{"id":"22b9fb91-e627-41f9-aa77-a53804733b17","arxiv_id":"2507.02399","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TABNet combines triplet augmentation self-recovery with loss-weighted boundary-aware pseudo-labels, reaching 89.1% and 91.1% average Dice on MSCMRseg and ACDC, near fully supervised nnU-Net.","lead":"This paper proposes a weakly supervised medical image segmentation framework, TABNet, that combines three data augmentations (intensity, cutout, jigsaw) with self-recovery training and boundary-aware pseudo-label supervision. On two cardiac MRI datasets it reports Dice scores close to those of fully supervised nnU-Net while using only scribble annotations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported SOTA margins may be inflated because Tables 3–5 select hyperparameters and the PL fusion pair on the MSCMR data without stating that the held-out validation split was used.","rationale":"I considered the reader's BAP concern in good faith. Equation 5 does use the scribble-region cross-entropy loss as a scalar proxy for branch quality, and this proxy is not tested on held-out regions. However, Table 4 shows that replacing loss weighting with average or random weighting costs only 0.7–0.8 Dice points (89.1 to 88.3/88.4), and even the averaged-fusion variant remains above the previous QMaxViT result on MSCMR (88.0). Thus, if the BAP proxy is imperfect, it weakens a specific sub-claim about the fusion rule but does not by itself overturn the main claim of being the best weakly supervised method. The more load-bearing issue is model selection: the lambda search and branch-pair selection are reported on 'the MSCMRseg dataset' without stating that the 5-case validation fold was used, while the headline numbers are on the 15-case test fold. The selection effect visible in Tables 4 and 5 is the same order of magnitude as the 1.1-point margin over QMaxViT, so the central performance claim depends on this omitted detail. The reader's verdict of CONDITIONAL is appropriate: the paper should be accepted only after the authors clarify the validation procedure and show that their selected configuration would have been chosen without test-fold access. I therefore keep the verdict unchanged rather than moving it, and I partially agree with the reader because both the BAP proxy and the selection issue are real, but the selection issue is the more decisive one.","tokens_in":14698,"tokens_out":13974,"duration_ms":170851,"concrete_test":"Request that the authors disclose which split (training/validation/test) was used for every row of Tables 3–5 and report the Dice on the fixed 5-case MSCMR validation fold for each configuration. Then retrain the final model using only the validation fold to choose lambda1, lambda2, lambda3 and the PL fusion pair, and evaluate the resulting model on the 15-case test fold. If the validation-best configuration differs from the reported one, or if the test margin over QMaxViT shrinks by more than roughly one Dice point, the headline superiority claim should be revised; if the reported configuration is also validation-best and the margin persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 fixes the MSCMR split as 25 training / 5 validation / 15 test cases, and Section 4.3 reports the headline results on what is evidently the 15-case test fold. Section 4.4.4 then reports the lambda ablation 'on the MSCMRseg dataset' without specifying that the 5-case validation fold was used, and Table 4 uses the same MSCMR data to select the PL branch pair. If Tables 3–5 were computed on the test fold, then the reported settings (lambda1=1.0, lambda2=0.3, lambda3=0.1) and the choice of PL(y_j,y_k) were selected using the test labels. The selection effect is large enough to matter: Table 5's lambda sweep alone spans 86.5 to 89.1 Dice, Table 4's branch-pair choices span 88.4 to 89.1, and the claimed margin over QMaxViT on MSCMR is only 1.1 points (89.1 vs 88.0). The ACDC margin (91.1 vs 87.6) is larger, but it is obtained with hyperparameters and module choices transferred from MSCMR, so it does not independently validate those choices. This is an omitted methodological detail rather than evidence of misconduct: if the authors used the validation fold for selection and simply failed to say so, the concern is discharged. Without that statement, the central claim that TABNet is the best weakly supervised method, and within 0.6% of fully supervised nnU-Net, is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents TABNet, a weakly supervised medical image segmentation framework combining a triplet augmentation self-recovery module (TAS) with boundary-aware pseudo-label supervision (BAP). TAS feeds intensity, cutout, and jigsaw-augmented views into a shared-weight segmentation network and trains with cross-entropy on scribble pixels; BAP fuses branch predictions with loss-based weights and adds a boundary Dice loss. The method is evaluated on MSCMRseg and ACDC with scribble supervision, reporting average Dice of 89.1% and 91.1%, outperforming prior weakly supervised methods and approaching fully supervised nnU-Net (89.7% and 91.7%). The paper includes detailed ablations of each module and hyperparameters.","tokens_in":14974,"tokens_out":4981,"duration_ms":52933,"significance":"If the reported numbers survive closer scrutiny, TABNet is a practically useful contribution: it gives a clear, reproducible framework (code released) with competitive performance on standard benchmarks, and the ablations support the design choices of the TAS and BAP modules. The paper's main weaknesses are methodological reporting ambiguities: the split used for hyperparameter selection is not stated, and no statistical significance tests are reported for the headline claims. These issues are fixable but currently prevent full confidence in the central claim of state-of-the-art scribble-supervised segmentation.","major_comments":[{"comment":"The ablation studies in Section 4.4 select the loss weights (λ1=1.0, λ2=0.3, λ3=0.1) and the pseudo-label fusion pair PL(y_j,y_k) on the MSCMRseg dataset, but the text does not state that the 5-case validation split described in Section 4.1 was used for this selection. Since Table 5's λ-sweep spans 86.5 to 89.1 Dice and Table 4's branch-pair choices span 88.4 to 89.1 Dice, while the reported margin over QMaxViT-Unet+ is only 1.1 points (89.1 vs 88.0), selection on the 15-case test fold would materially inflate the headline result. Please state explicitly that all model and hyperparameter selection was performed on the validation split and that the test fold was used only for the final evaluation; if this was not the case, the experiments should be re-run with selection on the validation split.","section":"Section 4.4 and Tables 3-5"},{"comment":"The abstract states that TABNet 'significantly outperforms' prior weakly supervised methods, but Table 1 reports no statistical significance tests for the main comparisons. Table 2 introduces a significance marker (p<0.05) without describing the statistical test used or reporting p-values. Please add appropriate case-level paired tests (e.g., Wilcoxon signed-rank or paired t-test) across the test cases and report the resulting p-values for the key comparisons in Table 1, or remove the word 'significantly' from the claims.","section":"Abstract and Table 1"},{"comment":"The loss-weighted fusion rule in Eq. (5) weights each branch's full-image prediction using the cross-entropy loss computed only on scribble pixels. Because scribble strokes typically pass through structure interiors, a branch with low scribble loss can still have poor boundary predictions elsewhere; the fused pseudo-label in Eq. (6) can then propagate those boundary errors through the Dice loss in Eq. (8). The only evidence supporting this proxy is the 0.7–0.8 point improvement of loss-weighted fusion over average and random weighting in Table 4. Please provide a direct analysis of the fusion rule's reliability (e.g., ablating with oracle or ground-truth-based weights, or reporting the correlation between loss-based weights and region-level segmentation quality), or explicitly discuss this limitation in the text.","section":"Equation (5) and Table 4"}],"minor_comments":[{"comment":"The term 'cropping augmentation' used in Table 4 and its surrounding text should be 'cutout augmentation' for consistency with the method description.","section":"Table 4"},{"comment":"The sentence 'closely methods the performance of fully supervised models' should read 'closely matches the performance of fully supervised models.'","section":"Conclusions"},{"comment":"The name 'TAB Net' appears with a space in the abstract, while the rest of the paper uses 'TABNet'; please unify the spelling.","section":"Abstract"},{"comment":"The notation 'w_j,w_k= ...' in Eq. (5) is ambiguous; please define each weight separately as w_j and w_k with explicit formulas.","section":"Equation (5)"},{"comment":"The claim that the model has 'standard deviations below ±7%' is unclear and inconsistent with the reported standard deviations, which are actually below 0.07 in Dice units; please rephrase as 'standard deviations below 0.07'.","section":"Section 4.4.4"},{"comment":"The term 'self-recovery' overstates the mechanism, since the cross-entropy loss in Eq. (3) is applied only to scribble pixels; consider clarifying that the model is trained to maintain correct predictions on labeled pixels under heavy augmentation rather than to reconstruct the complete mask.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The central risk is the undisclosed split for hyperparameter selection; if the authors confirm that the validation fold was used, the paper is likely acceptable after adding significance tests and addressing the pseudo-label proxy concern. The code availability and thorough ablations are strengths that support a revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a coherent, carefully ablated contribution to scribble-supervised medical image segmentation. The combination of triplet augmentation (intensity, cutout, jigsaw) with self-recovery and loss-weighted pseudo-label fusion is not in prior work, and the ablations in Tables 2–5 support each design choice. The loss-weighted fusion rule in Eq. 5 is a sensible alternative to random or fixed averaging, and the authors give credit to the relevant literature. Code is promised, which is good if it actually runs.\n\nThe major issue is the stress-test concern: the hyperparameters (λ1, λ2, λ3) and the PL fusion branch pair appear to have been selected on the MSCMRseg dataset without stating that the 5-case validation split was used. Tables 3–5 all report MSCMRseg numbers, and Table 5 alone spans 86.5 to 89.1 Dice. The claimed margin over the prior state of the art on MSCMR is only 1.1 points (89.1 vs 88.0). If those tables were produced on the 15-case test fold, the central claim is not established. This is an omitted methodological detail, not necessarily misconduct, but it needs to be stated plainly. The same concern applies to the ACDC result, which uses hyperparameters and module choices transferred from MSCMR, and so does not independently validate those choices.\n\nTwo smaller issues. The abstract says \"significantly outperforms,\" but Table 1 shows no significance tests; the per-class standard deviations are around 0.03–0.05, so a 1.1-point gap is not obviously significant. And the BAP fusion rule assumes the scribble-pixel cross-entropy loss ranks the quality of each branch's full-image prediction; that is plausible but untested on held-out regions.\n\nNone of this is fatal. The method is sensible, the ablations are careful, and the weaknesses are addressable. If the authors confirm that the validation fold was used for hyperparameter selection and add significance testing for the headline comparisons, this becomes a solid acceptance. As submitted, it deserves a serious referee, but I would not accept it on the current numbers without that clarification.\n\nThis paper is for researchers working on weakly supervised segmentation, especially scribble annotation. I would cite it as a strong baseline and bring it to a reading group as an example of thorough ablation studies, but the test-set tuning question needs to be raised.","headline":"A well-ablated weak-supervision paper with a plausible new design, but the headline gains may hinge on hyperparameters tuned on the test split rather than a held-out validation fold.","tokens_in":15607,"tokens_out":2003,"would_cite":true,"duration_ms":24836,"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":"TABNet's scribble-only training reaches 89.1% and 91.1% average Dice on cardiac MRI datasets, nearly matching fully supervised nnU-Net.","keywords":["scribble-supervised segmentation","medical image segmentation","triplet augmentation","self-recovery","boundary-aware pseudo-labels","loss-weighted fusion","cardiac MRI","weak supervision"],"falsifier":"Take a test slice, compute each branch's score on just the scribble pixels and its Dice score against the full ground truth; if the branch with the lower scribble-pixel error is not the branch with the higher full-image Dice in a substantial share of slices, loss-weighted fusion is not selecting the better prediction and the reported 89.1% vs 88.3% gap over average fusion should shrink or vanish.","tokens_in":14435,"feed_emoji":"🫀","tokens_out":7264,"duration_ms":72649,"temperature":0.7,"pith_summary":"Scribble annotations are sparse lines drawn through organs, but this paper claims it can train a cardiac MRI segmentation network to nearly full-supervision quality from them alone. The proposed framework, TABNet, couples a triplet augmentation self-recovery module with a boundary-aware pseudo-label module, and reports 89.1% average Dice on MSCMRseg and 91.1% on ACDC. Those numbers beat every compared weakly supervised method and sit within 0.6–0.8% of the fully supervised nnU-Net on the same splits. If the result holds, the main annotation bottleneck for cardiac segmentation—dense pixel masks—could be relaxed to quick scribbles.","feed_headline":"Scribble-trained network hits 91.1% Dice on cardiac MRI","feed_subtitle":"Triplet augmentation plus loss-weighted pseudo-labels close the gap to fully supervised nnU-Net on ACDC.","key_machinery":"The load-bearing mechanism is the pairing of two modules. The triplet augmentation self-recovery (TAS) module builds three augmented views $x_i=C(x)$ (cutout masking the scribble-bounding box), $x_j=J(x)$ (jigsaw patch shuffle), and $x_k=I(x)$ (intensity shift), runs them through one shared-weight network, un-shuffles the jigsaw prediction with the inverse operation $J'(\\cdot)$, and supervises each branch with cross-entropy only on scribble pixels. The boundary-aware pseudo-label (BAP) module then weights the jigsaw and intensity branch predictions by the reciprocal of their scribble cross-entropy losses, $\\omega_j = \\mathcal{L}_{ce}(y_j,s)/(\\mathcal{L}_{ce}(y_j,s)+\\mathcal{L}_{ce}(y_k,s))$ and symmetrically for $\\omega_k$, forms the fused pseudo-label by class-wise argmax, and applies Dice loss between the pseudo-label and each branch plus a boundary Dice loss computed from soft-eroded boundary maps $B=\\mathrm{ReLU}(y-\\mathrm{MinPooling}(y))$. The final loss is $\\mathcal{L}=\\lambda_1\\mathcal{L}_{TAS}+\\lambda_2\\mathcal{L}_{PL}+\\lambda_3\\mathcal{L}_{BD}$ with weights 1.0, 0.3, 0.1.","core_discovery":"The paper's central claim is that two complementary mechanisms—forcing a shared network to reconstruct full masks from intensity-shifted, cutout, and jigsaw-shuffled inputs, and supervising it with pseudo-labels formed by loss-weighted fusion of the jigsaw and intensity branches plus a soft boundary consistency term—recover most of the information that dense labels provide. On MSCMRseg the model reaches 89.1% average Dice versus 89.7% for nnU-Net; on ACDC it reaches 91.1% versus 91.7%. The authors attribute the gain to the self-recovery objective for semantic completeness and to the boundary-aware pseudo-label supervision for contour precision, and the ablations support this by showing each term adds a step: base 81.9%, plus TAS 83.4%, plus pseudo-label Dice 87.8%, plus boundary loss 89.1%.","pith_inferences":["A natural next test is to move the same two modules to other sparse-supervision regimes, such as point-click or bounding-box labels; nothing in the design is cardiac-specific, but the benefit may depend on the scribble stroke pattern.","The drop from three-branch to two-branch fusion suggests the cutout branch's predictions are systematically less reliable under occlusion; gating branches by predicted uncertainty instead of scribble loss might recover the extra view's value.","The result implies the gap between weakly and fully supervised segmentation may be more about annotation geometry than about the label budget: scribbles happen to run through organ interiors, which is exactly what the fusion loss rewards, so a different sparse label layout could change the ranking of methods.","A cheaper practical extension would be to apply the loss-weighted fusion to two checkpoints of the same network instead of two augmented branches, which would test whether view diversity or model diversity is the active ingredient."],"forward_implications":["Scribble-only training becomes a practical substitute for dense annotation in cardiac MRI: on these benchmarks the gap to the best fully supervised method is under one Dice point.","The full triplet of intensity, cutout, and jigsaw augmentations is worth more than any pair: the three-branch TAS reaches 89.1% versus 88.5% for the best pair, and jigsaw alone is the strongest single branch at 86.9%.","Loss-weighted fusion is the right way to combine pseudo-label branches: it beats average fusion by 0.8 Dice points (89.1% vs 88.3%) and random fusion by 0.7 points (89.1% vs 88.4%).","Pseudo-labels should be built from jigsaw and intensity branches only: adding the cutout branch to the fusion drops performance from 89.1% to 88.5%, so the cutout view helps as a training signal but hurts as a pseudo-label source.","Boundary-aware supervision contributes at small weight: adding $\\mathcal{L}_{BD}$ with $\\lambda_3=0.1$ improves RV and Myo Dice by about two points over the no-boundary setting."],"supporting_citations":[{"why":"Supplies the ACDC dataset, its cine-MRI volumes, and the standard evaluation protocol used for the 91.1% claim.","marker":"[1]"},{"why":"Provides the fully supervised nnU-Net baseline that defines the near-oracle ceiling (89.7% MSCMRseg, 91.7% ACDC).","marker":"[11]"},{"why":"ScribFormer is one of the comparison methods TABNet must beat to support the state-of-the-art claim.","marker":"[18]"},{"why":"DMPLS is the direct pseudo-labeling predecessor whose dynamic mixed pseudo-label strategy the BAP module builds on and outperforms.","marker":"[20]"},{"why":"QMaxViT-Unet+ is the prior best weakly supervised method; beating it by 1.1% and 3.5% anchors the headline comparisons.","marker":"[21]"},{"why":"CycleMix contributes the UNet+ baseline framework and the exact train/validation/test splits used for both datasets.","marker":"[32]"},{"why":"ShapePU is a competing PU-learning method in the comparison table, supplying one of the baselines for the 89.1% claim.","marker":"[33]"},{"why":"MSCMRseg is the dataset with scribble annotations that the method's 89.1% result is measured on.","marker":"[38]"}],"fun_headline_variants":["Scribbles alone train cardiac MRI segmentation near full supervision","Boundary-aware pseudo-labels lift scribble segmentation to nnU-Net level","Triplet augmentation recovers full-mask info from scribble labels","Weakly supervised cardiac MRI segmentation rivals fully supervised","Triplet augmentation and boundary loss close scribble-label gap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pseudo-label fusion trick rests on assuming that how well a branch matches the few labeled scribble lines tells which branch's full-image prediction is better; a branch can match those lines while misplacing edges elsewhere, and the fused label would then copy that mistake.","fun_headline_variants_meta":{"raw":{"variants":["Scribbles alone train cardiac MRI segmentation near full supervision","Boundary-aware pseudo-labels lift scribble segmentation to nnU-Net level","Triplet augmentation recovers full-mask info from scribble labels","Weakly supervised cardiac MRI segmentation rivals fully supervised","Triplet augmentation and boundary loss close scribble-label gap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000922,"raw_usage":{"total_tokens":3991,"prompt_tokens":1020,"completion_tokens":2971,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":2885}},"tokens_in":636,"tokens_out":2971,"duration_ms":21868,"temperature":1.0,"reasoning_tokens":2885,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:31:11.873156+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a test slice, compute each branch's score on just the scribble pixels and its Dice score against the full ground truth; if the branch with the lower scribble-pixel error is not the branch with the higher full-image Dice in a substantial share of slices, loss-weighted fusion is not selecting the better prediction and the reported 89.1% vs 88.3% gap over average fusion should shrink or vanish.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ACDC dataset, its cine-MRI volumes, and the standard evaluation protocol used for the 91.1% claim."},{"cited_title":"Scribformer: Transformer makes CNN work better for scribble-based medical image segmentation","cited_arxiv_id":null,"evidence_quote":"ScribFormer is one of the comparison methods TABNet must beat to support the state-of-the-art claim."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DMPLS is the direct pseudo-labeling predecessor whose dynamic mixed pseudo-label strategy the BAP module builds on and outperforms."},{"cited_title":"QMaxViT-Unet+: A query-based MaxViT-Unet with edge enhancement for scribble- supervised segmentation of medical images","cited_arxiv_id":null,"evidence_quote":"QMaxViT-Unet+ is the prior best weakly supervised method; beating it by 1.1% and 3.5% anchors the headline comparisons."},{"cited_title":"Computers in Biology and Medicine 168, 107744","cited_arxiv_id":null,"evidence_quote":"CycleMix contributes the UNet+ baseline framework and the exact train/validation/test splits used for both datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ShapePU is a competing PU-learning method in the comparison table, supplying one of the baselines for the 89.1% claim."},{"cited_title":"Multivariate mixture model for myocardial seg- mentation combining multi-source images","cited_arxiv_id":null,"evidence_quote":"MSCMRseg is the dataset with scribble annotations that the method's 89.1% result is measured on."}],"review_version":1}