Pith. sign in

REVIEW 3 major objections 7 minor 23 references

HES-UNet: A U-Net for Hepatic Echinococcosis Lesion Segmentation

T0 review · 3 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read HES-UNet, a U-Net variant with multi-directional downsampling and multi-scale aggregation, segments hepatic echinococcosis lesions in CT scans with a Dice Similarity Coefficient of 89.21%, outperforming six baselines including TransUNet…

desk verdict A plausible U-Net variant for hepatic echinococcosis segmentation, but the headline SOTA claim is contradicted by the paper's own ablation table. read the letter →

arxiv 2412.06530 v1 pith:L3IUJSQM submitted 2024-12-09 eess.IV cs.AIcs.CV

Jiayan Chen , Kai Li , Zhanjin Wang , Zhan Wang , Jianqiang Huang This is my paper · ORCID
classification eess.IVcs.AIcs.CV
keywords hepaticechinococcosismedicalimagesegmentationU-Netcomputedtomographymulti-scalefeaturefusionattentionmechanismdeepsupervisionDicesimilaritycoefficient
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

This paper aims to establish that a U-Net variant called HES-UNet, built around three new feature-fusion blocks, can segment hepatic echinococcosis lesions in abdominal CT scans more accurately than six existing segmentation models. On a private dataset of CT slices from 268 patients, HES-UNet reports a Dice Similarity Coefficient of 89.21%, outperforming the strongest baseline, TransUNet, by 1.09 percentage points. The authors argue that the gain comes from preserving high- and low-frequency detail during downsampling, aggregating features across all encoder levels, and letting each decoder stage receive multi-scale global features. If the result holds, it would give a practical tool for automated lesion delineation in regions where echinococcosis is common but specialized medical staff are scarce.

What carries the argument

The load-bearing mechanism is a set of three feature-fusion blocks grafted onto the U-Net skeleton. The multi-directional downsampling block (MDB) applies a Haar wavelet transform to split each feature map into low-frequency and three directional high-frequency parts, then concatenates and compresses them with a 1x1 convolution, so downsampling does not discard boundary detail. The multi-scale aggregation block (MAB) uses adaptive average pooling to resize all encoder outputs to one resolution, adds them, and passes the deepest scale through a GHPA attention module, producing a single global feature tensor. The multi-scale upsampling block (MUB) expands that tensor with group-wise attention and pixel shuffle, generating multi-scale global features that are fused with encoder features and intermediate predictions inside five global attention modules (GAMs). These modules, plus deep supervision, carry the claimed accuracy gain.

What would settle it

A concrete check is to re-split the dataset by patient (e.g., 80% of the 268 patients for training, 10% validation, 10% test) and rerun the identical HES-UNet training protocol; if the DSC drops well below 89.21%, the reported state-of-the-art result is an artifact of slice-level leakage rather than a property of the model.

Watch

Extended reading notes

Core claim

The central claim is that HES-UNet, a symmetric encoder-decoder with 32 to 512 channels, achieves state-of-the-art segmentation of hepatic echinococcosis lesions on a collected CT dataset, reaching a DSC of 89.21%, precision of 88.14%, and recall of 89.60%. These numbers beat U-Net, U-Net++, TransUNet, Swin-UNet, Res-UNet, and EGE-UNet on the same test slices. The paper attributes the improvement to three components added to the U-Net backbone: a multi-directional downsampling block using Haar wavelet decomposition to keep detail, a multi-scale aggregation block that pools encoder features of all five scales into a global representation, and a multi-scale upsampling block with pixel shuffle that feeds multi-scale global features into skip connections. A deep supervision module computes Dice plus binary cross-entropy losses at five decoder stages. The paper concludes that the architecture is an effective auxiliary diagnostic tool for echinococcosis in resource-poor areas.

Load-bearing premise

The CT slices were randomly split 80/10/10 without grouping by patient, so slices from the same patient may appear in both training and test sets, meaning the reported Dice could reflect memorization of patients rather than generalization to new patients.

Editorial extensions

If this is right

  • If the reported DSC is reproducible, HES-UNet could serve as a computer-aided detection tool for hepatic echinococcosis, reducing the time and expertise needed to outline lesions on CT.
  • The MDB's wavelet-based downsampling could be applied to other segmentation tasks with fine boundaries and low contrast, such as liver tumors or kidney stones, where max-pooling loses detail.
  • The MAB-MUB pairing suggests a general recipe for feeding multi-scale global context back into skip connections, potentially improving any U-Net-style architecture without a full Transformer.
  • Deep supervision at five decoder stages provides a training signal that may make the model converge faster and more stably on small medical datasets.

Reading between the lines

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

  • Because the dataset split is by slice rather than by patient, slices from the same patient likely appear in both training and test sets; the reported 89.21% DSC is therefore probably optimistic for unseen-patient generalization, and the comparison against baselines may be biased in the model's favor if it memorizes patient-specific appearance.
  • The 1.09-point lead over TransUNet is small relative to typical slice-level variance, so the practical significance of the improvement would need confirmation on a patient-level split or an external dataset.
  • The proposed modules are architecture-agnostic, so one could test them on public datasets such as liver or spleen segmentation to see whether the gain generalizes beyond this private echinococcosis collection.
  • If the authors release slice-level metadata with patient identifiers, independent groups could quantify the leakage and re-run the comparison with a grouped split; without that, the state-of-the-art claim remains tied to a single site's data.
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

3 major / 7 minor

Summary. The paper proposes HES-UNet, a U-Net-style architecture for segmenting hepatic echinococcosis (HE) lesions in CT images. The architecture combines convolutional and attention components, introducing three modules: a multi-directional downsampling block (MDB) based on Haar wavelets, a multi-scale aggregation block (MAB), and a multi-scale upsampling block (MUB) that feeds a global feature filtering module and a progressive fusion decoder with deep supervision. The authors collected a private CT dataset from 268 HE patients (137 CE, 131 AE), performed a slice-level 80/10/10 train/validation/test split, and report that HES-UNet achieves a DSC of 89.21%, claiming a 1.09% improvement over TransUNet (88.12%). They include comparisons with U-Net, U-Net++, TransUNet, Swin-UNet, Res-UNet, and EGE-UNet, plus an ablation study over the three proposed modules.

Significance. Hepatic echinococcosis segmentation is a clinically relevant problem, especially in resource-limited pastoral regions, and a well-performing automated method would be of practical value. The paper proposes a reasonably motivated architecture that combines wavelet-based downsampling, multi-scale aggregation, global attention filtering, and deep supervision; if the reported results were credible, the work would provide a useful reference for applying such hybrid CNN-attention designs to HE lesions. The paper also supplies a project page and describes hyperparameters and the training protocol, which aids reproducibility. However, the central performance claim is currently not substantiated because of an internal inconsistency between the two main result tables, and the evaluation protocol has design issues (slice-level split, no statistical tests) that affect the validity of the claimed state-of-the-art margin.

major comments (3)
  1. [Abstract / Section V / Tables I and II] The headline claim is undermined by an internal inconsistency between the two result tables. Table I reports HES-UNet at DSC 89.21%, precision 88.14%, recall 89.60%, and the 1.09% margin over TransUNet is derived from these numbers. Table II, which ablates the proposed modules on the same dataset, reports the full model (MDB ✓, MUB ✓, MAB ✓) at DSC 88.13%, precision 87.84%, recall 88.37%. The difference of 1.08 percentage points is almost exactly the claimed margin over TransUNet. Notably, the U-Net baseline row is identical in both tables (86.58 DSC), suggesting the same evaluation protocol was used. The authors provide no explanation for which number is correct. If Table II is correct, the claimed state-of-the-art advantage collapses to a 0.01% difference over TransUNet. This is a load-bearing discrepancy that must be resolved before the central claim can be accepted; the authors should report the source of each number, the protocol used, and ideally per-run or per-patient variation.
  2. [Section IV.A (Dataset and Implementation Details)] The data are split at the slice level, not the patient level: 'we saved the CT slice data, using 80% for the training set and 10% each for the testing and validation sets.' In a multi-slice CT dataset, slices from the same patient will appear in both training and test sets. This introduces potential patient-specific leakage: the model may memorize appearance or lesion characteristics that are shared across slices of the same patient, inflating the reported DSC and making the comparison measure intra-patient slice prediction rather than generalization to new patients, which is the clinically relevant task. The authors should perform a patient-level split (or otherwise ensure that no patient contributes slices to both training and test), and they should report the number of patients and slices in each split.
  3. [Tables I and II / Section IV.B] No error bars, standard deviations, or statistical significance tests are reported for any metric. The claimed improvements over baselines are small (1.09% over TransUNet, 2.63% over U-Net, 2.45% over U-Net++), and with a single evaluation run on a private dataset, these differences could easily be within run-to-run or patient-subset variation. At minimum, the authors should report the mean and standard deviation over multiple training seeds and a paired statistical test (e.g., Wilcoxon signed-rank at the patient level, or a bootstrap over patients) to establish that the differences are not due to chance.
minor comments (7)
  1. [Section IV.B] The sentence reporting improvements is inaccurate: 'HES-UNet achieved improvements of 2.63%, 1.56%, and 1.09% in the DSC metric compared to CNN-based segmentation methods (U-Net and U-Net++), the Transformer-based Swin-UNet, and the hybrid architecture TransUNet, respectively.' The U-Net++ difference is 2.45%, not 1.56%; the three listed percentages correspond to U-Net, Swin-UNet, and TransUNet, so the model list and values do not align.
  2. [Table I caption] The caption contains a typo: 'THE COMPARISON RESULTS BETWEEN OUR HES-UN ET AND OTHER MODELS' has an unwanted space in 'HES-UNET' and inconsistent capitalization.
  3. [Section III.A] There is a formatting issue: 'H = 512and W = 512represent' should be 'H = 512 and W = 512 represent'.
  4. [Section IV.A] The manuscript states that patient-identifying metadata were removed but does not mention ethics approval or institutional review board clearance for the collection and use of patient CT data. For a medical-imaging study, an explicit statement about ethics approval (or a justification for its absence) is expected.
  5. [Section III.B] The MDB is described in the contributions as 'lossless feature downsampling,' but the module includes a 1x1 convolution after concatenation of wavelet sub-bands, which is not invertible in general. The claim is overstated; consider using a term such as 'detail-preserving' or 'information-preserving.'
  6. [References [19], [20]] The design of MUB and GAM is inspired by two self-citations ([19], [20]) from prior speech-separation work. The authors should state more explicitly how these modules are adapted from the cited works and how the adaptation differs, to help readers assess novelty.
  7. [Section III.E] The deep supervision loss weights λ_i are set to {0.1, 0.2, 0.3, 0.4, 0.5, 1} without any sensitivity analysis or description of how these values were chosen. A brief justification or a sensitivity check would strengthen the presentation, although this is not a blocking issue.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central SOTA claim is an empirical comparison, and the architecture's components are fully specified rather than derived from their own outputs.

full rationale

The paper's central claim is empirical: HES-UNet is reported at 89.21% DSC against named baselines on a private 268-patient CT dataset (Section IV.B, Table I). No parameter is fitted to a subset and then renamed a prediction, and no equation reduces to its own input. The architecture is specified constructively: MDB uses Haar wavelet transforms and 1x1 convolution (Section III.B), MAB uses adaptive pooling, 1x1 convolution, and matrix addition (Section III.B), MUB uses GHPA and pixel shuffle (Section III.C), and GAMs use channel-wise grouping, concatenation, layer normalization, and dilated convolution (Section III.C). The loss (Eq. 3) is a standard weighted Dice plus BCE deep-supervision objective. Self-citations [19] and [20] by coauthor Kai Li are used only as design references for MDB and GAMs; the modules are implemented and ablated in this paper, so the citations are not load-bearing and do not force the SOTA result. The important weaknesses are empirical, not circular: the slice-level rather than patient-level split (Section IV.A) risks information leakage, and Table I (89.21% for HES-UNet) disagrees with the full-ablation row of Table II (88.13% for the same architecture), which would erase the claimed 1.09% margin over TransUNet if Table II is correct. These are correctness and reproducibility concerns, not circular derivation, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on hand-chosen preprocessing and loss weights, plus assumptions about annotation reliability and the appropriateness of the windowing range. No new physical or mathematical entities are introduced; the named modules are compositions of existing operations.

free parameters (2)
  • Deep supervision loss weights lambda_i = {0.1, 0.2, 0.3, 0.4, 0.5, 1}
    Hand-chosen in Eq. (3) to balance losses from different decoder stages; affects which model checkpoint is selected by early stopping.
  • CT window width and level = 150 HU width, 35 HU level
    Chosen with physician guidance in Section IV.A to standardize HU values to [-150, 35]; changes the input representation and therefore all results.
assumptions (3)
  • standard math Haar wavelet transform separates an image into four directional subbands and preserves information sufficient for the subsequent 1x1 convolution to reconstruct a useful downsampled feature map.
    Invoked in Section III.B as the basis of the MDB downsampling block; the lossless claim is not formally proven and is contradicted by the lossy conv-BN-ReLU steps.
  • domain assumption The ground-truth lesion annotations by three physicians are reliable and consistent.
    Section IV.A states the dataset was collected and labeled under physician guidance, but no inter-observer agreement or annotation protocol is reported.
  • domain assumption The chosen windowing range [-150, 35] preserves all lesion-relevant information while suppressing bone and noise.
    Section IV.A states this range was chosen for liver contours, lesions, and bone, but no sensitivity analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HES-UNet: A U-Net for Hepatic Echinococcosis Lesion Segmentation." pith.science (2026). https://pith.science/paper/L3IUJSQM

@misc{pith2026241206530,
  author       = {Pith},
  title        = {Pith review of: HES-UNet: A U-Net for Hepatic Echinococcosis Lesion Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L3IUJSQM}},
  note         = {Machine review of arXiv:2412.06530}
}
read the original abstract

Hepatic echinococcosis (HE) is a prevalent disease in economically underdeveloped pastoral areas, where adequate medical resources are usually lacking. Existing methods often ignore multi-scale feature fusion or focus only on feature fusion between adjacent levels, which may lead to insufficient feature fusion. To address these issues, we propose HES-UNet, an efficient and accurate model for HE lesion segmentation. This model combines convolutional layers and attention modules to capture local and global features. During downsampling, the multi-directional downsampling block (MDB) is employed to integrate high-frequency and low-frequency features, effectively extracting image details. The multi-scale aggregation block (MAB) aggregates multi-scale feature information. In contrast, the multi-scale upsampling Block (MUB) learns highly abstract features and supplies this information to the skip connection module to fuse multi-scale features. Due to the distinct regional characteristics of HE, there is currently no publicly available high-quality dataset for training our model. We collected CT slice data from 268 patients at a certain hospital to train and evaluate the model. The experimental results show that HES-UNet achieves state-of-the-art performance on our dataset, achieving an overall Dice Similarity Coefficient (DSC) of 89.21%, which is 1.09% higher than that of TransUNet. The project page is available at https://chenjiayan-qhu.github.io/HES-UNet-page.

Figures

Figures reproduced from arXiv: 2412.06530 by the authors.

Figure 1
Figure 1. The overall pipeline of HES-UNet. HES-UNet consists of four main components: MFSI encoder, MGF module, PF decoder, and DS module. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The structure of each module in HES-UNet: (A) Encoder Block, (B) Decoder Block, (C) Multi-scale Aggregation Block, (D) Multi-scale Upsampling [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Our segmentation results compared with other models. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 13 canonical work pages

  1. [1]

    Echinococcosis of the liver,

    Benedikt V Czermak, Okan Akhan, Renate Hiemetzberger, Bettina Zelger, Wolfgang V ogel, Werner Jaschke, Michael Rieger, Sang Yoon Kim, and Jae Hoon Lim, “Echinococcosis of the liver,” Abdominal imaging, vol. 33, pp. 133–143, 2008

    work page 2008

  2. [2]

    Advances in liver echinococcosis: diagnosis and treat- ment,

    Dimitrios Filippou, Dimitrios Tselepis, Georgios Filippou, and Vassilios Papadopoulos, “Advances in liver echinococcosis: diagnosis and treat- ment,” Clinical Gastroenterology and Hepatology , vol. 5, no. 2, pp. 152–159, 2007

    work page 2007

  3. [3]

    A computational approach to edge detection,

    John Canny, “A computational approach to edge detection,” IEEE Transactions on pattern analysis and machine intelligence , , no. 6, pp. 679–698, 1986

    work page 1986

  4. [4]

    Seeded region growing,

    Rolf Adams and Leanne Bischof, “Seeded region growing,” IEEE Transactions on pattern analysis and machine intelligence , vol. 16, no. 6, pp. 641–647, 1994

    work page 1994

  5. [5]

    Gradient-based learning applied to document recognition,

    Yann LeCun, L ´eon Bottou, Yoshua Bengio, and Patrick Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

    1998

  6. [6]

    U-net: Con- volutional networks for biomedical image segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Con- volutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 . Springer, 2015, pp. 234–241

    2015

  7. [7]

    Transunet: Transformers make strong encoders for medical image segmentation,

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou, “Transunet: Transformers make strong encoders for medical image segmentation,” arXiv preprint arXiv:2102.04306, 2021

    arXiv 2021

  8. [8]

    Attention is all you need,

    A Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

Show all 23 references
  1. [9]

    Swin-unet: Unet-like pure transformer for medical image segmentation,

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang, “Swin-unet: Unet-like pure transformer for medical image segmentation,” in European conference on computer vision. Springer, 2022, pp. 205–218

    2022

  2. [10]

    3d u-net: learning dense volumetric seg- mentation from sparse annotation,

    ¨Ozg¨un C ¸ ic ¸ek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger, “3d u-net: learning dense volumetric seg- mentation from sparse annotation,” in Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, At...

    2016

  3. [11]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation,

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi, “V-net: Fully convolutional neural networks for volumetric medical image segmentation,” in 2016 fourth international conference on 3D vision (3DV). Ieee, 2016, pp. 565–571

    2016

  4. [12]

    Unet++: A nested u-net architecture for medical image segmentation,

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang, “Unet++: A nested u-net architecture for medical image segmentation,” in Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th Interna- tional Workshop,...

    2018

  5. [13]

    Msa2net: Multi-scale adaptive attention-guided network for medical image segmentation,

    Sina Ghorbani Kolahi, Seyed Kamal Chaharsooghi, Toktam Khatibi, Afshin Bozorgpour, Reza Azad, Moein Heidari, Ilker Hacihaliloglu, and Dorit Merhof, “Msa2net: Multi-scale adaptive attention-guided network for medical image segmentation,” arXiv e-prints, pp. arXiv–2407, 2024

    2024

  6. [14]

    Haar wavelet downsampling: A simple but effective downsampling module for semantic segmentation,

    Guoping Xu, Wentao Liao, Xuan Zhang, Chang Li, Xinwei He, and Xinglong Wu, “Haar wavelet downsampling: A simple but effective downsampling module for semantic segmentation,” Pattern Recognition, vol. 143, pp. 109819, 2023

    2023

  7. [15]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    Alexey Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  8. [16]

    Hepatic echinococcosis: clinical and therapeutic aspects,

    Giuseppe Nunnari, Marilia R Pinzone, Salvatore Gruttadauria, Benedetto M Celesia, Giordano Madeddu, Giulia Malaguarnera, Piero Pavone, Alessandro Cappellani, and Bruno Cacopardo, “Hepatic echinococcosis: clinical and therapeutic aspects,” World journal of gastroenterology: WJG...

    2012

  9. [17]

    Ege-unet: an efficient group enhanced unet for skin lesion segmentation,

    Jiacheng Ruan, Mingye Xie, Jingsheng Gao, Ting Liu, and Yuzhuo Fu, “Ege-unet: an efficient group enhanced unet for skin lesion segmentation,” in International conference on medical image computing and computer-assisted intervention . Springer, 2023, pp. 481–490

    2023

  10. [18]

    Cbam: Convolutional block attention module,

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon, “Cbam: Convolutional block attention module,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 3–19

    2018

  11. [19]

    Springer, 2016, pp. 424–432

    2016

  12. [20]

    An efficient encoder-decoder architecture with top-down attention for speech separation,

    Kai Li, Runxuan Yang, and Xiaolin Hu, “An efficient encoder-decoder architecture with top-down attention for speech separation,” arXiv preprint arXiv:2209.15200, 2022

    2022 arXiv

  13. [21]

    Iianet: An intra- and inter-modality attention network for audio-visual speech separation,

    Kai Li, Runxuan Yang, Fuchun Sun, and Xiaolin Hu, “Iianet: An intra- and inter-modality attention network for audio-visual speech separation,” in Forty-first International Conference on Machine Learning , 2024

    2024

  14. [22]

    Decoupled weight decay regularization,

    I Loshchilov, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  15. [23]

    Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data,

    Foivos I Diakogiannis, Franc ¸ois Waldner, Peter Caccetta, and Chen Wu, “Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 162, pp. 94–114, 2020

    2020

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.