Pith. sign in

REVIEW 4 major objections 7 minor 5 cited by

SegKAN: High-Resolution Medical Image Segmentation with Long-Distance Dependencies

T0 review · 4 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read SegKAN turns image patch positions into time steps to lift vessel segmentation Dice to 69.02%.

desk verdict A plausible but under-specified architecture paper: the reported Dice gain on hepatic vessel segmentation cannot be attributed to the proposed temporal-positional mechanism until the patch ordering is specified and the code is released. read the letter →

arxiv 2412.19990 v2 pith:E5BWRKKR submitted 2024-12-28 eess.IV cs.CV

Shengbo Tan , Rundong Xue , Shipeng Luo , Zeyu Zhang , Xinran Wang , Lei Zhang , Daji Ergu , Zhang Yi
show 2 more authors
Yang Zhao Ying Cai
This is my paper · ORCID
classification eess.IVcs.CV
keywords hepaticvesselsegmentationposition-temporalsequencenetworkFourierKANconvolution3Dmedicalimagelong-distancedependenciesgatedrecurrentCT
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

SegKAN is a medical image segmentation model designed for elongated structures such as hepatic vessels in CT scans, where fragmentation and noise break long-range continuity. The paper claims that converting the spatial relationships between 3D image patches into temporal sequence relationships, processed by a gated recurrent cell, recovers positional information that Vision Transformers miss. It also replaces the standard embedding with a Fourier-based KAN convolution that smooths local noise and stabilizes training. On the MSD Hepatic Vessel dataset, SegKAN reaches an average Dice of 69.02%, improving on the previous state of the art TransUNet by 1.78 percentage points. The significance would be a parameter-light way to model long-distance dependencies in high-resolution medical volumes without full self-attention.

What carries the argument

The two load-bearing modules are PTSN and FKAC. PTSN is a gated recurrent cell, defined by Eqs. 1-5, that mixes a sigmoid-gated copy of the input with a tanh-transformed version, $hid = h_{input} \odot G + (1-G) \odot H$, and updates a memory cell $c$; it operates over a sequence of 3D patches so that spatial position becomes temporal order. FKAC is an embedding layer in which each kernel element is the learnable Fourier activation $\phi_F(x) = \sum_{i=1}^{d} \sum_{k=1}^{g} (\cos(k x_i) a_{ik} + \sin(k x_i) b_{ik})$, wrapped by 1x1 compression and expansion convolutions with a residual connection. The Fourier basis functions replace B-splines, reducing the computational load of KAN convolutions while the convolutions smooth noise before the recurrent stage.

What would settle it

Train SegKAN on the same Hepatic Vessel data with the patch sequence order shuffled (for example, a random permutation instead of the paper's fixed order); if average Dice stays at 69.02%, the temporal encoding is not carrying the positional information, and the reported gain would not be due to the spatial-to-temporal conversion claimed.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a position-temporal sequence network (PTSN) combined with a Fourier-based KAN convolution (FKAC) outperforms established transformer-based segmenters on hepatic vessel and tumor segmentation. By feeding 3D patches into a gated recurrent cell in sequence, the model encodes each patch's place in the volume as a step in time, so the recurrence accumulates context along the ordering; the Fourier basis in the FKAC embedding replaces spline-based KAN activations with cosine and sine terms, cutting complexity while preserving nonlinear fitting. The reported results are an average Dice of 69.02%, with vessel Dice 65.90% and tumor Dice 72.14%, compared with TransUNet's 67.24%, 64.58%, and 69.89% respectively, and the ablation attributes the gain to the PTSN-FKAC combination.

Load-bearing premise

The claim depends on the assumption that a fixed sequential ordering of the 3D patches, fed into the gated recurrent cell, faithfully represents spatial position and long-range vessel structure; the paper does not specify the ordering rule or show that recurrence over that order recovers spatial locality better than attention.

Editorial extensions

If this is right

  • If the 69.02% average Dice holds, SegKAN becomes the reported state of the art on the MSD Hepatic Vessel benchmark, ahead of TransUNet by 1.78 points.
  • The 16-patch setting outperforms both 8 and 32 patches, implying that a moderate patch count balances context and detail; models tuned for elongated structures should search this split.
  • Because PTSN outperforms LSTM, SimpTime, and Mamba inside the same SegKAN backbone (69.02 versus 68.55, 66.83, and 67.10), the specific gated mechanism matters, not just recurrent processing.
  • The ablation attributes a 2.54-point gain to combining FKAC and PTSN over the baseline, so the two modules are complementary rather than redundant.
  • The approach is designed for elongated structures, so a direct corollary is that similar gains may appear on other thin or tubular anatomies such as retinal vessels or airways if the same modules are reused.

Reading between the lines

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

  • A likely testable extension is replacing the unspecified patch order with a learned or anatomically informed ordering (for example, tracing vessel branches), which could push Dice further if positional encoding is the active ingredient.
  • Because the Fourier FKAC replaces splines, the same module could transfer to other high-resolution 3D segmentation tasks where noise and gradient instability are bottlenecks, not just hepatic vessels.
  • If recurrence over patches is the mechanism, the approach might be combined with lightweight attention to see whether the two ways of modeling long-range context complement each other at lower compute than full transformers.
  • The paper leaves open how the patch sequence is ordered; testing whether a random order preserves the reported Dice would separate the recurrence's effect from the specific spatial encoding.
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

4 major / 7 minor

Summary. The manuscript proposes SegKAN, a network for hepatic vessel and tumor segmentation from 3D CT images. Two components are introduced: PTSN, a gated recurrent cell that converts spatial relationships among 3D patches into temporal sequence relationships (Eqs. 1-5), and FKAC, a Fourier-based KAN convolution embedding that smooths noise and mitigates gradient issues (Eqs. 6-8). Experiments on the MSD Hepatic Vessel dataset report an average Dice of 69.02%, outperforming TransUNet (67.24%) by 1.78 points, with ablations on the two modules, alternative temporal processors, and the number of patches. The paper claims the gains demonstrate that transforming spatial patch relations into temporal relationships improves long-distance vessel segmentation.

Significance. If the reported gains are reproducible, SegKAN would offer a modest but useful improvement over a strong Transformer baseline on a challenging elongated-structure segmentation task, and the PTSN idea of temporal encoding of patch order is a plausible alternative to positional embeddings. The paper is honest about its limited scope: a single dataset, no code release yet, and no statistical significance analysis. The ablation comparing PTSN with LSTM and Mamba is a useful sanity check. However, the significance is tempered by the absence of any description of the patch partitioning and ordering rule, which is the core mechanism being claimed; without that, the central attribution of the improvement to temporal-positional encoding cannot be assessed. The Fourier-based KAN embedding is a reasonable engineering variation of existing KAN convolution ideas, but its novelty over prior KAN and Fourier-network work is incremental.

major comments (4)
  1. [III-A and Figure 3] The PTSN mechanism is defined only through the gated equations (1)-(5); the manuscript never states how a 3D volume is partitioned into patches, what patch size or stride is used, or in what spatial order the patches are fed into the recurrent cell. Since the abstract and Section I attribute the performance gain to 'inputting the patch sequence into the model in a temporal sequence manner' to 'reinforce the positional relationship between patches,' the ordering rule is load-bearing. A fixed 1D raster order cannot faithfully represent full 3D adjacency, and different orderings (raster, space-filling curve, random) would yield different effective neighborhood structures. The reported Dice improvement cannot currently be attributed to the proposed temporal-positional mechanism, and the experiment is irreproducible without this specification.
  2. [IV-D and Table IV] The number of patches (8, 16, 32) is selected using the same evaluation metric (average Dice) and the same five-fold cross-validation protocol as the headline result in Table I. This is a selection-on-the-test-set procedure: the reported 69.02% is the best of three configuration choices, so the comparison against TransUNet is optimistically biased by the configuration search. The paper should either report all configurations against all baselines, or use a separate validation split for hyperparameter selection, or provide a corrected significance statement.
  3. [IV-D, Table II and the accompanying text] The ablation text states that 'the model's DSC improving by 2.54% compared to the baseline,' but Table II shows baseline average 66.57 and full model average 69.02, a difference of 2.45 points. The 2.54% figure appears if one takes the relative improvement (2.45/66.57 ≈ 3.68%) or possibly another arithmetic path, but as written it is inconsistent with the table. This inconsistency, combined with the absence of error bars, makes it impossible to judge whether the 2.45-point difference is meaningful relative to run-to-run variance; the paper should report standard deviations or per-fold results.
  4. [IV-C and Table I] The comparison table reports results for TransUNet (67.24% average) and other methods, but the text does not describe how these baselines were trained: the same data splits, preprocessing, loss function, and training schedule are not stated. Given that the reported gain over TransUNet is 1.78 points, and given the configuration-selection issue in Table IV, the reader cannot rule out that the improvement comes from training details rather than the proposed modules. The authors should provide the baseline training protocol and, ideally, code or a detailed reproducibility appendix.
minor comments (7)
  1. [Abstract and Section I] The abstract says 'the Dice score improved by 1.78%,' but the conventional reading of a percentage-point improvement from 67.24 to 69.02 is 1.78 percentage points, not 1.78 percent. Please rephrase for clarity.
  2. [Section III-B, Eq. (6)] Eq. (6) contains a formatting error: 'φ = wb · SiLU(x) = x / (1 + e^{-x}) + ws · Spline(x)' presents an equality between the whole expression and a term that is only part of the expression. This should be rewritten, for example as φ(x) = w_b·SiLU(x) + w_s·Spline(x), to be mathematically coherent.
  3. [Section III-B, Eq. (8)] The Fourier basis in Eq. (8) sums over cos(k x_i) and sin(k x_i) with coefficients a_{ik} and b_{ik}, but the dependency of the coefficients on the feature index i and frequency k is clear only implicitly. The notation should be defined explicitly, and the relationship to the compression/expansion convolution described in the text should be clarified.
  4. [Section IV-B] The training protocol says 'the model was trained for 1000 iterations,' which is unusually short for a 3D medical segmentation task with a 443-case dataset and five-fold cross-validation. If this is intentional, please explain; if it is a typo (e.g., 1000 epochs), correct it. The learning rate of 0.01 also seems high for a Transformer-based model; please state the optimizer, scheduler, and loss function.
  5. [Section IV-D, Table IV] Table IV reports inference time in 'T' units without defining what 'T' denotes (presumably throughput or time per case). Please define the unit and report the hardware setting for the timing, since the paper claims that increasing patch quantity increases inference time.
  6. [References and related work] The related-work section cites KAN, FAN, KAT, MULTKAN, Wav-KAN, and XNET, but does not cite or compare against recent KAN-based medical segmentation methods (e.g., U-KAN, KAN-UNet). Adding such comparisons would strengthen the positioning of FKAC and avoid the impression of incompleteness.
  7. [General] There are numerous grammatical and stylistic issues throughout (e.g., 'the Hepatic appears as tiny spots' should be 'vessels in the hepatic region appear as tiny spots'). A thorough language edit is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; SegKAN's reported Dice improvement is an empirical benchmark result, not a consequence derived from its own equations or self-citations.

full rationale

SegKAN is an empirical architecture paper. Its central claim (a 1.78% average Dice gain over TransUNet on the MSD Hepatic Vessel dataset) is a measured benchmark outcome, not a quantity derived from the model equations. Equations (1)-(8) define the PTSN gated recurrent cell and the FKAC Fourier-basis convolution; none of these definitions contains the target Dice value or is solved for it, so there is no self-definitional circularity. The self-citations in the reference list (e.g., the authors' own SegStitch and MedDet papers) appear only as related-work pointers and are not used to justify the effectiveness of PTSN or FKAC. The selection of 16 patches in Table IV is a hyperparameter choice made on the same evaluation metric, which is a limitation regarding statistical optimism and reproducibility, but it is not a reduction-by-construction: the reported 69.02% Dice is an empirical outcome of a trained model, not an algebraic consequence of the chosen patch count. The comparison is self-contained against external benchmarks such as nnU-Net and TransUNet. The unspecified patch-ordering rule in Section III-A is an under-specification that threatens reproducibility, but it does not make the claim circular. Overall, no load-bearing step reduces to its own inputs.

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

No new physical or external entities are introduced. The paper's central claim depends on two strong assumptions: fair baseline comparability and the validity of treating spatial patch order as a temporal sequence. The only manually selected hyperparameter with demonstrated effect is the patch quantity, chosen using the test metric. The trained network weights are fitted to data in the standard supervised-learning sense, which is not treated as circular here.

free parameters (2)
  • Number of 3D patches (patch quantity) = 16
    Chosen from ablation Table IV because 16 gives the best average Dice (69.02%), while 8 and 32 score lower; this selection is made on the evaluation metric rather than on a separate held-out validation set.
  • Fourier coefficient counts d and g in Eq. 8 = Not reported
    The capacity and behavior of FKAC depend on the number of cosine and sine terms, but the paper never states these values, leaving the exact architecture underspecified.
assumptions (3)
  • domain assumption The baseline results in Table I are directly comparable to SegKAN.
    Section IV-C compares SegKAN to prior published models without stating whether they were retrained under the same protocol, same fold split, same preprocessing, and same compute. The entire superiority claim rests on this comparability assumption.
  • domain assumption A fixed sequential ordering of 3D patches preserves spatial position information for a recurrent network.
    Section III-A assumes that feeding patches in a temporal sequence lets the network infer relative spatial positions, but no ordering rule or positional encoding is specified, and no direct analysis shows that the recurrence recovers spatial locality.
  • domain assumption Dice similarity coefficient on this dataset reflects clinically meaningful segmentation quality.
    The evaluation uses only Dice as the primary metric; no clinical validation or assessment of downstream usability is provided, yet the conclusion is framed as improving segmentation performance for medical use.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SegKAN: High-Resolution Medical Image Segmentation with Long-Distance Dependencies." pith.science (2026). https://pith.science/paper/E5BWRKKR

@misc{pith2026241219990,
  author       = {Pith},
  title        = {Pith review of: SegKAN: High-Resolution Medical Image Segmentation with Long-Distance Dependencies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5BWRKKR}},
  note         = {Machine review of arXiv:2412.19990}
}
read the original abstract

Hepatic vessels in computed tomography scans often suffer from image fragmentation and noise interference, making it difficult to maintain vessel integrity and posing significant challenges for vessel segmentation. To address this issue, we propose an innovative model: SegKAN. First, we improve the conventional embedding module by adopting a novel convolutional network structure for image embedding, which smooths out image noise and prevents issues such as gradient explosion in subsequent stages. Next, we transform the spatial relationships between Patch blocks into temporal relationships to solve the problem of capturing positional relationships between Patch blocks in traditional Vision Transformer models. We conducted experiments on a Hepatic vessel dataset, and compared to the existing state-of-the-art model, the Dice score improved by 1.78%. These results demonstrate that the proposed new structure effectively enhances the segmentation performance of high-resolution extended objects. Code will be available at https://github.com/goblin327/SegKAN

Figures

Figures reproduced from arXiv: 2412.19990 by the authors.

Figure 1
Figure 1. Comparison of Tumor and Vessel Segmentation Results Across [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The core algorithm framework of SegKAN, illustrates the 3D image slices and encoding of 3D patches, as well as the learning of spatial relationships [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. PTSN Structure replaces the feedforward neural network (FNN) component in LSTM with XNet, successfully capturing complex data features. This model surpasses traditional MLP and KAN in both speed and accuracy, demonstrating its great potential. III. METHODS A. paradigm—Position-Time Series Network (PTSN) The core advantage of Recurrent Neural Networks (RNNs) lies in their recursive structure, which uses the output of… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FKAC Structure B. Fourier-based KAN Convolution (FKAC) Based on previous analysis, the limitations of PTSN are pri￾marily attributed to the recursive structure of RNNs, especially the issue of gradient explosion during the backpropagation process. In contrast, convolut…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

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

  1. GAMED-Snake: Gradient-aware Adaptive Momentum Evolution Deep Snake Model for Multi-organ Segmentation

    cs.CV 2025-01 conditional novelty 5.0 of 10

    GAMED-Snake combines deep snake contour evolution with distance energy map priors, differential convolution, and cross-iteration attention to improve multi-organ segmentation Dice by roughly 2% over baselines.

  2. SSS: Semi-Supervised SAM-2 with Efficient Prompting for Medical Imaging Segmentation

    cs.CV 2025-06 reject novelty 4.0 of 10

    SSS applies SAM-2 with a Discriminative Feature Enhancement mechanism and a physical-constraint sliding-window prompt generator, reporting Dice scores of 53.15 on BHSD and 89.34 to 91.21 on ACDC.

  3. ProjectedEx: Enhancing Generation in Explainable AI for Prostate Cancer

    eess.IV 2025-01 conditional novelty 4.0 of 10

    A GAN with feature pyramid encoder and cross-channel mixing produces lower-FID counterfactual explanations for prostate MRI classification than the StylEx baseline.

  4. A Novel Convolutional-Free Method for 3D Medical Imaging Segmentation

    eess.IV 2025-02 reject novelty 3.0 of 10

    The paper proposes a convolution-free transformer pipeline and a thick-to-thin joint loss but reports no experiments and no performance numbers.

  5. MedConv: Convolutions Beat Transformers on Long-Tailed Bone Density Prediction

    cs.CV 2025-02 reject novelty 3.0 of 10

    A 3D ResNet with reweighted loss and logit adjustment reaches 65.38% accuracy on a private 389-patient CT dataset for three-class T-score prediction.

Reference graph

Works this paper leans on

54 extracted references · 27 canonical work pages · cited by 5 Pith papers

  1. [1]

    Bhsd: A 3d multi-class brain hemorrhage segmentation dataset,

    Biao Wu, Yutong Xie, Zeyu Zhang, Jinchao Ge, Kaspar Yaxley, Suzan Bahadir, Qi Wu, Yifan Liu, and Minh-Son To, “Bhsd: A 3d multi-class brain hemorrhage segmentation dataset,” in International Workshop on Machine Learning in Medical Imaging . Springer, 2023, pp. 147–156

    2023

  2. [2]

    Computational methods for liver vessel segmentation in medical imaging: A review,

    Marcin Ciecholewski and Michał Kassja ´nski, “Computational methods for liver vessel segmentation in medical imaging: A review,” Sensors, vol. 21, no. 6, pp. 2027, 2021

    work page 2027

  3. [3]

    Deep learning: new computational modelling techniques for genomics,

    G ¨okcen Eraslan, ˇZiga Avsec, Julien Gagneur, and Fabian J Theis, “Deep learning: new computational modelling techniques for genomics,” Nature Reviews Genetics , vol. 20, no. 7, pp. 389–403, 2019

    work page 2019

  4. [4]

    Deep learning modelling techniques: current progress, applications, advantages, and challenges,

    Shams Forruque Ahmed, Md Sakib Bin Alam, Maruf Hassan, Mahtabin Rodela Rozbu, Taoseef Ishtiak, Nazifa Rafa, M Mofijur, ABM Shawkat Ali, and Amir H Gandomi, “Deep learning modelling techniques: current progress, applications, advantages, and challenges,” Artificial Intelligence Review , vol. 56, no. 11, pp. 13521–13617, 2023

    work page 2023

  5. [5]

    A survey on deep learning: Algorithms, techniques, and applications,

    Samira Pouyanfar, Saad Sadiq, Yilin Yan, Haiman Tian, Yudong Tao, Maria Presa Reyes, Mei-Ling Shyu, Shu-Ching Chen, and Sundaraja S Iyengar, “A survey on deep learning: Algorithms, techniques, and applications,” ACM computing surveys (CSUR), vol. 51, no. 5, pp. 1–36, 2018

    work page 2018

  6. [6]

    Chaos challenge-combined (ct-mr) healthy abdominal organ segmentation,

    A Emre Kavur, N Sinem Gezer, Mustafa Barıs ¸, Sinem Aslan, Pierre- Henri Conze, Vladimir Groza, Duc Duy Pham, Soumick Chatterjee, Philipp Ernst, Savas ¸ ¨Ozkan, et al., “Chaos challenge-combined (ct-mr) healthy abdominal organ segmentation,” Medical Image Analysis , vol. 69, pp. 101950, 2021

    work page 2021

  7. [7]

    Abdomenct- 1k: Is abdominal organ segmentation a solved problem?,

    Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, et al., “Abdomenct- 1k: Is abdominal organ segmentation a solved problem?,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 10, pp. 6695–6714, 2021

    work page 2021

  8. [8]

    Cross- modal hybrid architectures for gastrointestinal tract image analysis: A systematic review and futuristic applications,

    Praneeth Nemani, Venkata Surya Sundar Vadali, Prathistith Raj Medi, Ashish Marisetty, Satyanarayana V ollala, and Santosh Kumar, “Cross- modal hybrid architectures for gastrointestinal tract image analysis: A systematic review and futuristic applications,” Image and Vision Computing, p. 105068, 2024

    work page 2024

Show all 54 references
  1. [9]

    On the segmentation of vascular geometries from medical images,

    AG Radaelli and J Peiro, “On the segmentation of vascular geometries from medical images,” International Journal for Numerical Methods in Biomedical Engineering, vol. 26, no. 1, pp. 3–34, 2010

    2010

  2. [10]

    Laplacian salience- gated feature pyramid network for accurate liver vessel segmentation,

    Zhan Gao, Qiuhao Zong, Yiqi Wang, Yan Yan, Yuqing Wang, Ning Zhu, Jin Zhang, Yunfu Wang, and Liang Zhao, “Laplacian salience- gated feature pyramid network for accurate liver vessel segmentation,” IEEE Transactions on Medical Imaging, vol. 42, no. 10, pp. 3059–3068, 2023

    2023

  3. [11]

    Automated liver tissues delineation techniques: A systematic survey on machine learning current trends and future orientations,

    Ayman Al-Kababji, Faycal Bensaali, Sarada Prasad Dakua, and Yassine Himeur, “Automated liver tissues delineation techniques: A systematic survey on machine learning current trends and future orientations,” Engineering Applications of Artificial Intelligence, vol. 117, pp. 105532, 2023

    2023

  4. [12]

    Review of liver segmentation and computer assisted detection/diagnosis methods in computed tomography,

    Mehrdad Moghbel, Syamsiah Mashohor, Rozi Mahmud, and M Iqbal Bin Saripan, “Review of liver segmentation and computer assisted detection/diagnosis methods in computed tomography,” Artificial Intelligence Review , vol. 50, pp. 497–537, 2018

    2018

  5. [13]

    A deep learning approach to diabetes diagnosis,

    Zeyu Zhang, Khandaker Asif Ahmed, Md Rakibul Hasan, Tom Gedeon, and Md Zakir Hossain, “A deep learning approach to diabetes diagnosis,” in Asian Conference on Intelligent Information and Database Systems . Springer, 2024, pp. 87–99

    2024

  6. [14]

    Jointvit: Modeling oxygen saturation levels with joint supervision on long-tailed octa,

    Zeyu Zhang, Xuyin Qi, Mingxi Chen, Guangxi Li, Ryan Pham, Ayub Qassim, Ella Berry, Zhibin Liao, Owen Siggs, Robert Mclaughlin, et al., “Jointvit: Modeling oxygen saturation levels with joint supervision on long-tailed octa,” in Annual Conference on Medical Image Understand- in...

    2024

  7. [15]

    Xlip: Cross-modal attention masked modelling for medical language-image pre-training,

    Biao Wu, Yutong Xie, Zeyu Zhang, Minh Hieu Phan, Qi Chen, Ling Chen, and Qi Wu, “Xlip: Cross-modal attention masked modelling for medical language-image pre-training,” arXiv preprint arXiv:2407.19546, 2024

    2024 arXiv

  8. [16]

    A landmark-based approach for instability predic- tion in distal radius fractures,

    Yang Zhao, Zhibin Liao, Yunxiang Liu, Koen Oude Nijhuis, Britt Barvelink, Jasper Prijs, Joost Colaris, Mathieu Wijffels, Max Reijman, Zeyu Zhang, et al., “A landmark-based approach for instability predic- tion in distal radius fractures,” in 2024 IEEE International Symposium o...

    2024

  9. [17]

    Can rotational thromboelastometry rapidly identify theragnostic targets in isolated traumatic brain injury?,

    Abhiram D Hiwase, Christopher D Ovenden, Lola M Kaukas, Mark Finnis, Zeyu Zhang, Stephanie O’Connor, Ngee Foo, Benjamin Reddi, Adam J Wells, and Daniel Y Ellis, “Can rotational thromboelastometry rapidly identify theragnostic targets in isolated traumatic brain injury?,” Emerg...

    2024

  10. [18]

    Meddet: Generative adversarial distillation for efficient cervical disc herniation detection,

    Zeyu Zhang, Nengmin Yi, Shengbo Tan, Ying Cai, Yi Yang, Lei Xu, Qingtai Li, Zhang Yi, Daji Ergu, and Yang Zhao, “Meddet: Generative adversarial distillation for efficient cervical disc herniation detection,” arXiv preprint arXiv:2409.00204 , 2024

    2024 arXiv

  11. [19]

    Msdet: Receptive field enhanced multiscale detection for tiny pulmonary nodule,

    Guohui Cai, Ying Cai, Zeyu Zhang, Daji Ergu, Yuanzhouhan Cao, Binbin Hu, Zhibin Liao, and Yang Zhao, “Msdet: Receptive field enhanced multiscale detection for tiny pulmonary nodule,” arXiv preprint arXiv:2409.14028, 2024

    2024 arXiv

  12. [20]

    Medical ai for early detection of lung cancer: A survey,

    Guohui Cai, Ying Cai, Zeyu Zhang, Yuanzhouhan Cao, Lin Wu, Daji Ergu, Zhinbin Liao, and Yang Zhao, “Medical ai for early detection of lung cancer: A survey,” arXiv preprint arXiv:2410.14769 , 2024

    2024 arXiv

  13. [21]

    Segreg: Segmenting oars by registering mr images and ct annotations,

    Zeyu Zhang, Xuyin Qi, Bowen Zhang, Biao Wu, Hien Le, Bora Jeong, Zhibin Liao, Yunxiang Liu, Johan Verjans, Minh-Son To, et al., “Segreg: Segmenting oars by registering mr images and ct annotations,” in 2024 IEEE International Symposium on Biomedical Imaging (ISBI) . IEEE, 2024...

    2024

  14. [22]

    Segstitch: Multidimensional transformer for robust and efficient medical imaging segmentation,

    Shengbo Tan, Zeyu Zhang, Ying Cai, Daji Ergu, Lin Wu, Binbin Hu, Pengzhang Yu, and Yang Zhao, “Segstitch: Multidimensional transformer for robust and efficient medical imaging segmentation,” arXiv preprint arXiv:2408.00496 , 2024

    2024 arXiv

  15. [23]

    Thin-thick adapter: Segmenting thin scans using thick annotations,

    Zeyu Zhang, Bowen Zhang, Abhiram Hiwase, Christen Barras, Feng Chen, Biao Wu, Adam James Wells, Daniel Y Ellis, Benjamin Reddi, Andrew William Burgan, Minh-Son To, Ian Reid, and Richard Hartley, “Thin-thick adapter: Segmenting thin scans using thick annotations,” OpenReview, 2023

    2023

  16. [24]

    Esa: Annotation-efficient active learning for semantic segmentation,

    Jinchao Ge, Zeyu Zhang, Minh Hieu Phan, Bowen Zhang, Akide Liu, and Yang Zhao, “Esa: Annotation-efficient active learning for semantic segmentation,” arXiv preprint arXiv:2408.13491 , 2024

    2024 arXiv

  17. [25]

    Robust hyperspectral image classifi- cation using a multi-scale transformer with long-short-distance spatial- spectral cross-attention,

    Danyang Peng, Haoran Feng, Jun Wu, Yi Wen, Tingting Han, Yuanyuan Li, Guangyu Yang, and Lei Qu, “Robust hyperspectral image classifi- cation using a multi-scale transformer with long-short-distance spatial- spectral cross-attention,” IEEE Transactions on Geoscience and Remote ...

    2024

  18. [26]

    Csffnet: Lightweight cross-scale feature fusion network for salient object detection in remote sensing images,

    Longbao Wang, Chong Long, Xin Li, Xiaodan Tang, Zhipeng Bai, and Hongmin Gao, “Csffnet: Lightweight cross-scale feature fusion network for salient object detection in remote sensing images,” IET Image Processing, vol. 18, no. 3, pp. 602–614, 2024

    2024

  19. [27]

    The gan spatiotemporal fusion model based on multi-scale convolution and attention mechanism for remote sensing images,

    Youping Xie, Jun Hu, Kang He, Li Cao, Kaijun Yang, and Luo Chen, “The gan spatiotemporal fusion model based on multi-scale convolution and attention mechanism for remote sensing images,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2024

    2024

  20. [28]

    A multi-scale cascaded cross-attention hierarchical network for change detection on bitemporal remote sensing images,

    Xiaofeng Zhang, Liejun Wang, and Shuli Cheng, “A multi-scale cascaded cross-attention hierarchical network for change detection on bitemporal remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    2024

  21. [29]

    Pgn: The rnn’s new successor is effective for long-range time series forecasting,

    Yuxin Jia, Youfang Lin, Jing Yu, Shuo Wang, Tianhao Liu, and Huaiyu Wan, “Pgn: The rnn’s new successor is effective for long-range time series forecasting,” arXiv preprint arXiv:2409.17703 , 2024

    2024 arXiv

  22. [30]

    Tltnet: A novel transscale cascade layered transformer network for enhanced retinal blood vessel segmentation,

    Chengwei Wu, Min Guo, Miao Ma, and Kaiguang Wang, “Tltnet: A novel transscale cascade layered transformer network for enhanced retinal blood vessel segmentation,” Computers in Biology and Medicine, vol. 178, pp. 108773, 2024

    2024

  23. [31]

    Global transformer and dual local attention network via deep-shallow hierarchical feature fusion for retinal vessel segmentation,

    Yang Li, Yue Zhang, Jing-Yu Liu, Kang Wang, Kai Zhang, Gen-Sheng Zhang, Xiao-Feng Liao, and Guang Yang, “Global transformer and dual local attention network via deep-shallow hierarchical feature fusion for retinal vessel segmentation,” IEEE Transactions on Cybernetics , vol. 5...

    2022

  24. [32]

    Towards large-scale small object detection: Survey and benchmarks,

    Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han, “Towards large-scale small object detection: Survey and benchmarks,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2023

    2023

  25. [33]

    Timesnet: Temporal 2d-variation modeling for general time series analysis,

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long, “Timesnet: Temporal 2d-variation modeling for general time series analysis,” arXiv preprint arXiv:2210.02186 , 2022

    2022 arXiv

  26. [34]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” arXiv preprint arXiv:2211.14730 , 2022

    2022 arXiv

  27. [35]

    Attention is all you need,

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

    2017

  28. [36]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in Proceedings of the AAAI conference on artificial intelligence , 2021, vol. 35, pp. 11106–11115

    2021

  29. [37]

    nnformer: V olumetric medical image segmentation via a 3d transformer,

    Hong-Yu Zhou, Jiansen Guo, Yinghao Zhang, Xiaoguang Han, Lequan Yu, Liansheng Wang, and Yizhou Yu, “nnformer: V olumetric medical image segmentation via a 3d transformer,” IEEE Transactions on Image Processing, 2023

    2023

  30. [38]

    Xiangyu Meng, Xudong Zhang, Gan Wang, Ying Zhang, Xin Shi, Huanhuan Dai, Zixuan Wang, and Xun Wang, “Exploiting full resolution feature context for liver tumor and vessel segmentation via integrate framework: Application to liver tumor and vessel 3d reconstruction under embedd...

    2021 arXiv

  31. [39]

    Kan: Kolmogorov-arnold networks,

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Solja ˇci´c, Thomas Y Hou, and Max Tegmark, “Kan: Kolmogorov-arnold networks,” arXiv preprint arXiv:2404.19756, 2024

    2024 arXiv

  32. [40]

    Kolmogorov-arnold convolutions: Design principles and empirical studies,

    Ivan Drokin, “Kolmogorov-arnold convolutions: Design principles and empirical studies,” arXiv preprint arXiv:2407.01092 , 2024

    2024 arXiv

  33. [41]

    Fan: Fourier analysis networks,

    Yihong Dong, Ge Li, Yongding Tao, Xue Jiang, Kechi Zhang, Jia Li, Jing Su, Jun Zhang, and Jingjing Xu, “Fan: Fourier analysis networks,” arXiv preprint arXiv:2410.02675 , 2024

    2024

  34. [42]

    Kolmogorov-arnold transformer,

    Xingyi Yang and Xinchao Wang, “Kolmogorov-arnold transformer,” arXiv preprint arXiv:2409.10594 , 2024

    2024 arXiv

  35. [43]

    Kan/multkan with physics-informed spline fitting (kan-pisf) for ordinary/partial differential equation discov- ery of nonlinear dynamic systems,

    Ashish Pal and Satish Nagarajaiah, “Kan/multkan with physics-informed spline fitting (kan-pisf) for ordinary/partial differential equation discov- ery of nonlinear dynamic systems,” arXiv preprint arXiv:2411.11801 , 2024

    2024 arXiv

  36. [44]

    Wav-kan: Wavelet kolmogorov- arnold networks,

    Zavareh Bozorgasl and Hao Chen, “Wav-kan: Wavelet kolmogorov- arnold networks,” arXiv preprint arXiv:2405.12832 , 2024

    2024 arXiv

  37. [45]

    Model comparisons: Xnet outperforms kan,

    Xin Li, Zhihong Jeff Xia, and Xiaotao Zheng, “Model comparisons: Xnet outperforms kan,” arXiv preprint arXiv:2410.02033 , 2024

    2024 arXiv

  38. [46]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature methods, vol. 18, no. 2, pp. 203–211, 2021

    2021

  39. [47]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,

    Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu, “Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,” in International MICCAI brainlesion workshop. Springer, 2021, pp. 272–284

    2021

  40. [48]

    A robust volumetric transformer for accurate 3d tumor segmentation,

    Himashi Peiris, Munawar Hayat, Zhaolin Chen, Gary Egan, and Mehrtash Harandi, “A robust volumetric transformer for accurate 3d tumor segmentation,” in International conference on medical image computing and computer-assisted intervention. Springer, 2022, pp. 162– 172

    2022

  41. [49]

    nnformer: Interleaved transformer for volumetric segmentation,

    Hong-Yu Zhou, Jiansen Guo, Yinghao Zhang, Lequan Yu, Liansheng Wang, and Yizhou Yu, “nnformer: Interleaved transformer for volumetric segmentation,” arXiv preprint arXiv:2109.03201 , 2021

    2021 arXiv

  42. [50]

    Deformux-net: Exploring a 3d foundation backbone for medical image segmentation with depthwise deformable convolution,

    Ho Hin Lee, Quan Liu, Qi Yang, Xin Yu, Shunxing Bao, Yuankai Huo, and Bennett A Landman, “Deformux-net: Exploring a 3d foundation backbone for medical image segmentation with depthwise deformable convolution,” arXiv preprint arXiv:2310.00199 , 2023

    arXiv

  43. [51]

    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

    2021 arXiv

  44. [52]

    Lstm: A search space odyssey,

    Klaus Greff, Rupesh K Srivastava, Jan Koutn ´ık, Bas R Steunebrink, and J ¨urgen Schmidhuber, “Lstm: A search space odyssey,” IEEE transactions on neural networks and learning systems , vol. 28, no. 10, pp. 2222–2232, 2016

    2016

  45. [53]

    Timemachine: A time se- ries is worth 4 mambas for long-term forecasting,

    Md Atik Ahamed and Qiang Cheng, “Timemachine: A time se- ries is worth 4 mambas for long-term forecasting,” arXiv preprint arXiv:2403.09898, 2024

    2024 arXiv

  46. [54]

    Mamba: Linear-time sequence modeling with selective state spaces,

    Albert Gu and Tri Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.