Pith. sign in

REVIEW 4 major objections 4 minor 42 references

Dual Attention Residual U-Net for Accurate Brain Ultrasound Segmentation in IVH Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that an attention-augmented residual U-Net reaches Dice 89.04% and IoU 81.84% for ventricle segmentation on neonatal brain ultrasound, outperforming six comparison models.

desk verdict Honest incremental architecture study whose SOTA claim is undercut by an ambiguous frame-level split and missing variance estimates; fixable in revision. read the letter →

arxiv 2505.17683 v2 pith:UZPMCANY submitted 2025-05-23 eess.IV cs.AIcs.CV

classification eess.IVcs.AIcs.CV
keywords intraventricularhemorrhagebrainultrasoundventriclesegmentationResidualU-NetCBAMsparseattentionhybridlayermedicalimage
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

Intraventricular hemorrhage in premature infants is usually tracked by measuring ventricle dilation in cranial ultrasound, so accurate ventricle segmentation is the clinical bottleneck this paper targets. The paper argues that standard U-Nets capture local detail but miss global context, while full self-attention spends computation on irrelevant regions, and proposes a residual U-Net with two complementary attention modules: CBAM for channel and spatial refinement, and a hybrid layer that combines sparse and dense attention. Its central claim is that this combination reaches Dice 89.04% and IoU 81.84% on the Brain US dataset, the best among seven compared methods, and that the right loss mixture is part of the recipe. If the claim holds, the architecture is a direct candidate for computer-aided IVH screening in neonatal units.

What carries the argument

The load-bearing mechanism is the serial CBAM-plus-HAL insertion between the encoder and decoder of a residual U-Net. CBAM applies channel attention followed by spatial attention with a $7\times7$ dilated kernel; HAL splits attention into two branches, a sparse branch $\mathrm{ReLU}^2(QK^T/\sqrt{d_K}+B)$ and a dense branch $\mathrm{softmax}(QK^T/\sqrt{d_K}+B)$, fusing them as $\sum_i \omega_i \mathrm{Att}_i V$. This gives a convolutional encoder–decoder both local feature refinement and global token correlations while suppressing irrelevant correlations, which is the argument for the reported gains.

What would settle it

Split the Brain US data by neonate rather than by frame—train on frames from 16 infants and test on the 4 remaining infants—and recompute Dice and IoU for the proposed model and for Medical Transformer. If the claimed 0.20-point Dice advantage disappears or reverses, the state-of-the-art claim does not hold.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a residual U-Net whose skip connections pass through CBAM and a Hybrid Attention Layer (HAL) produces the best ventricle segmentation results on the Brain US dataset among the seven methods tested. CBAM refines features along channel and spatial dimensions, while HAL computes self-attention over image patches with a sparse branch, $\mathrm{ReLU}^2(QK^T/\sqrt{d_K}+B)$, that zeroes out low-confidence query–key pairs, and a dense softmax branch that preserves global context; the two branches are fused with learned weights. The reported headline numbers are Dice 89.04% and IoU 81.84%, and the ablation attributes roughly one point of Dice to HAL on top of CBAM. The paper also shows that the Dice + BCE + Focal loss combination is necessary for the best score and that the same architecture runs on the Nuclei microscopy dataset, though with visually worse segmentations.

Load-bearing premise

The reported advantage rests on the evaluation being a fair, subject-independent test; the paper describes a 1300/329 frame split from only 20 neonates without stating that the same baby never contributes frames to both sets, and it also describes a different 70/10/20 split, so the scores may not generalize to new infants.

Editorial extensions

If this is right

  • If the reported numbers reproduce, the CBAM+HAL residual U-Net outperforms FCN, U-Net, U-Net++, Res-UNet, Axial Attention U-Net, and Medical Transformer on the Brain US test protocol.
  • The ablation implies the sparse branch is doing real work: CBAM alone yields Dice 87.97%, and adding HAL brings the score to 89.04%.
  • Because the three-term loss (Dice + BCE + Focal) beats the best two-term combination by about 1.7 Dice points, the loss choice is load-bearing for the headline result.
  • The same architecture with identical hyperparameters segments nuclei microscopy images, suggesting the attention insertion transfers beyond ultrasound even though those results are visually noisier.

Reading between the lines

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

  • A subject-independent split is the natural next test: the data come from 20 neonates, and the paper never states that training and test frames came from different babies; leave-one-subject-out evaluation would show whether the 0.20-point Dice margin over Medical Transformer generalizes.
  • The paper does not isolate the relative position bias $B$ or the branch weights $\omega_i$ in ablations, so it is an open question whether the gain comes from sparsity itself or from the extra parameters of the dual-branch design.
  • If sparse attention is the active ingredient, the same HAL block could be dropped into other segmentation backbones and compared at matched parameter count, a transfer test the paper does not run.
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 / 4 minor

Summary. The paper proposes a brain ultrasound segmentation method for IVH detection, combining a Residual U-Net backbone with two attention mechanisms: CBAM (channel and spatial attention) and a hybrid attention layer (HAL) composed of a sparse attention branch and a dense attention branch. Experiments on the Brain US dataset (1629 annotated frames from 20 neonates) report a Dice score of 89.04% and IoU of 81.84%, which the authors claim is state-of-the-art compared with FCN, U-Net, U-Net++, Res-UNet, Axial Attention U-Net, and Medical Transformer. An ablation study, loss-function combination experiments, failure-case analysis, and an additional Nuclei dataset experiment are also presented.

Significance. If the reported results are reliable, the paper would provide a modest but useful empirical contribution: it shows that combining CBAM with a sparse/dense dual-branch attention layer inside a Res-UNet can improve ventricle segmentation in neonatal brain ultrasound. The manuscript is honest in including failure cases and a supplementary dataset, and it releases code, which aids reproducibility. However, the central claim of state-of-the-art performance is load-bearing and rests on the evaluation protocol. The paper currently has an explicit inconsistency between the split described in Section 4.1 (1300/329) and the split described in Section 4.4 (70%/10%/20%), no patient-level split is stated despite only 20 neonates, no error bars or multiple-run statistics are reported, and the loss combination is selected from Table 1 without clarifying that selection was performed on a validation set. These issues directly affect whether the 0.20 Dice point advantage over Medical Transformer is a genuine model-quality effect or an artifact of the evaluation procedure.

major comments (4)
  1. [Section 4.1 and Section 4.4] The data split is described inconsistently. Section 4.1 states that 1300 of 1629 images are used for training and 329 for testing (approximately 79.8%/20.2%), while Section 4.4 states that the data are divided into 70% train, 10% test, and 20% validation. These two descriptions are incompatible: 70% of 1629 is about 1140, not 1300, and 10% is about 163, not 329. The paper must state the exact split actually used, including how many frames and how many patients are in each partition.
  2. [Section 4.1, Table 3] The paper does not state that the 1300 training frames and 329 test frames come from disjoint neonates. With only 20 infants total, a frame-level random split can place frames from the same neonate in both training and test sets; because consecutive ultrasound frames from one infant are highly correlated, this can inflate Dice scores and differentially favor a model that overfits frame-specific texture. This concern is load-bearing for the state-of-the-art claim, since the reported advantage over Medical Transformer is only 0.20 Dice (89.04 vs. 88.84). The authors should report a patient-disjoint split or a leave-two-patients-out cross-validation, and report per-patient statistics.
  3. [Section 4.5.2, Table 1] The final loss combination (Dice + BCE + Focal) is selected after comparing the combinations in Table 1, but the paper does not state that this selection was made on a validation set rather than on the test set. If the test set was used to select the loss, the reported Dice and IoU for the final model are optimistically biased. In addition, no error bars or multiple runs with different random seeds are reported anywhere, so the 0.20-point difference from Medical Transformer in Table 3 cannot be distinguished from run-to-run noise. The authors should report mean and standard deviation over at least three runs and, where possible, a paired significance test.
  4. [Section 4.6, Table 3] The comparison with baseline models is not fully specified. The paper does not state whether FCN, U-Net, U-Net++, Res-UNet, Axial Attention U-Net, and Medical Transformer were trained under the same protocol as the proposed method (same loss function, optimizer, learning rate, number of epochs, image size, and data split). The caption of Fig. 4 says that 'All the results are obtained by running their respective code,' which suggests the baselines may have used their original, possibly different, configurations. For a fair state-of-the-art comparison, each baseline should be retrained or fine-tuned under the same evaluation protocol, with the exact settings reported.
minor comments (4)
  1. [Section 4.2] There is a typo: 'Tow metrics are used' should be 'Two metrics are used.'
  2. [Table 2] In Table 2, the column header lists 'U-Net[17]' and 'Res-UNet[27]', but reference [17] is V-Net and reference [27] is the brain anatomy segmentation paper; the correct references for U-Net and Res-UNet appear to be [15] and [28], respectively. This should be corrected.
  3. [Section 3.3.2, Eq. (4)] Equation (4) writes Att2 = ReLU^2(QK^T/sqrt(d_K)+B), but the notation ReLU^2 is not defined; it could mean the square of the ReLU output or a two-time application. Please define the operation explicitly.
  4. [Section 4.4] The hyperparameter section states a 70%/10%/20% split, but no validation accuracy is reported anywhere in the paper; it would be helpful to report how the validation set was used (e.g., for early stopping or hyperparameter selection).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported Dice/IoU are genuine empirical outputs of trained networks, the architecture is assembled from external published components, and no predicted quantity is defined in terms of a fitted parameter.

full rationale

This paper is an empirical architecture comparison, not a derivation. The proposed CBAM+HAL Residual U-Net is built from externally published building blocks (CBAM [33], sparse attention [1], softmax dense attention), and its reported Dice 89.04 and IoU 81.84 are measured segmentation metrics obtained by training and evaluating the model on the Brain US dataset. No equation in Sections 3.1-3.4 defines the target metric as a function of a fitted input, and no ablation value is a renamed fit. The loss-combination selection in Section 4.5.2 compares several discrete loss combinations and adopts the best; this is model selection, not fitting a parameter to the reported metric. It could in principle have been performed on the test set, which would be selection bias affecting the reliability of the SOTA claim, but it is not circularity because the final Dice is still an output of a trained model. The dataset-split inconsistency between Section 4.1 (1300 training / 329 testing) and Section 4.4 (70/10/20 split) and the absence of a stated patient-disjoint split are evaluation-protocol concerns, again correctness risks rather than circularity. No self-citation is load-bearing, no uniqueness theorem is imported from the authors' prior work, and no known result is renamed as a new contribution. The central empirical claim therefore has independent content and does not reduce to its own inputs.

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

The central claim is an empirical performance number. It rests on the accuracy of the manual annotations, on an independent train/test split, and on the standard deep-learning training procedure. No new physical entities are introduced.

free parameters (3)
  • Loss weights alpha, beta, lambda = 1/3 each (average) in the final model; selected from 7 combinations
    Table 1 searches combinations and the best (BCE + Dice + Focal) is used for all experiments. If selection was made on the test split, the reported metrics are optimistic.
  • HAL branch weights omega_1, omega_2 = Learned during training
    Equation (5) introduces learned normalized weights modulating the sparse and dense branches. Their final values are not reported, and they are fit to data.
  • HAL patch size M = Not reported
    The HAL tokenization in Section 3.3.2 depends on M x M patches. M is never given, so the architecture is underspecified and performance may depend on it.
assumptions (3)
  • domain assumption Brain US manual annotations from [27,30] are accurate ground truth.
    The dataset is cited from prior work and used without quality control or inter-observer agreement checks.
  • domain assumption Training and test images are independent at the patient level.
    With only 20 neonates, random image split may put frames from the same baby in both training and testing. The paper never states a patient-level split.
  • domain assumption Standard deep learning training with Adam and the described losses produces the reported numbers.
    The paper relies on ordinary supervised training but gives no seed, batch size, or hardware details beyond a single GPU mention.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual Attention Residual U-Net for Accurate Brain Ultrasound Segmentation in IVH Detection." pith.science (2026). https://pith.science/paper/UZPMCANY

@misc{pith2026250517683,
  author       = {Pith},
  title        = {Pith review of: Dual Attention Residual U-Net for Accurate Brain Ultrasound Segmentation in IVH Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZPMCANY}},
  note         = {Machine review of arXiv:2505.17683}
}
read the original abstract

Intraventricular hemorrhage (IVH) is a severe neurological complication among premature infants, necessitating early and accurate detection from brain ultrasound (US) images to improve clinical outcomes. While recent deep learning methods offer promise for computer-aided diagnosis, challenges remain in capturing both local spatial details and global contextual dependencies critical for segmenting brain anatomies. In this work, we propose an enhanced Residual U-Net architecture incorporating two complementary attention mechanisms: the Convolutional Block Attention Module (CBAM) and a Sparse Attention Layer (SAL). The CBAM improves the model's ability to refine spatial and channel-wise features, while the SAL introduces a dual-branch design, sparse attention filters out low-confidence query-key pairs to suppress noise, and dense attention ensures comprehensive information propagation. Extensive experiments on the Brain US dataset demonstrate that our method achieves state-of-the-art segmentation performance, with a Dice score of 89.04% and IoU of 81.84% for ventricle region segmentation. These results highlight the effectiveness of integrating spatial refinement and attention sparsity for robust brain anatomy detection. Code is available at: https://github.com/DanYuan001/BrainImgSegment.

Figures

Figures reproduced from arXiv: 2505.17683 by the authors.

Figure 1
Figure 1. Block diagram of the proposed model used for medical image segmentation. An input image with dimensions 128 × 128 × 1 undergoes feature extraction through the encoder, and the decoder then performs upsampling on the encoded features to predict a binary mask of size 128 × 128 × 1. The in-between connections of the encoder and the decoder are accompanied by the addition of CBAM and HAL to enhance the performance. for … view at source ↗
Figure 2
Figure 2. An illustration of CBAM and HAL blocks. CBAM and HAL are applied to the input feature F. The [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Results of the ablation study indi￾cate the improvement in model performance with each experimental modification. GT and PM are ground truth and predicted mask, respec￾tively. Heatmaps demonstrate the improvement of the model’s performance in focusing on the region of interest.The SAL branch filters out the negative impact of query-key matching scores so that the heatmap looks monotonic in the uninformative re￾gions… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results of the proposed segmentation model on images of the Brain dataset. All the results are obtained by running their respective code [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Illustration of some of the failure cases of our model. The enclosed regions are the misclas￾sified segmented masks. GT and PM represent the Ground Truth and Predicted Mask, respectively [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The visual results show that the overall per [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 35 canonical work pages

  1. [2]

    A study on ReLu and Softmax in transformer

    K.Shenet al., “A study on ReLu and Softmax in transformer” 2023, arXiv:2302.06461

  2. [3]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A.Dosovitskiyet al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” inProc. ICLR,Vienna, Austria, 2021

  3. [4]

    Medical Transformer: Gated Axial-Attention for Medical Image Seg- mentation

    J.M.J Valanarasuet al., “Medical Transformer: Gated Axial-Attention for Medical Image Seg- mentation” 2021, arXiv:2104.07012

  4. [5]

    UNet 3+: A Full-Scale Con- nected UNet for Medical Image Segmentation,

    H.Huanget al., “UNet 3+: A Full-Scale Con- nected UNet for Medical Image Segmentation,” inProc. ICASSP,Barcelona, Spain, 2020, pp. 1055–1059

  5. [6]

    A comparative study of pre-trained convolutional neural networks for se- mantic segmentation of breast tumors in ultra- sound,

    G.F Wilfridoet al., “A comparative study of pre-trained convolutional neural networks for se- mantic segmentation of breast tumors in ultra- sound,”Computers in Biology and Medicine., vol. 126, no. 8, 2020

  6. [7]

    SegNet: A Deep Con- volutional Encoder-Decoder Architecture for Im- age Segmentation,

    V.Badrinarayananet al., “SegNet: A Deep Con- volutional Encoder-Decoder Architecture for Im- age Segmentation,”IEEE Transactions on Pat- tern Analysis and Machine Intelligence., vol. 39, no. 12, pp. 2481 - 2495, 2017

  7. [8]

    Attention U-Net: Learn- ing Where to Look for the Pancreas

    O.Oktayet al., “Attention U-Net: Learn- ing Where to Look for the Pancreas” 2018, arXiv:1804.03999

  8. [9]

    Road Extraction by Deep Residual U-Net,

    Z.Zhanget al., “Road Extraction by Deep Residual U-Net,”IEEE Geoscience and Remote Sensing Letters., vol. 15, no. 5, pp. 749-753, 2018

Show all 42 references
  1. [10]

    Spatial and Channel Attention Modulated Network for Medical Image Segmen- tation,

    W.Fanget al., “Spatial and Channel Attention Modulated Network for Medical Image Segmen- tation,” inProc. ACCV,2020

  2. [11]

    Squeeze-and-Excitation Net- works,

    J.Huet al., “Squeeze-and-Excitation Net- works,” inProc. CVPR,2018, pp. 7132–7141

  3. [12]

    GC-Net: Global Context Net- work for Medical Image Segmentation,

    J.Niet al., “GC-Net: Global Context Net- work for Medical Image Segmentation,”Com- puter Methods and Programs in Biomedicine., vol. 190, pp. 105-121, 2019

  4. [13]

    MDA-Unet: A Multi-Scale Di- lated Attention U-Net for Medical Image Seg- mentation,

    A.Ameret al., “MDA-Unet: A Multi-Scale Di- lated Attention U-Net for Medical Image Seg- mentation,”Applied Sciences., vol. 12, pp. 36-76, 2022

  5. [14]

    FCPFNet: Feature Complementa- tion Network with Pyramid Fusion for Semantic Segmentation,

    J.Leiet al., “FCPFNet: Feature Complementa- tion Network with Pyramid Fusion for Semantic Segmentation,”Neural Processing Letters., vol. 56, no. 2, 2024

  6. [15]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    O.Ronnebergeret al., “U-Net: Convolutional Networks for Biomedical Image Segmentation,” inProc. MICCAI,2015

  7. [16]

    UNet++: A Nested U-Net Ar- chitecture for Medical Image Segmentation,

    Z.Zhouet al., “UNet++: A Nested U-Net Ar- chitecture for Medical Image Segmentation,” in Proc. 4th DLMIA,, pp. 3–11, 2018

  8. [17]

    V-Net: Fully Convolutional Neural Networks for Volumetric Medical Im- age Segmentation,

    F.Milletariet al., “V-Net: Fully Convolutional Neural Networks for Volumetric Medical Im- age Segmentation,” inProc. 2016 fourth inter- national conference on 3D vision (3DV),2016, pp.565–571

  9. [18]

    Axial-DeepLab: Stand-Alone Axial-Attention for Panoptic Segmentation

    H.Wanget al., “Axial-DeepLab: Stand-Alone Axial-Attention for Panoptic Segmentation” 2020, arXiv:2003.07853

  10. [19]

    Deep Learning in Medical Ul- trasound Image Segmentation: a Review

    Z.Wanget al., “Deep Learning in Medical Ul- trasound Image Segmentation: a Review” 2020, arXiv:2002.07703

  11. [20]

    Convolutional neu- ral networks in medical image understanding: a survey,

    D.R. Sarvamangalaet al., “Convolutional neu- ral networks in medical image understanding: a survey,”Evolutionary Intelligence., vol.15, no. 1, pp. 1-22, 2022

  12. [21]

    MRL-Net: Multi-Scale Represen- tation Learning Network for COVID-19 Lung CT Image Segmentation,

    S.Liuet al., “MRL-Net: Multi-Scale Represen- tation Learning Network for COVID-19 Lung CT Image Segmentation,”Journal on Biomed- ical and Health Informatics (J-BHI)., vol.27, no.9, pp.4317-4328, 2023

  13. [22]

    IMAGPose: A Unified Conditional Framework for Pose-Guided Person Generation,

    F. Shen and J. Tang, “IMAGPose: A Unified Conditional Framework for Pose-Guided Person Generation,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 37, 2024, pp. 6246–6266

  14. [23]

    IMAGDressing-v1: Customizable Virtual Dressing,

    F. Shen, X. Jiang, X. He, H. Ye, C. Wang, X. Du, Z. Li, and J. Tang, “IMAGDressing-v1: Customizable Virtual Dressing,” inProc. AAAI Conf. Artif. Intell., vol. 39, no. 7, 2025, pp. 6795– 6804

  15. [24]

    A Deep Feature Selection Method for Tumor Classification in Breast Ul- trasound Images,

    P.Pramaniket al., “A Deep Feature Selection Method for Tumor Classification in Breast Ul- trasound Images,” inProc. International Con- ference on Mathematics and its Applications in new Computer Systems,2022

  16. [25]

    KiU-Net: Towards Accurate Segmentation of Biomedical Images using Over- complete Representations,

    J.M Joseet al., “KiU-Net: Towards Accurate Segmentation of Biomedical Images using Over- complete Representations,” inProc. MICCAI, 2020, pp.363–373. 9

  17. [26]

    DAU-Net: Dual attention- aided U-Net for segmenting tumor in breast ul- trasound images,

    Pramanik Pet al., “DAU-Net: Dual attention- aided U-Net for segmenting tumor in breast ul- trasound images,”PLOS ONE., 19(5), 2024

  18. [27]

    Learning to Segment Brain Anatomy from 2D Ultrasound with Less Data,

    J.M.J. Valanarasuet al., “Learning to Segment Brain Anatomy from 2D Ultrasound with Less Data,”IEEE Journal of Selected Topics in Sig- nal Processing., vol.14, no.6, pp.1221 - 1234, 2020

  19. [28]

    Weighted Res-UNet for High- Quality Retina Vessel Segmentation,

    X.Xiaoet al., “Weighted Res-UNet for High- Quality Retina Vessel Segmentation,” inProc. 9th ITME,2018, pp.327-331

  20. [29]

    Attention-Enriched Deep Learning Model for Breast Tumor Segmentation in Ultrasound Images,

    A. Vakanskiet al., “Attention-Enriched Deep Learning Model for Breast Tumor Segmentation in Ultrasound Images,”Ultrasound in medicine and biology., vol.46, no.10, pp.2819-2833, 2020

  21. [30]

    Automatic real-time CNN-based neonatal brain ventricles segmentation,

    P.Wanget al., “Automatic real-time CNN-based neonatal brain ventricles segmentation,” inProc. 15th ISBI,2018, pp.716-719

  22. [31]

    Born Too soon: the global epi- demiology of 15 million preterm births,

    K. Yuniset al., “Born Too soon: the global epi- demiology of 15 million preterm births,”Repro- ductive Health., vol.10, pp.1-14, 2013

  23. [32]

    AAU-net: An Adaptive At- tention U-net for Breast Lesions Segmentation in Ultrasound Images,

    G.Chenet al., “AAU-net: An Adaptive At- tention U-net for Breast Lesions Segmentation in Ultrasound Images,”IEEE Transactions on Medical Imaging., vol.42, no.5, pp.1289-1300, 2022

  24. [33]

    CBAM: Convolutional Block At- tention Module,

    S.Wooet al., “CBAM: Convolutional Block At- tention Module,”Proc. ECCV., pp.3-9, 2018

  25. [34]

    An improved resnet algorithm based on cbam,

    Y.Luoet al., “An improved resnet algorithm based on cbam,”Proc. ICCNEA., pp.121–125, 2021

  26. [35]

    Neonatal posthemorrhagic hydrocephalus from prematurity: Pathophysiol- ogy and current treatment concepts: A review,

    S.Robinsonet al., “Neonatal posthemorrhagic hydrocephalus from prematurity: Pathophysiol- ogy and current treatment concepts: A review,” Journal of neurosurgery. Pediatrics., vol.9, no.3, pp.242-258, 2012

  27. [36]

    Adapt or Perish: Adap- tive Sparse Transformer with Attentive Feature Refinement for Image Restoration,

    S. Zhouet al., “Adapt or Perish: Adap- tive Sparse Transformer with Attentive Feature Refinement for Image Restoration,” inProc. CVPR,, pp. 2952–2963, 2024

  28. [37]

    IMAGGarment-1: Fine-Grained Gar- ment Generation for Controllable Fashion De- sign,

    F. Shen, J. Yu, C. Wang, X. Jiang, X. Du, and J. Tang, “IMAGGarment-1: Fine-Grained Gar- ment Generation for Controllable Fashion De- sign,” 2025, arXiv:2504.13176

  29. [38]

    Long-Term TalkingFace Generation via Motion-Prior Conditional Diffu- sion Model,

    F. Shen, C. Wang, J. Gao, Q. Guo, J. Dang, J. Tang, and T.-S. Chua, “Long-Term TalkingFace Generation via Motion-Prior Conditional Diffu- sion Model,” 2025, arXiv:2502.09533

  30. [39]

    A Bag of Tricks for Fine-Grained Roof Extrac- tion,

    J. Hu, Z. Huang, F. Shen, D. He, and Q. Xian, “A Bag of Tricks for Fine-Grained Roof Extrac- tion,” inProc. IGARSS, IEEE Int. Geosci. Re- mote Sens. Symp., 2023

  31. [40]

    Enhancing Land- slide Segmentation with Guide Attention Mech- anism and Fast Fourier Transformer,

    K. Yan, F. Shen, and Z. Li, “Enhancing Land- slide Segmentation with Guide Attention Mech- anism and Fast Fourier Transformer,” inProc. Int. Conf. Intell. Comput., Springer, 2024, pp. 296–307

  32. [41]

    Advancing Pose-Guided Image Synthesis with Progressive Conditional Diffusion Models,

    F. Shen, H. Ye, J. Zhang, C. Wang, X. Han, and W. Yang, “Advancing Pose-Guided Image Synthesis with Progressive Conditional Diffusion Models,” 2023, arXiv:2310.06313

  33. [42]

    Classifying Breast Cancer Histopathological Images Using a Robust Artifi- cial Neural Network Architecture,

    X.Zhanget al., “Classifying Breast Cancer Histopathological Images Using a Robust Artifi- cial Neural Network Architecture,”Proc. IWB- BIO., pp.204–215, 2019

  34. [43]

    Multiscale U-Net with Spatial Po- sitional Attention for Retinal Vessel Segmenta- tion,

    C.Liuet al., “Multiscale U-Net with Spatial Po- sitional Attention for Retinal Vessel Segmenta- tion,”Journal of Healthcare Engineering., vol.1, 2022. 10

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.