consistency without any gradient updates.","core_discovery":"On the HVSMR dataset (eight cardiac targets across 60 MRI scans), the paper's central claim is that iteratively enlarging the support set with each slice's own prediction produces segmentations that are more consistent across neighboring slices and more accurate in five of eight anatomical regions than the same model without in-context updates. The gain is attributed to information propagating forward and backward through the sequence, so that anatomical structure visible in neighboring slices acts as context for each query. Where ICS loses, notably the left ventricle, it tends to over-segment, adding false positives; the paper treats this as a known failure mode of self-training-style pseudo-label feedback to be mitigated with confidence metrics.","pith_inferences":["The paper does not test other modalities; if the consistency gain is due to pseudo-label feedback rather than dataset memorization, the same cascade should transfer to CT, ultrasound, and other sequential modalities.","The position sensitivity suggests a cheap pre-annotation step, analogous to cold-start active learning: cluster slice appearances and pick a central or representative slice as the initial support before any labels are drawn.","A confidence-gated update, appending only masks above a per-slice confidence threshold, could reduce the left-ventricle over-segmentation the paper observed; this is a concrete extension the paper only gestures at with future-work language.","The five regions with significant gains (LA, RA, AO, PA, SVC) are generally smaller or geometrically complex, whereas the non-significant LV, RV, and IVC are larger cavities; whether benefit scales with boundary complexity is worth testing explicitly."],"forward_implications":["Fewer labeled slices suffice: with m=5 initial slices, ICS propagates labels through entire volumes, reducing annotation load for sequential CT and MRI data.","Boundary consistency improves in complex structures such as the pulmonary artery, where per-slice baseline inference produces discontinuous predictions and missing areas.","The number and spatial placement of initial support slices materially change accuracy; more supports generally help, but the effect plateaus and costs inference time and memory.","ICS reduces slice-to-slice variance in Dice score relative to independent per-slice inference, giving more stable volume-level segmentations."],"supporting_citations":[{"why":"Supplies the frozen few-shot segmenter UniverSeg whose support set ICS modifies; every ICS output goes through this pretrained model without additional training.","marker":"[19]"},{"why":"Provides the HVSMR-2.0 dataset with 60 cardiac MRI scans and ground-truth masks for eight anatomical regions used in all experiments.","marker":"[20]"},{"why":"The closest prior sequential pseudo-label approach (4S), used as the self-training reference point for the over-segmentation behavior that ICS shows in some regions.","marker":"[40]"}],"fun_headline_variants":["Cardiac segmentation improves in 5 of 8 regions with self-feedback","In-context cascade boosts cardiac MRI segmentation across slices","Feeding own predictions back lifts cardiac segmentation in most cases","Self-referencing context sharpens cardiac segmentation, cuts workload","Iterative support growth yields cardiac gains in 5 of 8 regions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes HVSMR was not among the 53 public datasets used to pretrain UniverSeg; if that assumption is wrong, the reported gains could come from memorization rather than from the cascade mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Cardiac segmentation improves in 5 of 8 regions with self-feedback","In-context cascade boosts cardiac MRI segmentation across slices","Feeding own predictions back lifts cardiac segmentation in most cases","Self-referencing context sharpens cardiac segmentation, cuts workload","Iterative support growth yields cardiac gains in 5 of 8 regions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000347,"raw_usage":{"total_tokens":1863,"prompt_tokens":869,"completion_tokens":994,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":906}},"tokens_in":485,"tokens_out":994,"duration_ms":9494,"temperature":1.0,"reasoning_tokens":906,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:15:06.483966+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check whether HVSMR-2.0 appears in UniverSeg's MegaMedical pretraining corpus; if it does, rerun the comparison on a held-out in-house volume or a dataset released after UniverSeg's training cutoff and see whether the significant DSC gains in LA, RA, AO, PA, and SVC persist.","supporting_citations":[{"cited_title":"Sabuncu, John Guttag, and Adrian V","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen few-shot segmenter UniverSeg whose support set ICS modifies; every ICS output goes through this pretrained model without additional training."},{"cited_title":"Pace, Hannah T","cited_arxiv_id":null,"evidence_quote":"Provides the HVSMR-2.0 dataset with 60 cardiac MRI scans and ground-truth masks for eight anatomical regions used in all experiments."},{"cited_title":"Sequential semi-supervised segmentation for serial electron microscopy image with small number of labels","cited_arxiv_id":null,"evidence_quote":"The closest prior sequential pseudo-label approach (4S), used as the self-training reference point for the over-segmentation behavior that ICS shows in some regions."}],"review_version":1}