Pith. sign in

REVIEW 4 major objections 6 minor 9 references

UNet-3D with Adaptive TverskyCE Loss for Pancreas Medical Image Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict A plausible adaptive loss-fusion idea, but the experiments don't isolate it; worth a referee, not worth taking on faith. read the letter →

arxiv 2505.01951 v2 pith:BINCJDRS submitted 2025-05-04 eess.IV

Xubei Zhang , Mikhail Y. Shalaginov , Tingying Helen Zeng This is my paper · ORCID
classification eess.IV
keywords pancreassegmentationTverskylosscross-entropyadaptiveweightingUNet-3DDilatedabdominalCTclassimbalance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

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.

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 (4)
  1. [§4.4, Table 1] 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.
  2. [§4.1] 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.
  3. [Abstract, §4.4, Fig. 3] 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.
  4. [§3.2C, Eqs. (8)–(11)] 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.
minor comments (6)
  1. [Title/Abstract] The title and abstract contain missing spaces and garbled symbols, e.g., 'UNet-3DwithAdaptiveTverskyCELoss' and 'the ?2 score'; these need copyediting.
  2. [Eq. (3)] 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.
  3. [§4.2] 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.
  4. [§4.1/References] 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.
  5. [§4.1] No details are given on preprocessing, resampling, intensity normalization, or patch extraction for the CT volumes, which are important for reproducibility.
  6. [Fig. 3] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: reported DSC/F2 are held-out test measurements, and the adaptive weights are online training losses, not fitted to the reported metrics.

full rationale

The paper's central comparison in Table 1 reports Dice and F2 scores on a fixed 16-volume test split; these values are not used to define the loss in Eqs. (8)-(11), nor are they recovered by construction from any fitted parameter. The adaptive weights lambda_Tversky and lambda_CE are updated from the previous epoch's training-loss values in Eqs. (9)-(10), which is a recursive training heuristic rather than a definition of the test metric in terms of itself. The baseline comparison uses the same UNet-3D architecture with a different loss, and the Dilated V-Net row cites independent prior work [12]; there is no load-bearing self-citation chain. The absence of a fixed-weight TverskyCE ablation is a real experimental-design limitation affecting attribution of the reported gain, but it does not make the reported metric equivalent to an input by construction. No pattern of self-definition, fitted-input-called-prediction, or renaming of a known result is present, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The paper's contribution rests on four fitted choices: the adaptive weighting rule itself (which is the method), the alpha/beta settings, the unspecified learning-rate schedule, and a single random data split. No hyperparameter sensitivity, no ablations, and no formal analysis of the weighting rule are provided. The invented entity is the adaptive weighting scheme, and its only support is the single-split performance table.

free parameters (4)
  • Tversky hyperparameters alpha and beta = alpha=0.3, beta=0.7 and alpha=beta=0.5
    Two settings are explored and the better one (alpha=beta=0.5) is selected by validation/test performance, so the best reported result includes a hyperparameter choice made on the same data.
  • Adaptive loss proportions omega_Tversky and omega_CE = Not reported; each epoch t weights are L_Tversky(t-1)/(L_Tversky(t-1)+L_CE(t-1)) and L_CE(t-1)/(same denominator)
    These are the central learnable weights of the method (Eqs. 9-10). They are fitted to the training dynamics of each run, not derived from theory, and their values are not reported in the paper.
  • Initial learning rate and decay schedule details = lr 0.005, 'adaptive decay schedule' unspecified
    These training choices affect the loss trajectory and hence the adaptive weights, but the schedule itself is not specified, making it an unquantified free choice.
  • Train/validation/test split selection (56/8/16 of 80 patients) = Not identified by patient ID; random selection reported
    The test split is a single random draw, and its composition determines all reported metrics. Without seeds or patient IDs it is an uncontrolled choice.
assumptions (4)
  • ad hoc to paper The ratio of previous-epoch loss values is a meaningful indicator of which loss the model should emphasize next.
    This is the design premise of Eqs. 9-10. The paper provides heuristic justification only, with no proof, gradient analysis, or ablation that this ratio improves convergence.
  • domain assumption Tversky loss gradient fluctuation (Eqs. 6-7) is the reason for instability, and adding CE with adaptive weights fixes it.
    The paper cites [19] for gradient fluctuation but does not measure the gradients or the loss trajectory of UNet-3D to confirm the mechanism.
  • ad hoc to paper A single random 56/8/16 split of NIH Pancreas-CT is a sufficient evaluation protocol.
    The standard NIH Pancreas-CT evaluation is four-fold cross-validation; the paper departs from this without justification, which weakens comparability.
  • domain assumption The reported DSC and F2 metrics are computed on a held-out test set with consistent preprocessing.
    The paper states the split and metrics but omits preprocessing and inference details, so the reader cannot verify that all six configurations were evaluated identically.
invented entities (1)
  • Adaptive TverskyCE loss weighting rule
    purpose: Automatic reweighting of Tversky and CE losses each epoch based on previous loss proportions.
    This is a new training mechanism with no external falsifiable handle; its only evidence is the reported test-set Dice on one split.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UNet-3D with Adaptive TverskyCE Loss for Pancreas Medical Image Segmentation." pith.science (2026). https://pith.science/paper/BINCJDRS

@misc{pith2026250501951,
  author       = {Pith},
  title        = {Pith review of: UNet-3D with Adaptive TverskyCE Loss for Pancreas Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BINCJDRS}},
  note         = {Machine review of arXiv:2505.01951}
}
read the original abstract

Pancreatic cancer, which has a low survival rate, is one of the most challenging cancers to diagnose and treat effectively. Early detection through abdominal computed tomography (CT) scans is crucial, yet complicated by the pancreas' obscure anatomical position, small size, and frequent occlusion by surrounding organs. These factors make the pancreas particularly difficult to identify and segment accurately. While deep learning (DL) models have shown promise for segmentation tasks, their performance still requires significant improvement to address these challenges. In this research, we propose a novel adaptive TverskyCE loss for DL model training, which combines Tversky loss with cross-entropy loss through learnable weights. Our method enables automatic adjustment of loss contributions during training, dynamically optimizing the objective function for improved performance. All experiments were conducted on the National Institutes of Health (NIH) Pancreas-CT dataset. We evaluated the adaptive TverskyCE loss on the UNet-3D and Dilated UNet-3D, and our method achieved a Dice Similarity Coefficient (DSC) of 85.59%, with peak performance up to 95.24%, and the score of 85.14%. DSC and the score score were improved by 9.47% and 8.98% respectively compared with the baseline UNet-3D with Tversky loss for pancreas segmentation. Keywords: Pancreas segmentation, Tversky loss, Cross-entropy loss, UNet-3D, Dilated UNet-3D

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 5 canonical work pages

  1. [1]

    M3BUNet: Mobile Mean Max UNet for Pancreas Segmentation on CT-Scans

    Juwita, J., Hassan, G. M., Akhtar, N., Datta, A. M3BUNet: Mobile Mean Max UNet for PancreasSegmentationonCT-Scans,2024,arXiv:2401.10419[eess.IV]

    work page Pith review arXiv 2024

  2. [2]

    L., Lee, M., Heinrich, M., Misawa, K., Mori, K.,McDonagh, S., Hammerla, N

    Oktay, O., Schlemper, J., Folgoc, L. L., Lee, M., Heinrich, M., Misawa, K., Mori, K.,McDonagh, S., Hammerla, N. Y., Kainz, B., Glocker, B., Rueckert, D. Attention U- Net:LearningWheretoLookforthePancreas,2018,arXiv:1804.03999[cs.CV]

    arXiv 2018

  3. [3]

    An optimized two stage U-Net approach for segmentation of pancreas and pancreatic tumor, MethodsX, vol

    Ghorpade, H., Kolhar, S., Jagtap, J., Chakraborty, J. An optimized two stage U-Net approach for segmentation of pancreas and pancreatic tumor, MethodsX, vol. 13, 2024, 102995

    work page 2024

  4. [4]

    Z., Che, H., Li, H., Qian, X

    Li, J., Lin, X. Z., Che, H., Li, H., Qian, X. H. Pancreas segmentation with probabilistic map guided bi-directional recurrent UNet, Physics in Medicine & Biology, 2021, vol.66, no.11,115010

    work page 2021

  5. [5]

    Kurnaz, E., Ceylan, R. Pancreas Segmentation in Abdominal CT Images with U-Net Model,202028thSignal ProcessingandCommunicationsApplicationsConference (SIU), Gaziantep,Turkey,2020,pp.1-4,doi:10.1109/SIU49456.2020.9302180

    arXiv 2020

  6. [6]

    TD-Net: Trans-Deformer network for automatic pancreas segmentation, Neurocomputing, vol 517, 2023, pp

    Dai, S., Zhu, Y., Jiang, X., Yu, F., Lin, J., Yang, D. TD-Net: Trans-Deformer network for automatic pancreas segmentation, Neurocomputing, vol 517, 2023, pp. 279-293, https://doi.org/10.1016/j.neucom.2022.10.060

    work page doi:10.1016/j.neucom.2022.10.060 2023

  7. [7]

    An Introduction to Convolutional Neural Networks, 2015, arXiv:1511.08458[cs.NE]

    O'Shea, K., Nash, R. An Introduction to Convolutional Neural Networks, 2015, arXiv:1511.08458[cs.NE]

    arXiv 2015

  8. [8]

    Ronneberger, O., Fischer, P., Brox, T., U-Net: Convolutional Networks for Biomedical Image Segmentation, 18th International Conference on Medical Image Computing and Computer-AssistedIntervention(MICCAI),Munich,Germany,October5-9,2015

    work page 2015

Show all 9 references
  1. [9]

    Springer, Cham

    Peng,Y.,Sonka,M.,Chen,D.Z.U-Netv2:RethinkingtheSkipConnectionsofU-Netfor MedicalImageSegmentation,2024,arXiv:2311.17791[eess.IV] [10]Çiçek,Ö.,Abdulkadir,A.,Lienkamp,S.S.,Brox,T.,Ronneberger,O.3DU-Net: LearningDenseVolumetricSegmentationfromSparseAnnotation,In:Ourselin,S., Josk...

    2017 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.