Pith. sign in

REVIEW 3 major objections 7 minor 28 references

EAGLE: An Efficient Global Attention Lesion Segmentation Model for Hepatic Echinococcosis

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A state-space U-shaped network reaches 89.76% Dice on hepatic echinococcosis CT segmentation, beating the best prior model by 1.61%.

desk verdict Slice-level split and missing patient-level validation undermine the reported DSC in this otherwise competently assembled Mamba-U-Net for hepatic echinococcosis. read the letter →

arxiv 2506.20333 v1 pith:VIWHFRGD submitted 2025-06-25 eess.IV cs.CV

classification eess.IVcs.CV
keywords hepaticechinococcosismedicalimagesegmentationstate-spacemodelU-NetHaarwavelettransformCTimagingdeeplearning
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 tries to establish that a U-shaped network combining convolutional and state-space modeling can segment hepatic echinococcosis lesions in CT scans more accurately than existing CNN, Transformer, and state-space baselines. It reports a Dice Similarity Coefficient of 89.76% on a private 260-patient dataset, surpassing MSVM-UNet, the strongest comparison model, by 1.61%. The motivation is practical: HE is widespread in pastoral areas with scarce medical resources, and accurate automated segmentation could support diagnosis where radiologists are scarce. The architecture's two named components—CVSSB for local-global feature fusion and HWTB for lossless downsampling—are presented as the reasons for the gain.

What carries the argument

The load-bearing objects are two modules. The Convolutional Vision State Space Block (CVSSB) pairs an SS2D block, which scans the image in four directions with separate state-space equations to capture long-range dependencies, with a Depthwise-Aware FeedForward Network (DA-FFN) built from parallel depthwise convolutions with kernel sizes 1, 3, and 5 fused by channel attention. The Haar Wavelet Transform Block (HWTB) applies a Haar wavelet transform to split the feature map into low-frequency and horizontal, vertical, and diagonal detail sub-bands, concatenates them along channels, and restores the channel count with a 1x1 convolution, giving resolution reduction without discarding detail.

What would settle it

Retrain EAGLE on the same 260-patient dataset with a strictly patient-level split, holding out whole volumes for validation and test, and compare the resulting DSC; if the margin over MSVM-UNet shrinks or disappears, the claimed advantage is partly an artifact of slice-level overlap.

Watch

Extended reading notes

Core claim

EAGLE is a U-shaped network with a Progressive Visual State Space (PVSS) encoder and a Hybrid Visual State Space (HVSS) decoder. The central claim is that this design captures both local detail and global context at linear complexity, while the Haar Wavelet Transform Block avoids the spatial information loss of ordinary pooling by compressing spatial resolution into the channel dimension. On the collected HE dataset the model achieves 89.76% DSC, 88.95% precision, and 89.47% recall, outperforming all compared methods on all three metrics; the ablation study attributes the improvement to the combination of CVSSB, DA-FFN, and HWTB.

Load-bearing premise

The load-bearing premise is that the 80/10/10 split into training, validation, and test slices keeps every patient's slices entirely inside one split; if slices from the same patient appear in both training and test, the reported 89.76% Dice is inflated by the model having seen that patient's anatomy during training.

Editorial extensions

If this is right

  • If the reported result is correct, EAGLE is the most accurate model on this HE benchmark, and hybrid state-space U-Nets become a stronger choice than plain CNN or Transformer baselines for this disease.
  • The wavelet downsampling mechanism implies that fine boundary detail survives aggressive resolution reduction, which should help lesions with indistinct edges such as alveolar echinococcosis.
  • Linear-complexity global modeling makes the approach more plausible than Transformer-based alternatives for deployment in low-resource clinical settings, since it avoids quadratic attention cost.
  • The combination of Dice and BCE losses, together with the architecture's feature fusion, is claimed to handle small lesion regions better, which is the setting where the compared CNN models misclassify spinal tissue.

Reading between the lines

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

  • The reported 1.61% margin may be optimistic if the train/validation/test split was performed on 2D slices rather than per patient; slices from the same patient are highly correlated, and a patient-level split would be a stricter test of generalization.
  • Because the CVSSB and HWTB are not specific to echinococcosis, the same architecture should be testable on other lesion segmentation tasks with blurred boundaries, such as pancreatic or liver tumor segmentation, to see whether the gain transfers.
  • The dataset is single-center and private; a multi-center study with varied CT scanners and window settings would clarify whether the margin holds outside the collecting hospital.
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

3 major / 7 minor

Summary. The paper proposes EAGLE, a U-shaped segmentation network combining state-space models, convolutional blocks, and Haar wavelet downsampling for hepatic echinococcosis (HE) lesion segmentation from CT images. The authors introduce a Progressive Visual State Space (PVSS) encoder and a Hybrid Visual State Space (HVSS) decoder, and report a Dice Similarity Coefficient (DSC) of 89.76% on a private dataset of 260 patients, surpassing MSVM-UNet by 1.61%. Quantitative comparisons with eight baselines and ablation studies are presented.

Significance. If the reported performance were reliable, EAGLE would represent a meaningful advance in SSM-based medical image segmentation, with a clinically relevant application to a neglected disease. The architecture design—particularly the integration of SS2D with a multi-branch feed-forward network and wavelet-based downsampling—is interesting and could inform future work. The paper also provides a relatively large private dataset and a systematic ablation study. However, the significance is currently limited by the lack of patient-level splitting, missing statistical analysis, and an inconsistency between the ablation and main results, which together prevent the claimed state-of-the-art performance from being accepted.

major comments (3)
  1. [Section 3.1] The dataset split is described as follows: 'all CT volumes were converted into 2D slices and split into training, validation, and test sets with a ratio of 80%, 10%, and 10%, respectively.' As written, this is a slice-level random split performed after converting the volumes to slices, rather than a patient-level split. Because adjacent slices from the same patient volume are highly correlated and the same lesion typically appears across many consecutive slices, a model trained on slices from a given patient can memorize patient-specific intensity and texture patterns and score artificially high on held-out slices of that same patient. This patient leakage positively biases the reported DSC. The manuscript does not state that patient volumes were kept intact across splits, nor does it report any per-patient statistics, confidence intervals, or repeated runs. Since the central claim—that EAGLE surpasses MSVM-UNet by 1.61% DSC (Table 1)—rests entirely on this comparison, the result is not yet supported. The authors should repeat the experiments with a patient-level split (or provide evidence that no patient appears in more than one split) and report per-patient mean and standard deviation for DSC.
  2. [Section 3.4, Table 3 versus Tables 1 and 2] There is a direct inconsistency between the ablation study and the main results. Table 3 reports that adding all three proposed modules (CVSSB, DA-FFN, HWTB) to a Vanilla UNet baseline yields DSC = 88.32%, whereas Table 1 reports that EAGLE, which uses exactly these modules, achieves DSC = 89.76%, and Table 2 reports the same 89.76% for the configuration [2, 2, 4, 2]. The paper does not explain why the 'all modules' row in Table 3 differs from the full EAGLE results. If the ablation was performed on a different backbone or with additional differences (e.g., no CBAM, different patch embedding), this must be stated explicitly. As it stands, the ablation's claim that the three modules together contribute 3.1% DSC over the baseline is inconsistent with the final model's reported performance, undermining the paper's validation of its architectural contributions.
  3. [Section 3.3 and Section 3.2] The comparison with state-of-the-art methods in Table 1 lacks crucial details about baseline training. The implementation details in Section 3.2 specify the training protocol for EAGLE (AdamW, learning rate 0.001, batch size 4, up to 100 epochs, early stopping), but no equivalent information is provided for U-Net, UNet++, ResUNet, Attention UNet, TransUNet, Swin-UNet, HMT-UNet, or MSVM-UNet. Without identical, or at least reported, training settings for the baselines (optimizer, learning rate, epochs, loss function, data augmentation), the observed differences in DSC cannot be attributed to architectural superiority rather than to tuning asymmetries. The authors should either train all methods under the same protocol and state this clearly, or provide the exact training configurations used for each baseline.
minor comments (7)
  1. [Section 2.2, HWTB] The term 'lossless downsampling' is inaccurate because the 1x1 convolution after the Haar transform reduces the channel dimension from 4C to C, which discards information. The wavelet transform itself is invertible, but the subsequent channel reduction is lossy. Please revise the terminology or clarify the information-theoretic claim.
  2. [Section 3.1 heading] The subsection title contains a typo: 'Deatils' should be 'Details'.
  3. [Figure 1] The abbreviation 'BI' appears in the decoder stages but is not defined in the caption or text; presumably it stands for bilinear interpolation. Please define it explicitly.
  4. [Section 2.4] The loss weights λ1 and λ2 are said to be 'empirically set to 1 by default'; please state whether any hyperparameter search was performed for these weights or if they were fixed without tuning.
  5. [General] The paper does not provide a data availability statement or code release, which limits reproducibility. Even if the dataset cannot be shared for privacy reasons, releasing the code and trained model weights would facilitate verification.
  6. [Section 1, Related Work] The related-work section includes several references from the speech/audio domain (e.g., [8], [9], [10], [11], [12], [26]) that are not directly relevant to medical image segmentation; consider replacing them with more pertinent SSM-based segmentation papers.
  7. [Section 3.3] The evaluation metrics (Precision, Recall) are not fully defined; please specify whether they are computed per-slice, per-patient, or pooled over all test slices.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: EAGLE's performance claims rest on an empirical benchmark with standard ablation-based hyperparameter selection, not on a fitted parameter or a self-citation chain.

full rationale

The paper's central claim is an empirical segmentation accuracy (DSC 89.76%, Table 1) obtained by training EAGLE on a private CT dataset and evaluating it on a held-out test split. This is not a derivation from an input that already contains the output: the reported DSC is a measured outcome, not a fitted constant or a renamed input. Hyperparameters such as the CVSSB layer counts L=[2,2,4,2] are selected through ablations on the validation set (Tables 2 and 3), which is standard model selection rather than circular fitting; the final test result is reported for the chosen configuration. The manuscript contains several self-citations to co-author Kai Li's audio/speech processing work (refs. [8]-[12], [26]), and some of these citations appear misplaced (e.g., [11] is attached to MSVM-UNet in the introduction, and [12] is listed alongside [25] for Haar wavelet downsampling), but these citations do not carry the argument: the SS2D backbone is attributed to the independent VMamba work [14], and the wavelet downsampling to HWD [25]. No uniqueness theorem or prior result by the authors is invoked to force the architecture. The slice-level train/test split described in Section 3.1 ('all CT volumes were converted into 2D slices and split into training, validation, and test sets') raises a potential data-leakage and correctness concern rather than a circularity concern: leaking same-patient slices across splits would inflate the empirical numbers, but it would not make the reported DSC equivalent by construction to an input of the model. The paper is self-contained relative to its empirical benchmark; the absence of an external benchmark is a generalizability limitation, not circular reasoning. Therefore no circular step is identified.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The paper's central claim depends on four hand-chosen hyperparameters (network depth, loss weights, CT windowing, channel widths), on assumptions about the data split and representativeness, and on three newly named architectural modules that are validated only on a single private dataset.

free parameters (4)
  • CVSSB layer counts L = [2, 2, 4, 2]
    Chosen via ablation on the validation set (Table 2); the final configuration is reported as optimal.
  • Loss weights λ1 and λ2 = 1, 1
    Set 'empirically to 1 by default' in Eq. (5); no sensitivity analysis is provided.
  • CT window width/level = W=150, L=35
    Chosen with radiologists to map HU to [-40, 110]; this preprocessing affects all inputs and the results.
  • Channel widths per stage = [32, 64, 128, 256, 512]
    No justification given beyond convention; parameter count and capacity are not analyzed.
assumptions (5)
  • domain assumption The private dataset of 260 patients is representative of the target HE population in all deployment settings.
    Used in Section 3.1 to define the evaluation; no external validation or multi-center data is provided.
  • domain assumption Slices from the same patient are statistically independent, allowing a slice-level split to estimate generalization.
    This is assumed by the split described in Section 3.1; it is violated if the split is not patient-disjoint.
  • domain assumption The baseline models were trained with sufficiently tuned hyperparameters to represent their best performance.
    No baseline training protocols are given in Section 3.3; the comparison assumes fair and similarly optimized baselines.
  • domain assumption The Haar wavelet transform plus channel reduction preserves all clinically relevant information for segmentation.
    Invoked in Section 2.2 when the paper claims 'lossless downsampling.' The 1x1 convolution reduces channels, so the assumption is questionable.
  • standard math The SS2D state-space equations model long-range dependencies in 2D images effectively when combined with convolution.
    This is a design assumption inherited from VMamba [14]; the paper relies on it without new derivation.
invented entities (3)
  • Convolutional Vision State Space Block (CVSSB)
    purpose: Combine local convolutions with global state-space scanning in both encoder and decoder.
    Evaluated only via internal ablation (Table 3) on the private dataset; no external dataset or theoretical analysis supports its general superiority.
  • Haar Wavelet Transform Block (HWTB)
    purpose: Downsample feature maps while moving spatial information into channels.
    Claimed lossless but information is lost after channel reduction; only internal ablation evidence.
  • Depthwise-Aware Feed-Forward Network (DA-FFN)
    purpose: Enhance multi-scale features via parallel depthwise convolutions and channel attention.
    Similar to prior multi-branch FFN designs; only internal ablation evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EAGLE: An Efficient Global Attention Lesion Segmentation Model for Hepatic Echinococcosis." pith.science (2026). https://pith.science/paper/VIWHFRGD

@misc{pith2026250620333,
  author       = {Pith},
  title        = {Pith review of: EAGLE: An Efficient Global Attention Lesion Segmentation Model for Hepatic Echinococcosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VIWHFRGD}},
  note         = {Machine review of arXiv:2506.20333}
}
read the original abstract

Hepatic echinococcosis (HE) is a widespread parasitic disease in underdeveloped pastoral areas with limited medical resources. While CNN-based and Transformer-based models have been widely applied to medical image segmentation, CNNs lack global context modeling due to local receptive fields, and Transformers, though capable of capturing long-range dependencies, are computationally expensive. Recently, state space models (SSMs), such as Mamba, have gained attention for their ability to model long sequences with linear complexity. In this paper, we propose EAGLE, a U-shaped network composed of a Progressive Visual State Space (PVSS) encoder and a Hybrid Visual State Space (HVSS) decoder that work collaboratively to achieve efficient and accurate segmentation of hepatic echinococcosis (HE) lesions. The proposed Convolutional Vision State Space Block (CVSSB) module is designed to fuse local and global features, while the Haar Wavelet Transformation Block (HWTB) module compresses spatial information into the channel dimension to enable lossless downsampling. Due to the lack of publicly available HE datasets, we collected CT slices from 260 patients at a local hospital. Experimental results show that EAGLE achieves state-of-the-art performance with a Dice Similarity Coefficient (DSC) of 89.76%, surpassing MSVM-UNet by 1.61%.

Figures

Figures reproduced from arXiv: 2506.20333 by the authors.

Figure 1
Figure 1. The overall pipeline of EAGLE, which is composed of three main components: a Progressive Visual State Space (PVSS) Encoder, a Hybrid Visual State Space (HVSS) Decoder, and skip connections. F e i denotes the feature map output by the encoder block at the i-th stage, and F d i denotes the feature map output by the decoder block at the i-th stage. directional sequences, as shown in [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 2
Figure 2. The structure of the CVSSB module and its internal components: (a) the SS2D Block, (b) the proposed Convolutional Vision State Space Block (CVSSB), and (c) the proposed Depthwise-Aware FeedForward Network (DA-FFN). mechanism to emphasize informative feature channels while suppressing redun￾dant ones, ultimately improving the block’s capacity to represent diverse spatial patterns. The definition of DA-FFN can be form… view at source ↗
Figure 3
Figure 3. Illustration of 2D Selective Scan (SS2D). The input feature map is sequentially processed along four different scanning directions, with each directional sequence mod￾eled by a distinct discrete state-space equation to capture long-range dependencies. The directional outputs are then integrated and reshaped to recover the 2D feature map. Haar Wavelet Transform Block Conventional pooling-based downsampling operations… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The structure of the Haar Wavelet Transform Block. Specifically, the input feature map is denoted as Xm ∈ R C×H×W . It is first processed by a Haar wavelet transform denoted as T , which decomposes the fea￾ture map into four sub-bands: {Xm L , Xm HV , Xm HH, Xm HD} ∈ R…
Figure 5
Figure 5. Figure 5: Representative examples from the HE dataset. (a) CE; (b) AE. echinococcosis (CE) and 130 cases of alveolar echinococcosis (AE). Representa￾tive samples are shown in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison between our proposed EAGLE and other methods on CE lesion segmentation. True positives (TP) are shown in green, false negatives (FN) in hotpink, and false positives (FP) in blue. regions, it still struggles with lesions exhibiting blurred boundaries and high…
Figure 7
Figure 7. Figure 7: Comparison between our proposed EAGLE and other methods on AE lesion segmentation. True positives (TP) are shown in green, false negatives (FN) in hotpink, and false positives (FP) in blue. Proposed Modules To evaluate the individual contributions of the proposed CVSSB…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 15 canonical work pages

  1. [11]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    Li, K., Xie, F., Chen, H., Yuan, K., Hu, X.: An audio-visual speech separation model inspired by cortico-thalamo-cortical circuits. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

  2. [2]

    In: 2024 IEEE International Conference on Bioin- formatics and Biomedicine (BIBM)

    Chen, C., Yu, L., Min, S., Wang, S.: Msvm-unet: Multi-scale vision mamba unet for medical image segmentation. In: 2024 IEEE International Conference on Bioin- formatics and Biomedicine (BIBM). pp. 3111–3114. IEEE (2024)

  3. [1]

    In: European conference on computer vision

    Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin- unet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. pp. 205–218. Springer (2022)

  4. [3]

    arXiv preprint arXiv:2102.04306 (2021)

    Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)

  5. [4]

    Advances in parasitology 95, 315–493 (2017)

    Deplazes, P., Rinaldi, L., Rojas, C.A., Torgerson, P., Harandi, M.F., Romig, T., Antolova, D., Schurer, J., Lahmar, S., Cringoli, G., et al.: Global distribution of alveolar and cystic echinococcosis. Advances in parasitology 95, 315–493 (2017)

  6. [5]

    ISPRS Journal of Photogrammetry and Remote Sensing 162, 94–114 (2020)

    Diakogiannis, F.I., Waldner, F., Caccetta, P., Wu, C.: Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing 162, 94–114 (2020)

  7. [6]

    arXiv preprint arXiv:2312.00752 (2023)

    Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)

  8. [7]

    Proceedings of the IEEE 86(11), 2278–2324 (1998)

    LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278–2324 (1998)

Show all 28 references
  1. [8]

    arXiv preprint arXiv:2404.02063 (2024)

    Li, K., Chen, G., Hu, X.: Spmamba: State-space model is all you need in speech separation. arXiv preprint arXiv:2404.02063 (2024)

  2. [9]

    In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Li, K., Luo, Y.: Subnetwork-to-go: Elastic neural network with dynamic training and customizable inference. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6775–6779. IEEE (2024)

  3. [10]

    In: ICASSP 2025 (2025)

    Li, K., Luo, Y.: Apollo: Band-sequence modeling for high-quality audio restoration. In: ICASSP 2025 (2025)

  4. [12]

    arXiv preprint arXiv:2209.15200 (2022)

    Li, K., Yang, R., Hu, X.: An efficient encoder-decoder architecture with top-down attention for speech separation. arXiv preprint arXiv:2209.15200 (2022)

  5. [13]

    In: International Conference on Medical Image Computing and Computer- Assisted Intervention

    Liu, J., Yang, H., Zhou, H.Y., Xi, Y., Yu, L., Li, C., Liang, Y., Shi, G., Yu, Y., Zhang, S., et al.: Swin-umamba: Mamba-based unet with imagenet-based pretrain- ing. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 615–625. Spri...

  6. [14]

    Advances in neural information processing systems 37, 103031–103063 (2024)

    Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Jiao, J., Liu, Y.: Vmamba: Visual state space model. Advances in neural information processing systems 37, 103031–103063 (2024)

  7. [15]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)

  8. [16]

    arXiv preprint arXiv:1711.05101 (2017)

    Loshchilov, I.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

  9. [17]

    The lancet 362(9392), 1295–1304 (2003)

    McManus, D.P., Zhang, W., Li, J., Bartley, P.B.: Echinococcosis. The lancet 362(9392), 1295–1304 (2003)

  10. [18]

    arXiv preprint arXiv:1804.03999 (2018) 14 Chen et al

    Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al.: Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018) 14 Chen et al

  11. [19]

    In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

  12. [20]

    PLoS medicine 12(12), e1001920 (2015)

    Torgerson, P.R., Devleesschauwer, B., Praet, N., Speybroeck, N., Willingham, A.L., Kasuga, F., Rokni, M.B., Zhou, X.N., F` evre, E.M., Sripa, B., et al.: World health organization estimates of the global and regional disease burden of 11 foodborne parasitic diseases, 2010: a d...

  13. [21]

    PLoS neglected tropical diseases 4(6), e722 (2010)

    Torgerson, P.R., Keller, K., Magnotta, M., Ragland, N.: The global burden of alveolar echinococcosis. PLoS neglected tropical diseases 4(6), e722 (2010)

  14. [22]

    Advances in Neural Information Processing Systems (2017)

    Vaswani, A.: Attention is all you need. Advances in Neural Information Processing Systems (2017)

  15. [23]

    The Lancet Digital Health 5(11), e754–e762 (2023)

    Wang, Z., Bian, H., Li, J., Xu, J., Fan, H., Wu, X., Cao, Y., Guo, B., Xu, X., Wang, H., et al.: Detection and subtyping of hepatic echinococcosis from plain ct images with deep learning: a retrospective, multicentre study. The Lancet Digital Health 5(11), e754–e762 (2023)

  16. [24]

    In: Proceedings of the European conference on computer vision (ECCV)

    Woo, S., Park, J., Lee, J.Y., Kweon, I.S.: Cbam: Convolutional block attention module. In: Proceedings of the European conference on computer vision (ECCV). pp. 3–19 (2018)

  17. [25]

    Pattern Recognition 143, 109819 (2023)

    Xu, G., Liao, W., Zhang, X., Li, C., He, X., Wu, X.: Haar wavelet downsampling: A simple but effective downsampling module for semantic segmentation. Pattern Recognition 143, 109819 (2023)

  18. [26]

    In: ICLR 2025 (2025)

    Xu, M., Li, K., Chen, G., Hu, X.: Tiger: Time-frequency interleaved gain extraction and reconstruction for efficient speech separation. In: ICLR 2025 (2025)

  19. [27]

    arXiv preprint arXiv:2408.11289 (2024)

    Zhang, M., Chen, Z., Ge, Y., Tao, X.: Hmt-unet: a hybird mamba-transformer vi- sion unet for medical image segmentation. arXiv preprint arXiv:2408.11289 (2024)

  20. [28]

    Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: 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 In- ternational Workshop, DLMIA 2018, and 8th ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.