Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

Deep Skin Lesion Segmentation with Transformer-CNN Fusion: Toward Intelligent Skin Cancer Analysis

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read An improved TransUNet with boundary-guided attention reports the top scores for skin lesion segmentation on ISIC.

desk verdict A modest TransUNet variant whose reported gains rest on an unspecified training protocol and an under-described attention module; the comparison is not verifiable. read the letter →

arxiv 2508.14509 v1 pith:RA64ADTU submitted 2025-08-20 eess.IV cs.CV

Xin Wang , Xiaopei Zhang , Xingang Wang This is my paper · ORCID
classification eess.IVcs.CV
keywords skinlesionsegmentationTransUNettransformerboundary-guidedattentionISICdatasetsemanticdeeplearningmedicalimageanalysis
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 proposes a hybrid CNN-transformer segmentation model for skin lesions, built on the TransUNet design. It adds a boundary-guided attention module that supervises edge prediction and a multi-scale upsampling path, aiming to handle blurred boundaries, irregular shapes, and scale variation. The central claim is that, on the selected ISIC dataset, the model reaches mIoU 0.869, mDice 0.911, and mAcc 0.961, above Swin-Unet, SkinSam, Mask2former, and SegFormer. The paper also reports experiments on learning rate, optimizer, augmentation, split ratio, and batch size that support its training guidelines. If accurate, the method offers a stronger automatic segmentation tool for skin-cancer screening and could transfer to other medical image tasks.

What carries the argument

The central object is an improved TransUNet: a U-Net-style encoder-decoder that embeds a transformer for global self-attention while keeping a convolutional branch for local detail. A boundary-guided attention (BGA) module learns a boundary map under binary cross-entropy supervision and folds that boundary signal into the segmentation loss. A multi-scale upsampling path with residual fusion combines high-level transformer features with low-level convolutional features. The final loss mixes cross-entropy, Dice, and boundary terms, with weights lambda_1, lambda_2, lambda_3.

What would settle it

Retrain Swin-Unet, SkinSam, Mask2former, SegFormer, and the proposed model on the same ISIC subset with identical split, preprocessing, augmentation, and schedule, then compare mIoU/mDice/mAcc; if the proposed model's margins shrink or reverse, the central claim fails.

Watch

Extended reading notes

Core claim

The paper claims that an improved TransUNet — retaining a convolutional branch for local texture and edge details while embedding a vision transformer for global context, plus a boundary-guided attention module and multi-scale upsampling — achieves mIoU 0.869, mDice 0.911, and mAcc 0.961 on the ISIC skin-lesion dataset used. These are the highest numbers in its comparison table, ahead of Swin-Unet (0.812/0.864/0.941), SkinSam (0.825/0.875/0.946), Mask2former (0.837/0.881/0.951), and SegFormer (0.844/0.889/0.953). The paper attributes the improvement to better boundary localization from boundary supervision and to fusion of convolutional local features with transformer global modeling.

Load-bearing premise

The reported advantage assumes all comparison models were trained and evaluated under a strictly identical protocol (same ISIC subset, data split, preprocessing, augmentation, and training schedule); the paper gives no description of such a shared protocol.

Editorial extensions

If this is right

  • On the paper's evidence, the model achieves the best mIoU, mDice, and mAcc among the five compared methods on the ISIC subset.
  • The reported training recipe is concrete and reproducible: learning rate 0.001, AdamW, batch size 8, and an 85/10/5 train/validation/test split.
  • Data augmentation choices alter results substantially, with color jitter, MixUp, and Cutout giving larger gains than random rotation.
  • The hybrid design with boundary supervision is presented as transferable to other medical segmentation tasks such as retinal lesion detection and tumor boundary segmentation.

Reading between the lines

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

  • If the reported margins survive a controlled replication, the boundary-guided attention module is the most likely source of the gain; an ablation that removes only that module would quantify its contribution to the 0.869 mIoU.
  • The baselines' scores cluster closely (0.812–0.844 mIoU), so whether the 0.025–0.057 gain matters clinically depends on how the model generalizes beyond the chosen ISIC subset to other skin tones, devices, and lesion types.
  • The sensitivity results suggest the method is tuning-sensitive; a fair comparison would require per-model hyperparameter searches rather than using the proposed model's optimal settings for all methods.
  • A natural extension is to test the same boundary-supervision recipe on other boundary-blurred structures such as retinal vessels or tumor margins, where the same failure modes appear.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The manuscript proposes an improved TransUNet for skin lesion segmentation, combining convolutional and transformer branches with a boundary-guided attention (BGA) module and a multi-scale upsampling path. The authors report that the method achieves mIoU 0.869, mDice 0.911, and mAcc 0.961 on an ISIC dataset subset, outperforming Swin-Unet, SkinSam, Mask2former, and SegFormer. Additional sensitivity experiments on learning rate, optimizer, data augmentation, split ratio, and batch size are presented. The central claim is that the architectural modifications improve segmentation performance and robustness.

Significance. If the reported performance gains are real and reproducible, the work addresses a practically important problem: automated skin lesion segmentation with accurate boundary handling. The hybrid transformer-CNN direction is plausible and clinically relevant. However, the manuscript as written provides no code, no data/version specification, no controlled training/evaluation protocol, no ablation of the proposed modules, and no uncertainty quantification. The empirical evidence is therefore insufficient to establish the central claim. The paper does not currently meet the reproducibility standard expected of a journal submission, and the proposed contributions are not verified by the experiments.

major comments (5)
  1. [Section IV and Table 1] The comparison underlying the central claim is not reproducible. The text never specifies the ISIC dataset version, subset size, input resolution, train/validation/test split, augmentation ranges, training schedule, or loss weights used for Table 1. Section IV states only that images were resized, histogram-equalized, and augmented with rotation, scaling, and flipping. Figure 3 later suggests an 85/10/5 split, but it is not stated that this split was used for Table 1. Without a shared protocol, the 0.025 mIoU advantage over SegFormer (0.869 vs 0.844) cannot be attributed to the architecture rather than to training configuration or dataset selection.
  2. [Eq. (5)] The total loss is defined as L = lambda1 L_ce + lambda2 L_dice + lambda3 L_bce, but the values of lambda1, lambda2, and lambda3 are never reported. Since the comparison baselines may use different loss functions and weights, the comparison is uncontrolled. In addition, Eq. (4) defines the boundary BCE loss but does not specify how the boundary labels are generated from the segmentation masks, nor how the BGA module is integrated into the attention mechanism. These details are essential to understanding and reproducing the proposed method.
  3. [Section III and Table 1] No ablation isolates the proposed components. The claimed improvements come from the BGA module, multi-scale upsampling, and residual fusion, but Table 1 compares the full model only against unrelated baselines. A TransUNet baseline without BGA and without the multi-scale upsampling path is missing, so the contribution of the proposed modules is unverified. Without this ablation, the reported gains cannot be attributed to the specific architectural innovations.
  4. [Section V and Tables 2-3] The hyperparameter sensitivity analysis selects the 'optimal' learning rate, optimizer, and batch size based on the highest test-set metrics, with no separate validation set and no repeated runs. Tables 2 and 3 and Figures 2-4 report single performance values with no standard deviations, no number of seeds, and no confidence intervals. The final reported numbers may therefore reflect test-set overfitting during hyperparameter selection, and the claimed stability and robustness are not supported by the evidence.
  5. [Table 1 and Ref [12]] The baseline labeled 'Mask2former' cites Ref [12] as "Mask matching transformer for few-shot segmentation," which is not the Mask2Former model. If the baseline is not Mask2Former, the comparison label is misleading; if it is Mask2Former, the citation is incorrect. Either way, the comparison in Table 1 is not verifiable from the reference list.
minor comments (4)
  1. [Section V, paragraph after Fig. 4] The text says "As shown in Figure 5" when referring to the batch-size experiment; the corresponding figure is Figure 4. Figure/table callouts need to be checked throughout.
  2. [Equations (1)-(5)] Several equations contain garbled or inconsistent notation (e.g., Eq. (3)'s 1x1 convolution expression and Eq. (4)'s probability definitions). Please render the mathematics cleanly and define all symbols.
  3. [Section IV] The phrase "selected ISIC subset" needs an exact dataset version and size (e.g., ISIC 2016/2017/2018) so readers can assess representativeness and compare with existing results.
  4. [Throughout] The abbreviation "mIOU" is inconsistently written; use "mIoU" consistently. Also, the caption of Figure 2 mentions several augmentation strategies, but the corresponding text omits one (Cutout), and the figure is not clearly described.

Circularity Check

1 steps flagged · score 4.0 of 10

Reported final performance is the hyperparameter-search optimum on the same evaluation set, so the central comparison is partially circular.

  1. fitted input called prediction [Section V (Table 1 vs Tables 2-3 and Figures 3-4)]
    "The optimal setting of 0.001 achieves the highest scores—mIoU 0.869, mDice 0.911, and mAcc 0.961—balancing global structural understanding with fine boundary precision."

    The paper reports in Table 1 that the proposed model achieves mIoU 0.869, mDice 0.911, and mAcc 0.961. The same three numbers are then presented as the best results obtained by tuning the learning rate to 0.001 (Table 2), choosing AdamW as optimizer (Table 3), selecting batch size 8 (Figure 4), and using an 85–10–5 data split (Figure 3). No separate validation set or nested evaluation procedure is described. Thus the 'final' performance is not an independent prediction of the architecture; it is the maximum of a hyperparameter search evaluated on the same data used to report the result. The comparison against Swin-Unet, SkinSam, Mask2former, and SegFormer is therefore statistically forced in favor of the tuned configuration, making the claimed superiority at least partly an artifact of sel

full rationale

No definitional circularity, no load-bearing self-citation, and no ansatz-smuggled-in-via-citation was found. The architecture description is largely a re-assembly of known TransUNet components (transformer encoder, CNN decoder, attention, auxiliary boundary loss), and the performance claims are empirical rather than derived. However, one concrete circular step is present: the final reported numbers in Table 1 are exactly the best numbers found during hyperparameter sensitivity analysis on the same dataset, with no indication of a held-out validation set used to choose those hyperparameters. The reported superiority over baselines therefore reduces in part to test-set selection rather than to an independent architectural advantage. Because the central claim still has some independent content—the architecture could in principle be genuinely better—the circularity is partial, not total. This is a methodological-selection circularity, not a self-referential derivation, so the score is moderate.

Assumptions & free parameters 5 free parameters · 3 assumptions · 1 invented entities

The central claim relies on a set of chosen hyperparameters and several unstated protocol assumptions. The loss weights and training details are not given, and the fair comparison with baselines is assumed. The boundary-guided attention is introduced as an entity but its actual implementation is not specified, so its contribution is not independently evidenced.

free parameters (5)
  • Learning rate = 0.001
    Selected as best from sensitivity analysis (Table 2).
  • Optimizer = AdamW
    Selected as best from optimzer comparison (Table 3).
  • Batch size = 8
    Selected from batch size experiments (Figure 5).
  • Training/validation/test split ratio = 85/10/5
    Selected as optimal from partition ratio experiments (Figure 3).
  • Loss weights (lambda1, lambda2, lambda3) = Not reported
    Weights in total loss (Eq. 5) are never specified.
assumptions (3)
  • domain assumption The ISIC dataset provides reliable pixel-level segmentation labels annotated by dermatologists.
    The paper relies on this in Section IV to claim reliable supervision, but does not validate inter-observer variability or label quality.
  • domain assumption The boundary map derived from the segmentation mask provides useful supervision for the boundary-guided attention module.
    Eq. 4 defines a boundary loss but the mechanism by which the boundary map guides attention is not described; the assumption that a loss alone achieves guidance is unstated.
  • domain assumption All compared models were trained and evaluated under identical conditions.
    The comparison in Table 1 is meaningless without this assumption, but the paper never states the protocol used for baselines.
invented entities (1)
  • Boundary-Guided Attention (BGA) module
    purpose: Enhance boundary localization in skin lesion segmentation.
    The module is only described by a boundary loss equation (Eq. 4) and not by its internal mechanism or an ablation showing its contribution. Its effect is only claimed within this study.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Skin Lesion Segmentation with Transformer-CNN Fusion: Toward Intelligent Skin Cancer Analysis." pith.science (2026). https://pith.science/paper/RA64ADTU

@misc{pith2026250814509,
  author       = {Pith},
  title        = {Pith review of: Deep Skin Lesion Segmentation with Transformer-CNN Fusion: Toward Intelligent Skin Cancer Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RA64ADTU}},
  note         = {Machine review of arXiv:2508.14509}
}
read the original abstract

This paper proposes a high-precision semantic segmentation method based on an improved TransUNet architecture to address the challenges of complex lesion structures, blurred boundaries, and significant scale variations in skin lesion images. The method integrates a transformer module into the traditional encoder-decoder framework to model global semantic information, while retaining a convolutional branch to preserve local texture and edge features. This enhances the model's ability to perceive fine-grained structures. A boundary-guided attention mechanism and multi-scale upsampling path are also designed to improve lesion boundary localization and segmentation consistency. To verify the effectiveness of the approach, a series of experiments were conducted, including comparative studies, hyperparameter sensitivity analysis, data augmentation effects, input resolution variation, and training data split ratio tests. Experimental results show that the proposed model outperforms existing representative methods in mIoU, mDice, and mAcc, demonstrating stronger lesion recognition accuracy and robustness. In particular, the model achieves better boundary reconstruction and structural recovery in complex scenarios, making it well-suited for the key demands of automated segmentation tasks in skin lesion analysis.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Topology-Aware Graph Reinforcement Learning for Dynamic Routing in Cloud Networks

    cs.LG 2025-09 reject novelty 3.0 of 10

    A graph RL router with structure-aware state encoding and policy-driven edge rewiring reports improved throughput and latency on GEANT, based on single-run table comparisons.

Reference graph

Works this paper leans on

9 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Deep semantic segmentation and skin disease classification from dermoscopic images based on a modernized deep learning network with multi-feature extraction[J]

    Prabu M, MuthuKumar S, Sudhahar T N, et al. Deep semantic segmentation and skin disease classification from dermoscopic images based on a modernized deep learning network with multi-feature extraction[J]. Australian Journal of Electrical and Electronics Engineering,2024:1-26

    work page 2024

  2. [2]

    An artificial intelligence model for the semantic segmentation of neoplasms on imagesoftheskin[J].BiomedicalEngineering,2024,58(1):36-39

    Nikitaev V G, Pronichev A N, Nagornov O V, et al. An artificial intelligence model for the semantic segmentation of neoplasms on imagesoftheskin[J].BiomedicalEngineering,2024,58(1):36-39

    work page 2024

  3. [3]

    A modified deep semantic segmentationmodelforanalysisofwholeslideskinimages[J].Scientific Reports,2024,14(1):23489

    Asaf M Z, Rasul H, Akram M U, et al. A modified deep semantic segmentationmodelforanalysisofwholeslideskinimages[J].Scientific Reports,2024,14(1):23489

    work page 2024

  4. [4]

    Semantic segmentation in skin surface microscopicimageswithartifactsremoval[J].ComputersinBiologyand Medicine,2024,180:108975

    Aneesh R P, Zacharias J. Semantic segmentation in skin surface microscopicimageswithartifactsremoval[J].ComputersinBiologyand Medicine,2024,180:108975

    work page 2024

  5. [5]

    Convolutional neural network classification of cancer cytopathology images: taking breast cancer as an example,

    M. Xiao, Y. Li, X. Yan, M. Gao, and W. Wang, "Convolutional neural network classification of cancer cytopathology images: taking breast cancer as an example," Proceedings of the 2024 7th International Conference on Machine Vision and Applications, pp. 145–149, Singapore,Singapore,2024

    work page 2024

  6. [6]

    Improved Skin Disease Classification By Segmentation and Feature Mapping using Deep Learning[C]//2024 4th Asian Conference on Innovation in Technology (ASIANCON)

    Goindi S, Thakur K, Kapoor D S. Improved Skin Disease Classification By Segmentation and Feature Mapping using Deep Learning[C]//2024 4th Asian Conference on Innovation in Technology (ASIANCON). IEEE,2024:1-10

    work page 2024

  7. [7]

    YanY, LiJ,ZhangZ. Ahybridnetworkcombiningconvolutional neural networks and transformers for skin disease segmentation[C]//2025 4th International Symposium on Computer Applications and Information Technology(ISCAIT).IEEE,2025:896-899

    work page 2025

  8. [8]

    Singapore: Springer Nature Singapore, 2024: 372-385

    NiuZ,WangL,JiangZ,etal.ASemanticSegmentationMethodforSkin Lesion Images Based on ViT[C]//International Conference on Neural Information Processing. Singapore: Springer Nature Singapore, 2024: 372-385

    work page 2024

Show all 9 references
  1. [9]

    Survival prediction across diverse cancer types using neural networks

    X. Yan, W. Wang, M. Xiao, Y. Li, and M. Gao, "Survival prediction across diverse cancer types using neural networks", Proceedings of the 20247th International Conference on MachineVision andApplications, pp.134-138,2024. [10]CaoH, WangY, Chen J, etal. Swin-unet: Unet-likepure ...

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.