{"id":"48209dc6-5ffd-4c14-8eda-c0ce1f9346f2","arxiv_id":"2607.25432","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A bi-level framework that lets a learnable superpixel model and a few-shot scribble-supervised segmentation model teach each other achieves state-of-the-art mean Dice on ACDC and Prostate.","lead":"This paper presents BiSCL, a training framework that couples a learnable superpixel network with a segmentation network for medical images labeled with only a few scribbles. It reports large gains over prior scribble-supervised methods on two MRI benchmarks, but the evaluation rests on a single favorable split and hand-tuned spatial priors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed gains may be inflated by test-set hyperparameter selection and a single unrandomized few-shot split; a validation-based multi-split evaluation is needed before 'significantly outperforms' is supportable.","rationale":"The reader identifies the center-of-image prior as the weakest assumption. I do not think that is the most load-bearing issue for the paper's central claim: ACDC and Prostate organs are roughly centrally located, so the prior can hold on the two datasets for which the claim is made. The evaluation protocol is more fundamental because it determines whether the reported improvements are real even on those datasets. The paper explicitly tunes the key filter hyperparameters on the test set (Table 4 'best performance'), uses one non-random few-shot split, and provides no significance testing. These issues are addressable but currently prevent unconditional acceptance. No code is provided, so the reported numbers cannot be independently verified. I agree with the reader's final CONDITIONAL verdict, but for a different primary reason, hence partial agreement.","tokens_in":14552,"tokens_out":5029,"duration_ms":57380,"concrete_test":"Re-run the ACDC and Prostate comparisons under a pre-registered protocol: (1) choose sigma and gamma_center using only the validation split, or fix them before any test evaluation; (2) repeat the few-shot construction across 10 random draws of 5 labeled training cases; (3) report mean +/- std of Mean Dice and Mean HD95 across splits, and paired bootstrap 95% CIs for BiSCL vs QMaxViT-Unet+ (ACDC) and vs ScribbleVC (Prostate). If the lower confidence bound of the gain does not exceed zero, or the gain does not persist on a majority of splits, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim ('significantly outperforms existing state-of-the-art scribble-supervised methods') rests entirely on Tables 1-2, but the evaluation protocol makes the comparison optimistic in a way that directly affects the headline numbers. In Table 4, sigma and gamma_center are selected because they 'achieve the best performance' on the ACDC test set, and the final reported results use these test-set-tuned values. The few-shot construction is a single non-random split (first five training cases), with no seed variation or repeated splits, and the reported standard deviations are across test slices, not across training runs or split choices. With only five labeled cases, test-set hyperparameter tuning can exploit idiosyncratic variance in the test set, so the 5.2 pp Mean Dice gain over QMaxViT-Unet+ and 5.6 pp over ScribbleVC may not replicate under a fair protocol. The center prior (Eqs. 8-9, 13-14) is also an unvalidated anatomical assumption, and Table 3 shows the method is highly sensitive to the filter, but the more immediately load-bearing problem is that the claimed outperformance has not been demonstrated under a held-out protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BiSCL, a bi-level collaborative learning framework for few-shot scribble-supervised medical image segmentation. An upper-level learnable superpixel network (based on SSN) provides region-level structural priors; the lower-level segmentation network (U-Net in a Mean Teacher setting) uses superpixel-guided pseudo-label diffusion with a spatial-prior-guided adaptive filter to suppress unreliable expansion. Segmentation-derived semantic feedback is then used to refine the superpixel representation. The method is validated on ACDC and Prostate with only five scribble-annotated training cases, reporting Mean Dice improvements of 5.2 pp over QMaxViT-Unet+ on ACDC and 5.6 pp over ScribbleVC on Prostate, with corresponding HD95 reductions.","tokens_in":14853,"tokens_out":2566,"duration_ms":29842,"significance":"If the reported gains are reproducible, the framework is a useful contribution to weakly supervised medical image segmentation: it addresses a realistic setting where both annotation sparsity and annotated-sample scarcity coexist, and the superpixel/segmentation feedback loop is an interesting mechanism beyond static superpixel priors. The paper includes a clear formalization of the bi-level objective, a concrete training algorithm, and an informative ablation study showing that unconstrained superpixel diffusion hurts and the proposed filter restores performance. The central idea is therefore worth pursuing. However, the current experimental evidence is not sufficient to support the strong claim of 'significantly outperforms,' because hyperparameters are selected on the test set and the few-shot setting is evaluated on a single non-random split.","major_comments":[{"comment":"The final reported results (Tables 1 and 2) use hyperparameters sigma=0.3 and gamma_center=6, which are selected in Table 4 because they 'achieve the best performance' on the ACDC test set. Using the test set for model selection makes the headline comparison optimistic: the 5.2 pp gain over the second-best baseline on ACDC may be inflated by test-set-specific tuning. A proper validation-based selection (e.g., choosing sigma and gamma_center on the existing validation split and only then reporting test performance) is required before the 'significantly outperforms' claim is supported.","section":"§4.5, Table 4"},{"comment":"The few-shot evaluation uses only the first five training cases as labeled data, with no repeated random splits or multiple training seeds. The reported standard deviations are across test slices, not across training runs or split choices. With only five labeled cases, a single split can produce large variance in the comparison, so the 5.2–5.6 pp differences may not be stable. The authors should report results over multiple random few-shot splits and/or multiple seeds, and ideally include a paired significance test, to substantiate the claimed improvement over the second-best method.","section":"§4.1, Tables 1–2"},{"comment":"Both the spatial-prior-guided filter and the non-uniform superpixel initialization rely on an unvalidated assumption that target anatomical structures lie near the image center. Equation (9) limits allowable expansion proportionally to a Gaussian of the centroid's distance from the center, and Eq. (14) redistributes superpixel seeds toward the center. The ablation (Table 3) shows the filter is indispensable: without it Mean Dice drops from 0.599 to 0.311. If the target organ is off-center in many slices, the method could suppress valid pseudo-label expansion and lose the only dense supervision signal. The paper should provide evidence about the actual centroid distribution of the target structures in ACDC and Prostate, and report sensitivity to this prior (e.g., by comparing with a version that uses a learned or data-dependent spatial prior, or by evaluating on slices where the target is","section":"§3.4, Eqs. (8)–(9) and (13)–(14)"}],"minor_comments":[{"comment":"The formula for the expansion ratio r_k appears without a visible fraction bar in the text: r_k = |Ω_k| Σ ... . It should be written as a fraction with the denominator being the sum of high-confidence seeds. Please clarify.","section":"Eq. (7)"},{"comment":"The dominant seed class is computed as argmax over c in {1,...,C-1}, which excludes the background class (assumed to be 0). If scribble annotations never label background, this is fine, but it should be stated explicitly; otherwise the exclusion of background in the diffusion rule is not justified.","section":"Eq. (5)"},{"comment":"The reuse of theta_0 and w_0 as both initial parameters and loop-indexed variables is confusing. For example, line 8 sets theta_0 <- theta_hat while the outer loop still refers to theta_0; renaming would improve readability.","section":"Algorithm 1"},{"comment":"The paper states in a footnote that it has been accepted for publication at ACM MM 2026. This is not a technical issue, but it is unusual in a submitted manuscript and may raise editorial questions about prior disclosure; please verify venue policies.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and the ablation is instructive, but the empirical claim is load-bearing and currently rests on a single non-random split and test-set hyperparameter selection. I would not reject outright, but the authors need to re-run validation with a held-out protocol and multiple splits before the claim 'significantly outperforms' can be taken at face value. The center-prior assumption is also a substantive limitation that should be examined, not just mentioned as a design choice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me skip the formalities. This is a workmanlike method paper for the niche of few-shot scribble-supervised medical image segmentation. The combination of a learnable superpixel network (SSN) with segmentation feedback and a spatial-prior-guided filter is genuinely not present in the cited literature, and the authors make a real attempt to justify each component. The ablation is the best part: unconstrained superpixel diffusion collapses Mean Dice to 0.311, the filter restores it to 0.774, and full bilevel training reaches 0.828. That chain makes the mechanism plausible.\n\nThe soft spot is the evaluation. Table 4 selects sigma and gamma_center on the ACDC test set, and those same values feed the results in Tables 1–2. With only five labeled training cases, that is enough to inflate a 5-point Dice gain. The few-shot split is also the first five training cases — a single, non-random split — and the reported standard deviations are across test slices, not across training runs or split choices. No significance testing. No code released. The center-of-image prior in Eqs. (8)–(14) is transparent and may be sensible for cardiac MRI and prostate T2, but it is an assumption with no validation on off-center anatomies.\n\nI do not want to overstate the problem. The internal consistency is good, and the method is not a free lunch: the authors acknowledge the dropped implicit gradient term in Eq. (22) and justify it with references. The comparison includes strong recent baselines like QMaxViT-Unet+ and ScribbleVC. But the phrase “significantly outperforms” is supportable only if the protocol is tightened. A multi-split evaluation with hyperparameters chosen on the validation set — or at least a demonstration that the advantage holds across seeds — is the minimum needed before I would trust the magnitude of the gains.\n\nWho this is for: people actively working on scribble-supervised or weakly supervised medical segmentation will want to read it, and it is a reasonable new baseline. It is not a paradigm shift, and the impact stays inside that subfield. I would send it to peer review, not desk-reject it, but I would push for a redo of the evaluation with proper splits and no test-set tuning. The paper is honestly written and the mechanism is clear; it just needs a protocol that matches its claims.","headline":"Solid method paper for scribble-supervised segmentation, but headline gains rest on test-set-tuned hyperparameters and a single non-random few-shot split, so the size of the improvement is not yet trustworthy.","tokens_in":15326,"tokens_out":2028,"would_cite":true,"duration_ms":22852,"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 five scribble-annotated cases, a bi-level superpixel–segmentation loop reaches 0.828 Mean Dice on cardiac MRI and 0.534 on prostate MRI, surpassing prior scribble-supervised methods.","keywords":["medical image segmentation","scribble supervision","few-shot learning","superpixels","pseudo-label propagation","bilevel optimization","cardiac MRI","prostate MRI"],"falsifier":"Run the same five-case protocol on a dataset with off-center organs, such as liver or kidney segmentation from CT slices where organs appear at the periphery. If Mean Dice does not degrade relative to the centered cardiac/prostate results, the center prior is not load-bearing; if it collapses toward the no-filter ablation level (about 0.311 Dice), the assumption is confirmed as essential.","tokens_in":14451,"feed_emoji":"🫀","tokens_out":3480,"duration_ms":33738,"temperature":0.7,"pith_summary":"The paper argues that the dual challenge of few-shot and scribble supervision can be broken by coupling a learnable superpixel model with a segmentation model in a bi-level loop. The superpixel model supplies region-structural priors that turn sparse strokes into dense pseudo-labels; a spatial-prior-guided filter prevents those pseudo-labels from over-expanding into errors; and the segmentation model's learned semantics are fed back to reshape the superpixels so they align with anatomy rather than just low-level image edges. On ACDC and Prostate with only five scribble-annotated training cases, this raises Mean Dice by about five percentage points over prior methods and sharply lowers boundary error (HD95). A sympathetic reader would take this as evidence that structural priors, made learnable and task-aligned, can compensate for extreme supervision scarcity.","feed_headline":"Five scribbles reach 0.828 Dice on cardiac MRI","feed_subtitle":"Learnable superpixels turn sparse strokes into dense pseudo-labels, beating prior scribble methods on two medical benchmarks.","key_machinery":"The central mechanism is a bi-level loop: an upper-level differentiable Superpixel Sampling Network (SSN) generates soft pixel-to-superpixel assignments; the lower-level segmentation model (U-Net under Mean Teacher) uses those assignments to diffuse high-confidence pseudo-labels to whole superpixels; a spatial-prior-guided adaptive filter (Eqs. 8–10) caps each superpixel's expansion ratio as a Gaussian decay of its centroid's distance from the image center, and a center-dense seed reallocation (Eqs. 13–14) concentrates superpixels near the center. Semantic feedback from the segmentation model is added to the SSN's reconstruction objective, so superpixels evolve from generic boundary detector","core_discovery":"Under a five-case scribble-supervised setting, the proposed BiSCL reports Mean Dice of 0.828 on ACDC versus 0.776 for the best prior method, and 0.534 on Prostate versus 0.478, with Mean HD95 dropping from 21.01 mm to 5.57 mm and from 32.19 mm to 16.09 mm respectively. Ablations attribute the gain to the spatial-prior-guided filtering: unconstrained superpixel diffusion collapses Mean Dice from 0.599 to 0.311, filtering restores it to 0.774, and adding the bi-level semantic feedback raises it to 0.828. The claim is that making superpixels learnable and driving them with segmentation feedback yields region priors aligned with anatomical semantics, enabling reliable dense pseudo-label generati","pith_inferences":["A direct test of the center prior: run BiSCL on a dataset with off-center target organs (e.g., liver or kidneys in CT slices where structures appear at the periphery). If performance degrades toward the no-filter level, the center assumption is confirmed as the bottleneck; if not, the filter is more flexible than the paper suggests.","Because the bilevel training deliberately drops the implicit Jacobian term in the upper-level gradient, a natural extension is a small-scale comparison against a full hypergradient estimator to quantify the bias introduced by this approximation.","The framework's reliance on low-level reconstruction plus semantic feedback could be carried over to other sparse-supervision regimes, such as point-click or box annotations, without changing the bilevel structure.","The large HD95 reduction (about 16 mm on both datasets) implies boundary quality improves substantially; a clinic-facing extension would check whether these contour changes alter downstream volume measurements in cardiac or prostate assessment."],"forward_implications":["With only five scribble-annotated training cases, BiSCL outperforms existing scribble-supervised methods on ACDC and Prostate in both region overlap (Dice) and boundary error (HD95).","The spatial-prior-guided filter is load-bearing: removing it turns superpixel diffusion from a benefit into a liability, dropping Mean Dice from 0.599 to 0.311 on ACDC.","Bilevel semantic feedback further improves the filtered diffusion, from 0.774 to 0.828 Mean Dice, indicating that superpixels aligned to segmentation semantics produce more reliable pseudo-labels.","The segmentation backbone is a plain U-Net, so the reported gains come from the supervision machinery rather than a specialized decoder, suggesting the approach transfers across architectures."],"fun_headline_variants":["BiSCL: five scribbles lift cardiac MRI Dice to 0.828","Learnable superpixels turn sparse scribbles into dense priors","Bi-level learning beats prior scribble methods on ACDC and Prostate","Spatial-prior filtering is key: without it, Dice halves in few-shot","Few-shot scribble segmentation improved by bi-level collaboration"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The spatial-prior-guided filter and superpixel seeding assume target anatomical structures sit near the image center; if that assumption fails, the filter suppresses exactly the pseudo-label expansion the method relies on.","fun_headline_variants_meta":{"raw":{"variants":["BiSCL: five scribbles lift cardiac MRI Dice to 0.828","Learnable superpixels turn sparse scribbles into dense priors","Bi-level learning beats prior scribble methods on ACDC and Prostate","Spatial-prior filtering is key: without it, Dice halves in few-shot","Few-shot scribble segmentation improved by bi-level collaboration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1096,"prompt_tokens":774,"completion_tokens":322,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":518,"tokens_out":322,"duration_ms":4613,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:25:21.982691+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same five-case protocol on a dataset with off-center organs, such as liver or kidney segmentation from CT slices where organs appear at the periphery. If Mean Dice does not degrade relative to the centered cardiac/prostate results, the center prior is not load-bearing; if it collapses toward the no-filter ablation level (about 0.311 Dice), the assumption is confirmed as essential.","supporting_citations":[],"review_version":1}