{"id":"9dcbe7c1-3178-47a9-8353-90a7322b0333","arxiv_id":"2505.01951","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Adaptively weighted Tversky and cross-entropy losses improve UNet-3D pancreas segmentation Dice from 76.11% to 85.59% on the NIH Pancreas-CT dataset.","lead":"This paper combines Tversky loss with cross-entropy loss using learnable weights for 3D pancreas segmentation, reporting a Dice score of 85.59% on the NIH Pancreas-CT dataset, up from a 76.11% Tversky baseline. A generalist might read it because adaptive loss weighting is a cheap and potentially portable way to improve imbalanced medical image segmentation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No fixed-weight TverskyCE control: the claimed 9.47% DSC gain is not attributable to adaptivity, because Table 1 compares adaptive fusion only with pure Tversky loss.","rationale":"Read in good faith, the paper proposes a clear and simple mechanism: each epoch, set the fusion weights equal to the previous epoch's loss proportions, so the model emphasizes whichever loss is currently larger. This is plausible and potentially useful. For the central claim to hold, this adaptive schedule itself must improve over a fixed fusion of the same two losses at the same alpha/beta settings. The experiments do not test that: Table 1 has adaptive TverskyCE rows and a pure Tversky baseline row, but no fixed-weight TverskyCE row. The text's mention of DiceCE is unsupported by any table entry. Since CE is known to stabilize Tversky-type losses (the paper itself cites weighted fusion work [12]), the 76.11% to 85.59% jump is exactly what one might expect from adding CE and changing alpha/beta, independent of adaptivity. The reader's weakest assumption concerned the stability and usefulness of the ratio-based weights; my sharper formulation is that even if the ratio controller were perfectly stable, the evidence cannot distinguish it from constant weighting. The missing ablation is therefore load-bearing, not a minor omission. I agree with the REJECT verdict: the preprint does not demonstrate its central claim, though the underlying idea may be salvageable with a proper fixed-weight control, multiple seeds, and a release of code or at least the exact split and metrics. No ad hominem is intended; the issue is purely in the evidence structure. The verdict remains UNCHANGED relative to the reader because this concern reinforces rather than redirects the rejection.","tokens_in":8349,"tokens_out":4484,"duration_ms":48411,"concrete_test":"Retrain UNet-3D on the same 56/8/16 NIH Pancreas-CT split with Eq. (8) but constant weights lambda_Tversky = lambda_CE = 0.5 at alpha=beta=0.5, and also a second fixed setting such as lambda_Tversky = 0.7, lambda_CE = 0.3, using at least three seeds; report DSC and F2 on the same 16 test volumes. If either fixed-weight configuration reaches within roughly 0.5% DSC of 85.59%, the adaptive update in Eqs. (9)–(10) is not what drives the reported improvement, and the claim should be narrowed to 'adding CE to Tversky loss helps'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adaptive weighting via Eqs. (8)–(11) yields a 9.47% DSC improvement (85.59% vs 76.11%, Table 1). The necessary control is a fixed-weight TverskyCE baseline: Eq. (8) with constant lambdas and the same alpha/beta values. No such row appears in Table 1 or Table 2, although Section 4.4 states the method outperforms 'DiceCE'. The only external row, DilatedVNet-3D [12], uses a different architecture. Additionally, the adaptive winner uses alpha=beta=0.5 while the baseline Tversky loss uses alpha=0.3, beta=0.7, so the reported margin may reflect adding CE, changing alpha/beta, or both, rather than the adaptive schedule. The adaptive alpha=0.3/beta=0.7 row reaches 84.18% DSC, only 1.41% below the winner, so alpha/beta choice may matter as much as adaptivity. A related numerical inconsistency (abstract 95.24% vs Fig. 3 caption 95.53%) reinforces that the quantitative reporting is not yet reliable. The idea is plausible, but the experiments as presented underdetermine the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an adaptive TverskyCE loss for 3D pancreas segmentation, fusing Tversky loss and cross-entropy loss with epoch-dependent weights computed from the previous epoch's loss values (Eqs. 8–11). The method is evaluated on the NIH Pancreas-CT dataset with UNet-3D and Dilated UNet-3D backbones, using 56/8/16 train/validation/test splits. The best configuration, UNet-3D with adaptive TverskyCE loss at α=β=0.5, is reported to achieve DSC 85.59% and F2 85.14%, improving DSC by 9.47% and F2 by 8.98% over the Tversky-loss baseline (Section 4.4, Table 1). The paper claims that the adaptive weighting automatically balances the two losses and improves convergence.","tokens_in":8664,"tokens_out":3978,"duration_ms":38930,"significance":"If the central claim were fully supported, an adaptive fusion of Tversky and cross-entropy losses would be a simple and potentially useful contribution to imbalanced 3D medical image segmentation. The authors should be credited for using a public dataset, reporting six model/loss configurations, and providing standard metrics (DSC, F2, sensitivity, specificity, precision). However, the evidence as presented is insufficient: the experiment lacks a fixed-weight control, relies on a single random split with no statistical analysis, and contains numerical inconsistencies. The proposed weighting rule is heuristic and its behavior is not analyzed. The idea is plausible, but the experiments do not yet demonstrate that the adaptive mechanism itself is responsible for the reported gains.","major_comments":[{"comment":"The central claim that the adaptive schedule yields a 9.47% DSC improvement is not supported because no fixed-weight TverskyCE baseline is reported. The comparison between the adaptive α=β=0.5 TverskyCE result and the UNet-3D Tversky-loss baseline (α=0.3, β=0.7) conflates three changes: the addition of the cross-entropy term, the change in α/β values, and the adaptive weighting mechanism. A control row with constant weights, e.g., ω_Tversky = ω_CE = 0.5 and α=β=0.5, is required to attribute the gain to adaptivity. The paper also states in §4.4 that the method outperforms 'DiceCE', but no DiceCE row appears in Table 1 or Table 2.","section":"§4.4, Table 1"},{"comment":"The evaluation uses a single random split of 80 patients into 56/8/16 training/validation/test. With only 16 test volumes, the reported 9.47% DSC improvement could easily be an artifact of split selection. No cross-validation, bootstrap confidence intervals, or statistical significance testing is provided. The manuscript should report results over multiple splits or give per-case standard deviations and a paired test, such as a Wilcoxon signed-rank test, between the proposed method and the baseline.","section":"§4.1"},{"comment":"There are numerical inconsistencies that undermine confidence in the reported results. The abstract states peak DSC of 95.24%, while the caption of Fig. 3 states 95.53%. Additionally, §4.2 lists three loss configurations for each architecture, but Table 1 labels the Dilated UNet-3D α=β=0.5 row as 'TverskyCEloss' rather than 'adaptive TverskyCE loss', and Table 2 contains an entry '84' without a decimal value. These discrepancies need to be corrected and clarified.","section":"Abstract, §4.4, Fig. 3"},{"comment":"The adaptive weighting rule assumes that the ratio of the previous-epoch loss values is a meaningful and stable controller of training dynamics, but this assumption is not justified. The text itself notes that Tversky-loss gradients fluctuate (Eqs. 6–7), yet no analysis, ablation with fixed weights, or sensitivity study shows that the proposed ratio-based weights mitigate that fluctuation rather than amplify it. Furthermore, the abstract describes the weights as 'learnable', but they are computed deterministically from previous losses, not learned through backpropagation. The authors should provide an ablation, plot the evolution of ω_Tversky during training, and discuss under what conditions the ratio rule is stable.","section":"§3.2C, Eqs. (8)–(11)"}],"minor_comments":[{"comment":"The title and abstract contain missing spaces and garbled symbols, e.g., 'UNet-3DwithAdaptiveTverskyCELoss' and 'the ?2 score'; these need copyediting.","section":"Title/Abstract"},{"comment":"Equation (3) is difficult to parse because of poor formatting; the sums and Tversky index terms should be typeset with clear subscripts and separated terms.","section":"Eq. (3)"},{"comment":"The sentence 'In the place of max-pooling layers, the convolutional kernel size is 2x2x2' is unclear; presumably the authors mean strided convolutions replace max-pooling, but this should be stated explicitly.","section":"§4.2"},{"comment":"Reference [20] is cited for the NIH Pancreas-CT dataset, but the listed citation describes the TCGA-GBM collection. The correct dataset reference should be provided.","section":"§4.1/References"},{"comment":"No details are given on preprocessing, resampling, intensity normalization, or patch extraction for the CT volumes, which are important for reproducibility.","section":"§4.1"},{"comment":"The abstract reports a peak DSC of 95.24% while Fig. 3 shows 95.53%; the authors should clarify whether these refer to the same or different test cases.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible technical report, but the central claim requires substantial additional experiments: a fixed-weight TverskyCE control, multi-split or cross-validated evaluation with statistical tests, and an analysis of the adaptive weight dynamics. The numerical inconsistencies and the mismatched dataset reference suggest that the manuscript needs careful proofreading. The idea may be worth publishing if these points are addressed, but in its current form the evidence is not convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: the adaptive TverskyCE rule—weight each loss by its previous-epoch share of total loss—is genuinely new as far as I can tell, and the paper is honest about its lineage (Weighted Fusion Loss, adaptive t-vMF Dice). The idea is simple and plausible: if Tversky loss fluctuates and CE is smooth, letting the proportions move with loss magnitudes could stabilize training. The authors do not oversell the literature or bury prior work.\n\nWhat's good: they give the update equations, run six configurations on a public dataset, and report the usual metrics. The related-work handling is fair. No circularity: the test metrics are on a held-out split.\n\nBut the evidence for the central claim is thin. Table 1 compares adaptive TverskyCE only against pure Tversky loss (and against DilatedVNet from a different architecture). There is no fixed-weight TverskyCE row. So the reported 9.47% DSC gain could come from adding CE, from changing alpha/beta (the baseline uses alpha=0.3, beta=0.7 while the winner uses 0.5/0.5), or from the adaptive schedule. The adaptive alpha=0.3/beta=0.7 row lands at 84.18%, only 1.41% below the winner, so hyperparameter choice may matter as much as adaptivity. There's also a numerical inconsistency: abstract says peak DSC 95.24%, Fig. 3 caption says 95.53%. And there's a single 56/8/16 split with no significance or error bars. These aren't fatal to the idea, but they mean the paper's headline claim is underdetermined.\n\nMy take: the adaptive mechanism is worth testing, but this preprint doesn't demonstrate it. The minimal fix is a fixed-weight TverskyCE baseline at the same alpha/beta, plus a couple of seeds or a significance check.\n\nWho gets value: people working on loss design for imbalanced segmentation would find the idea worth a look, but as a training technique to adopt, not yet. I'd give it a referee slot, because the question—whether loss proportions are a useful control signal—is legitimate and cheap to test. I just wouldn't trust the numbers as they stand.\n\nRecommendation: send to peer review, but the authors will need to add the missing control before any acceptance.","headline":"A plausible adaptive loss-fusion idea, but the experiments don't isolate it; worth a referee, not worth taking on faith.","tokens_in":9184,"tokens_out":1848,"would_cite":false,"duration_ms":17776,"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":"An adaptive blend of Tversky and cross-entropy losses raises pancreas segmentation DSC from 76.1% to 85.6% on a 16-volume test split of the Pancreas-CT dataset.","keywords":["pancreas segmentation","Tversky loss","cross-entropy loss","adaptive loss weighting","UNet-3D","Dilated UNet-3D","abdominal CT","class imbalance"],"falsifier":"Train the identical UNet-3D on the same data split with a fixed-weight Tversky-plus-cross-entropy sum, for example constant 0.5/0.5 weights, and compare DSC; if the fixed fusion matches 85.59% within run-to-run noise, the adaptive weighting is not what produces the gain. A second check is to rerun the adaptive scheme under 5-fold cross-validation of the 80 volumes, in which case the claimed margin over the Tversky baseline would have to persist to keep the conclusion.","tokens_in":8129,"feed_emoji":"🩻","tokens_out":12233,"duration_ms":105203,"temperature":0.7,"pith_summary":"The paper seeks to show that a loss function adaptively mixing Tversky loss and cross-entropy loss improves 3D pancreas segmentation over either loss alone or a fixed fusion. The proposed adaptive TverskyCE loss reweights the two components each epoch according to their previous-epoch values, so the network automatically pays more attention to whichever term is lagging. On a 16-volume test split of the Pancreas-CT dataset, UNet-3D trained with this loss reaches a Dice Similarity Coefficient of 85.59% and an $F_2$ score of 85.14%, against 76.11% and 76.16% for the same network with Tversky loss alone. If the results hold, they suggest that dynamic loss weighting is a low-cost route to substantially better delineation of a small, low-contrast, class-imbalanced organ.","feed_headline":"Adaptive loss lifts pancreas segmentation Dice to 85.6%","feed_subtitle":"Adaptive loss weighting shifts between overlap and classification terms, gaining 9.5 Dice points on pancreas CT scans.","key_machinery":"The carrying mechanism is the adaptive TverskyCE loss: a weighted sum of Tversky loss and cross-entropy loss, $L_{\\mathrm{total}}(t)=w_{\\mathrm{T}}(t)L_{\\mathrm{T}}(t)+w_{\\mathrm{CE}}(t)L_{\\mathrm{CE}}(t)$, whose weights at epoch $t$ are set to the fractions of the two losses in the total at epoch $t-1$, summing to one. The Tversky term includes $\\alpha$ and $\\beta$, which control penalties for false positives and false negatives; $\\alpha=\\beta=0.5$ makes it the Dice coefficient, and the paper trains the Tversky-only baseline with $\\beta=0.7$ to favor recall. The adaptive weighting does the work of automatically reallocating learning effort: when one component's loss is large relative to the other, the next epoch increases that component's weight, so the model learns more about the task it is doing worse on.","core_discovery":"The central claim is that letting the fusion weights of Tversky and cross-entropy losses evolve during training yields a large improvement on pancreas segmentation. With $\\alpha=\\beta=0.5$, the adaptive TverskyCE loss gives UNet-3D a Dice Similarity Coefficient of 85.59% and an $F_2$ score of 85.14%, improvements of 9.47 and 8.98 percentage points over the Tversky-only baseline (76.11% DSC, 76.16% $F_2$) on the same test volumes. The authors attribute the gain to the cross-entropy term stabilizing the fluctuating Tversky gradients while the adaptive weighting shifts emphasis toward whichever loss is currently larger, letting the model address the dominant error mode. The best configuration also balances recall and precision (sensitivity 86.09%, precision 95.36%), supporting the claim that the loss mitigates both under- and over-segmentation.","pith_inferences":["Because the paper reports no fixed-weight fusion ablation, the 9.47-point gain cannot yet be assigned to the adaptivity rule itself; a constant 50/50 mixture of the same two losses might perform just as well.","With only 16 test volumes, the reported margin probably carries large variance; a standard 5-fold cross-validation on the 80 volumes, or evaluation on an external pancreas dataset, would be expected to shrink or reshape the margin.","A direct test of the mechanism would log both loss terms and the weight trajectory over training: a genuinely adaptive controller should visibly shift weight when one loss stagnates, rather than simply mirror relative magnitudes."],"forward_implications":["A pure loss-function change, with no architectural modification, lifts UNet-3D DSC by 9.47 points on the test split, indicating that loss design is a primary lever for small-organ segmentation.","The adaptive weights remove the need to hand-tune the fusion coefficient that fixed-ratio combinations such as TverskyCE or weighted fusion require.","The best configuration raises $F_2$ and precision together (95.36% vs 81.55% for the Tversky baseline), implying better boundary delineation rather than a simple precision-recall trade-off.","Dilated UNet-3D also improves with the adaptive loss (DSC 83.4% vs 72.8% for Tversky alone), so the benefit is not specific to one architecture."],"supporting_citations":[{"why":"Defines the Tversky loss/index that the adaptive loss builds on and that serves as the Tversky-only baseline.","marker":"[11]"},{"why":"Cited for the claim that Tversky-loss gradients fluctuate, which motivates adding a cross-entropy term for stability.","marker":"[19]"},{"why":"Provides the fixed-weight fusion-loss precedent (Dice plus cross-entropy) and the Dilated V-Net comparison; the new method contrasts its automatic weights with manually set ones.","marker":"[12]"},{"why":"Defines 3D U-Net, the base architecture in which the adaptive loss is evaluated.","marker":"[10]"},{"why":"Introduces the U-Net architecture and its encoder-decoder with skip connections, the design extended to 3D.","marker":"[8]"},{"why":"Cited as the source of the CT volumes used for all experiments.","marker":"[20]"}],"fun_headline_variants":["Adaptive loss fusion lifts pancreas Dice by 9.5 points","Learnable loss weights add 9.5 Dice points to pancreas segmentation","Dynamic loss weighting raises pancreas Dice to 85.6%","Adaptive TverskyCE loss adds 9.5 points to pancreas Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the ratio of the two previous-epoch loss values is a dependable guide to how much each loss should count in the next epoch; if that ratio mostly reflects how large the losses happen to be rather than how well the model is learning, the claimed 9.47-point gain collapses.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive loss fusion lifts pancreas Dice by 9.5 points","Learnable loss weights add 9.5 Dice points to pancreas segmentation","Dynamic loss weighting raises pancreas Dice to 85.6%","Adaptive TverskyCE loss adds 9.5 points to pancreas Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000803,"raw_usage":{"total_tokens":3564,"prompt_tokens":1016,"completion_tokens":2548,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":2470}},"tokens_in":632,"tokens_out":2548,"duration_ms":18303,"temperature":1.0,"reasoning_tokens":2470,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:04:26.389654+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical UNet-3D on the same data split with a fixed-weight Tversky-plus-cross-entropy sum, for example constant 0.5/0.5 weights, and compare DSC; if the fixed fusion matches 85.59% within run-to-run noise, the adaptive weighting is not what produces the gain. A second check is to rerun the adaptive scheme under 5-fold cross-validation of the 80 volumes, in which case the claimed margin over the Tversky baseline would have to persist to keep the conclusion.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the U-Net architecture and its encoder-decoder with skip connections, the design extended to 3D."}],"review_version":1}