Pith. sign in

REVIEW 4 major objections 4 minor 15 references

XAG-Net: A Cross-Slice Attention and Skip Gating Network for 2.5D Femur MRI Segmentation

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

Pith's one-line read A 2.5D U-Net that applies pixel-wise softmax attention across three stacked MRI slices, with attention-gated skip connections, achieves the best femur segmentation scores among the evaluated 2D, 2.5D, and 3D baseline models on the authors'

desk verdict Real engineering, but the main attention module is likely misdescribed and the baseline comparison is underpowered. read the letter →

arxiv 2508.06258 v1 pith:7GDND6BU submitted 2025-08-08 cs.CV

classification cs.CV
keywords femurMRIsegmentation2.5Dconvolutionalneuralnetworkscross-sliceattentiongatingmedicalimageanisotropicvolumetricdataskipconnections
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

XAG-Net is a 2.5D U-Net for femur MRI segmentation whose thesis is that two lightweight additions—pixel-wise cross-slice attention and skip-connection attention gating—close the gap between cheap 2D processing and expensive 3D volumetric models. On the authors' full-scan femur MRI test set it reports DSC 0.9535, IoU 0.9160, and HD95 0.92 px, beating all compared 2D, 2.5D, and 3D U-Net baselines while using 56.3% fewer parameters and 72% fewer FLOPs than 3D U-Net. The paper argues that the CSA module, which softmax-normalizes attention over three stacked slices at every pixel, is the decisive component, with skip-connection CSA contributing more than input CSA or attention gating alone. The result matters because it offers a computationally practical route to inter-slice context for anisotropic scans like femur MRI.

What carries the argument

The load-bearing mechanism is the CSA module: a 1x1 convolution predicts a per-pixel attention score for each of the three input slices, softmax normalizes those scores along the slice dimension so the three weights sum to one at every spatial location, and the normalized map is multiplied element-wise with the input and added back residually. This same module is inserted at the input and, in parallel with standard attention-gating blocks, on each skip connection; the decoder fuses the two streams as $\operatorname{Concat}(U, \operatorname{CSA}(S_{\mathrm{enc}}), \operatorname{AG}(S_{\mathrm{enc}}, g))$. The CSA provides inter-slice context; the AG suppresses irrelevant background using the

What would settle it

Retrain the 3D U-Net baseline on the same TD01–TD08 training slices with per-model hyperparameter search and early stopping until its validation DSC reaches the 0.94 level the paper's own related work reports for a 3D CNN on proximal femur MRI, then re-run the full-scan evaluation on TD09–TD10. If that 3D U-Net reaches or exceeds XAG-Net's DSC 0.9535, the central superiority claim is falsified; if XAG-Net still leads, the claim survives.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the cross-slice attention operation $$\operatorname{CSA}(X)=X+(X\odot\operatorname{Softmax}(W(X)))$$ applied over three stacked axial slices at both the input and the skip connections, together with decoder-gated attention on skip features, extracts enough volumetric context to outperform fully 3D and prior 2.5D attention models on femur segmentation. In the full-scan comparison, XAG-Net reports the best scores on all three metrics (DSC 0.9535, IoU 0.9160, HD95 0.92 px). Its ablation study isolates the contribution: skip-connection CSA alone lifts DSC from 0.9436 to 0.9520, whereas attention gating alone lowers it to 0.9341, and only when C

Load-bearing premise

The claim that XAG-Net beats the baselines assumes every baseline was trained to comparable competence under the same recipe; the 3D U-Net's unusually low score (DSC 0.8307) is what creates the headline 12.3-point gain, and a properly tuned 3D U-Net might erode much of that margin.

Editorial extensions

If this is right

  • If the paper's results hold, femur MRI segmentation can reach a full-scan DSC above 0.95 and a boundary error near 1 pixel without a 3D volume model.
  • The ablation ordering implies future 2.5D designs should put cross-slice attention on skip connections before adding gating units; AG alone can hurt accuracy.
  • The proximal femur, the region with the most complex 3D shape, shows the largest relative gain (DSC 0.9197 vs. 0.9126 for 2.5D U-Net), indicating the CSA module's inter-slice modeling targets the hardest part of the anatomy.
  • Because XAG-Net remains cheaper than CSAM and 3D U-Net in parameters and FLOPs, the approach is practical for anisotropic volumetric datasets where full 3D convolutions are too expensive.

Reading between the lines

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

  • Editorial inference: the 3D U-Net baseline in Table 1 scores far below what the paper's own related-work section reports for a 3D CNN on proximal femur MRI (0.94 DSC); this suggests the double-digit headline gain may be driven by an undertrained baseline rather than by XAG-Net's architecture alone.
  • Editorial inference: since the CSA softmax is computed over only three slices yet outperforms CSAM's broader cross-slice modeling on rigid cortical bone, the decisive factor may be pixel-wise normalization rather than slice count; varying the stack size (e.g., 5 or 7 slices) would be a direct test.
  • Editorial inference: the method's logic—pixel-normalized attention over a thin slice stack plus gated skip fusion—should transfer to other anisotropic modalities such as CT or ultrasound volumes with small rigid structures, but the paper only demonstrates femur MRI.
  • Editorial inference: the authors' single-institution adolescent dataset leaves open whether the 0.9535 DSC holds across field strengths, protocols, and age groups; the paper itself names external validation as future work.
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 XAG-Net, a 2.5D U-Net for femur MRI segmentation that combines a pixel-wise cross-slice attention (CSA) module at the input and skip connections with attention gating (AG) blocks at skip connections. The model is trained on 3,761 slices from eight patients and tested on 1,041 slices from two held-out patients. The authors report that XAG-Net outperforms 2D U-Net, Attention U-Net, 2.5D U-Net, CSAM, and 3D U-Net on full-scan DSC (0.9535), IoU (0.9160), and HD95 (0.92 px), with a claimed 'up to 12.3 percentage point gain' over 3D U-Net, while using fewer parameters and FLOPs than 3D U-Net and CSAM. Ablation studies indicate that CSA modules in skip connections provide the largest performance improvement.

Significance. If the architectural claims and results hold, XAG-Net would be a useful contribution to efficient 2.5D medical image segmentation, particularly for anisotropic MRI where inter-slice context matters. The held-out patient evaluation (TD09/TD10) and the regional analysis across proximal, shaft, and distal femur are legitimate strengths, as is the reporting of parameter counts and FLOPs. However, the central architectural novelty is compromised by a dimension mismatch in the CSA formulation at skip connections, and the empirical comparison lacks statistical grounding and a clearly described validation procedure. These issues must be resolved before the claimed advantages can be accepted.

major comments (4)
  1. [Section 4.3, Table 1] The CSA definition in Eq. (1) assumes an input tensor X in R^{H x W x C} with C equal to the number of slices (3) and applies softmax over the slice (channel) dimension. The paper states this module is applied 'within each skip connection' (Sec. 3.1) and that F_concat = Concat(U, CSA(S_enc), AG(S_enc, g)) in Eq. (4). However, the encoder feature maps at skip connections have shape H x W x F, where F is the number of filters (64, 128, 256, 512), not the number of input slices. After the first 2D convolution, the three input slices are mixed, so no slice dimension remains. Applying Eq. (1) verbatim to a skip feature map means the softmax operates over feature channels, which is standard channel attention (as in SE blocks), not cross-slice attention. The paper never redefines C or reshapes the skip tensor to recover slice information. Because Table 3 attributes the largest performance gain
  2. [Section 3.5 vs. Section 4.1] The 3D U-Net baseline obtains DSC 0.8307, far lower than the 0.94 reported for a 3D CNN on proximal femur MRI in the paper's own reference [7]. The headline '12.3 percentage point gain' is computed against this baseline. The authors attribute the underperformance to 3D U-Net's complexity and overfitting (Discussion), but no convergence curves, hyperparameter search, or repeated-seed results are provided to demonstrate that the baseline was trained to competence under a fair recipe. If the 3D U-Net is undertrained or misconfigured, the central comparison and the abstract's headline gain are misleading. Please provide evidence of baseline training adequacy, such as training/validation curves, hyperparameter tuning, or a comparison with the 3D result in [7] on the same data.
  3. [Tables 1-3] Section 3.5 states that 'the model with the lowest validation loss is utilized for evaluation and comparison,' but the paper never describes a validation split. The dataset description in Section 4.1 only mentions training on TD01-TD08 (with some shaft-dominant slices withdrawn) and testing on TD09-TD10. If validation was performed on a subset of the training patients, this should be stated; if it was performed on the test patients TD09/TD10, then model selection leaks test information and all reported results are optimistic. This is a load-bearing experimental-design detail that affects every table in the paper.
  4. [Section 4.3] All results are reported as point estimates from a single training run, with no error bars, confidence intervals, or significance tests. Several differences between XAG-Net and the strongest baselines are extremely small: in the femoral shaft, XAG-Net (DSC 0.9624) is actually slightly worse than 2.5D U-Net (DSC 0.9626), and in the distal femur, XAG-Net (0.9235) is worse than Attention U-Net (0.9256). Without repeated runs or statistical testing, the claims that XAG-Net 'surpasses all baseline models' and that module ablations are 'critical' are not justified. Reporting variance (e.g., mean +/- std over at least three seeds) or a paired significance test is necessary.
minor comments (4)
  1. [Section 4.3] The text says 'All models are tested on axial femur MRI dataset consisting of 3,761 slices from patients 1 through 8 (TD01 through TD08)' but then says evaluation is on 1,041 slices from TD09/TD10. The first statement should be 'trained on', not 'tested on'.
  2. [Section 3.3] The output mask is described as '(256 x 1 x 1)'; this appears to be a typo for (256 x 256 x 1), since the segmentation is per-slice.
  3. [Section 3.4] The loss weights (0.9 and 0.1) are taken from reference [13] by the same group. It would be helpful to state whether these weights were re-tuned on this dataset or transferred unchanged, since this is a hyperparameter choice.
  4. [General] The phrase 'up to 12.3 percentage point gain' in the abstract is tied to the 3D U-Net baseline. Given the concerns about that baseline's training, the claim should be qualified or revised.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity; the central result is an empirical comparison on held-out patients. Same-group citations are used for hyperparameters and prior context but are not load-bearing.

full rationale

The paper's central claim is an empirical evaluation, not a derivation from first principles. XAG-Net's DSC, IoU, and HD95 are measured on held-out patients TD09 and TD10 and compared against baseline 2D, 2.5D, and 3D U-Nets (Table 1). The CSA and AG modules are defined by Eqs. (1)-(4) from their inputs, and the reported segmentation scores are not defined by those equations or by the loss function, so the result does not reduce to the model definition by construction. The only same-group citations are (i) the boundary/Dice loss weights 0.9 and 0.1 taken from ref [13] in Eq. (8) and discussed in Sec. 3.4, and (ii) a prior femur MRI 2D U-Net DSC from ref [8] cited in Related Work. Neither is load-bearing: the loss weights are a hyperparameter choice, and changing them would not make XAG-Net's superiority tautological; the prior femur result is contextual motivation. The skeptical concern that CSA at skip connections may actually softmax over feature channels rather than the three input slices is a correctness/description issue, not a circularity issue, because even if the implementation is misdescribed, the reported accuracy is not a logical consequence of the equations. The paper also explicitly acknowledges its single-dataset limitation, which is an external-validity concern rather than a circularity concern. Overall, there is no circular derivation; the score reflects only minor non-load-bearing self-citations.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a private dataset, hyperparameters inherited from prior work, assumed annotation quality, assumed baseline fairness, and a distribution-shifting training rebalance. No new physical entities, forces, or conserved quantities are introduced.

free parameters (4)
  • Loss weight for Dice term = 0.9
    Chosen from prior study [13] by the same group; fixed in Eq. 8, not tuned in this paper.
  • Loss weight for boundary term = 0.1
    Chosen from prior study [13] by the same group; fixed in Eq. 8, not tuned in this paper.
  • Number of input slices = 3
    Chosen by hand; all inputs are three-slice stacks (Sections 3.3 and 4.3). No sweep over slice count is reported.
  • Dice smoothing constant epsilon = 1
    Added to Eq. 5 for numerical stability; not varied across experiments.
assumptions (4)
  • domain assumption The pixel-level ground truth masks, drawn slice-by-slice by experienced musculoskeletal annotators, are accurate.
    Section 4.1 states annotations serve as voxel-level ground truth, but no quality control, inter-rater agreement, or correction process is reported.
  • domain assumption Patients TD09 and TD10 are representative of the broader femur MRI population.
    Section 4.1 uses only 10 subjects from a single institution; generalization beyond this cohort is assumed but not tested on external data.
  • ad hoc to paper Withdrawing shaft-dominant slices from the training corpus to equalize anatomy proportions does not harm full-scan test performance.
    Section 4.1 describes rebalancing the training set, but the test set is a full-scan (imbalanced) distribution, so the train and test distributions differ.
  • domain assumption All baseline models were trained fairly and comparably with the same hyperparameter recipe.
    Section 4.3 says all models were tested on the same dataset for 100 epochs, but no per-model tuning, repeated runs, or performance checks against published baselines are described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XAG-Net: A Cross-Slice Attention and Skip Gating Network for 2.5D Femur MRI Segmentation." pith.science (2026). https://pith.science/paper/7GDND6BU

@misc{pith2026250806258,
  author       = {Pith},
  title        = {Pith review of: XAG-Net: A Cross-Slice Attention and Skip Gating Network for 2.5D Femur MRI Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7GDND6BU}},
  note         = {Machine review of arXiv:2508.06258}
}
read the original abstract

Accurate segmentation of femur structures from Magnetic Resonance Imaging (MRI) is critical for orthopedic diagnosis and surgical planning but remains challenging due to the limitations of existing 2D and 3D deep learning-based segmentation approaches. In this study, we propose XAG-Net, a novel 2.5D U-Net-based architecture that incorporates pixel-wise cross-slice attention (CSA) and skip attention gating (AG) mechanisms to enhance inter-slice contextual modeling and intra-slice feature refinement. Unlike previous CSA-based models, XAG-Net applies pixel-wise softmax attention across adjacent slices at each spatial location for fine-grained inter-slice modeling. Extensive evaluations demonstrate that XAG-Net surpasses baseline 2D, 2.5D, and 3D U-Net models in femur segmentation accuracy while maintaining computational efficiency. Ablation studies further validate the critical role of the CSA and AG modules, establishing XAG-Net as a promising framework for efficient and accurate femur MRI segmentation.

Figures

Figures reproduced from arXiv: 2508.06258 by the authors.

Figure 1
Figure 1. An architectural overview of the proposed XAG-Net [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An overview of XAG-Net’s CSA module: The softmax operation ensures that the at [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of segmentation output of XAG-Net, 3D U-Net, and CSAM. The green [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 14 canonical work pages

  1. [7]

    Segmentation of the proximal femur from mr images using deep convolutional neural networks,

    C. M. Deniz, S. Xiang, R. S. Hallyburton, A. Welbeck, J. S. Babb, S. Honig, K. Cho, and G. Chang, “Segmentation of the proximal femur from mr images using deep convolutional neural networks,” Scientific Reports , vol. 8, no. 1, Nov. 2018. [Online]. Available: http://dx.doi.org/10.1038/s41598-018-34817-6

  2. [1]

    A two-stage deep learning network for automated femoral segmentation in bilateral lower limb ct scans,

    W. Xie, P. Chen, Z. Li, X. Wang, C. Wang, L. Zhang, W. Wu, J. Xiang, Y. Wang, and D. Zhong, “A two-stage deep learning network for automated femoral segmentation in bilateral lower limb ct scans,” Scientific Reports , vol. 15, no. 1, p. 9198, 2025. [Online]. Available: https://doi.org/10.1038/s41598-025-94180-1

  3. [2]

    Medical image segmentation automatic quality control: A multi-dimensional approach,

    J. Fournel, A. Bartoli, D. Bendahan, M. Guye, M. Bernard, E. Rauseo, M. Y. Khanji, S. E. Petersen, A. Jacquier, and B. Ghattas, “Medical image segmentation automatic quality control: A multi-dimensional approach,” Medical Image Analysis , vol. 74, p. 102213, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1361841521002589

  4. [3]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer International Publishing, 2015, pp. 234–241

  5. [4]

    Bridging 2d and 3d segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5d solutions,

    Y. Zhang, Q. Liao, L. Ding, and J. Zhang, “Bridging 2d and 3d segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5d solutions,” Computerized Medical Imaging and Graphics , vol. 99, p. 102088, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0895611122000611

  6. [5]

    3d u-net: Learning dense volumetric segmentation from sparse annotation,

    O. C ¸ i¸ cek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, “3d u-net: Learning dense volumetric segmentation from sparse annotation,” in Medical Image Computing and 10 Computer-Assisted Intervention – MICCAI 2016 , S. Ourselin, L. Joskowicz, M. R. Sabuncu, G. Unal, and W. Wells, Eds. Cham: Springer International Publishing, 2016, pp. 424–432

  7. [6]

    A flexible 2.5d medical image segmentation approach with in-slice and cross-slice attention,

    A. Kumar, H. Jiang, M. Imran, C. Valdes, G. Leon, D. Kang, P. Nataraj, Y. Zhou, M. D. Weiss, and W. Shao, “A flexible 2.5d medical image segmentation approach with in-slice and cross-slice attention,” Computers in Biology and Medicine , vol. 182, p. 109173, 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0010482524012587

  8. [8]

    Performance Analysis of Deep Learning Models for Femur Segmentation in MRI Scan

    M. Liu, Y. Chen, A. Tian, X. Wu, M. Shen, T. Gong, and J. Lee, “Performance analysis of deep learning models for femur segmentation in mri scan,” 2025. [Online]. Available: https://arxiv.org/abs/2504.04066

Show all 15 references
  1. [9]

    CSAM: A 2.5D Cross-Slice Attention Module for Anisotropic Volumetric Medical Image Segmentation,

    Y. H. AL, H. Zheng, K. Zhao, X. Du, K. Pang, Q. Miao, S. S. Raman, D. Terzopoulos, and K. Sung, “CSAM: A 2.5D Cross-Slice Attention Module for Anisotropic Volumetric Medical Image Segmentation,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Visio...

  2. [10]

    Attention u-net: Learning where to look for the pancreas,

    O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz, B. Glocker, and D. Rueckert, “Attention u-net: Learning where to look for the pancreas,” 2018. [Online]. Available: https://arxiv.org/abs/1804.03999

  3. [11]

    A 2.5d cancer segmentation for mri images based on u-net,

    K. Hu, C. Liu, X. Yu, J. Zhang, Y. He, and H. Zhu, “A 2.5d cancer segmentation for mri images based on u-net,” in 2018 5th International Conference on Information Science and Control Engineering (ICISCE) , 2018, pp. 6–10

  4. [12]

    Rsanet: Recurrent slice-wise attention network for mul- tiple sclerosis lesion segmentation,

    H. Zhang, J. Zhang, Q. Zhang, J. Kim, S. Zhang, S. A. Gauthier, P. Spincemaille, T. D. Nguyen, M. Sabuncu, and Y. Wang, “Rsanet: Recurrent slice-wise attention network for mul- tiple sclerosis lesion segmentation,” in Proceedings of the International Conference on Medical Imag...

  5. [13]

    Medvis suite: A framework for mri visualization and u-net-based bone segmentation with in-depth evaluation,

    Liu, Mengyuan, Zhang, Di, Chen, Yixiao, Gong, Tianchou, Kainz, Hans, Song, Seungmoon, and Lee, Jeongkyu, “Medvis suite: A framework for mri visualization and u-net-based bone segmentation with in-depth evaluation,” BIO Web Conf. , vol. 163, p. 04001, 2025. [Online]. Available:...

  6. [14]

    A generalized surface loss for reducing the hausdorff distance in medical imaging segmentation,

    A. Celaya, B. Riviere, and D. Fuentes, “A generalized surface loss for reducing the hausdorff distance in medical imaging segmentation,” 2024. [Online]. Available: https://arxiv.org/abs/2302.03868

  7. [15]

    Semantic segmentation of thigh muscle using 2.5d deep learning network trained with limited datasets,

    H. Haque, M. Hashimoto, N. Uetake, and M. Jinzaki, “Semantic segmentation of thigh muscle using 2.5d deep learning network trained with limited datasets,” 2019. [Online]. Available: https://arxiv.org/abs/1911.09249 11

Pith tools

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