Pith. sign in

REVIEW 3 major objections 4 minor 28 references

A 3D transformer that routes whole MRI windows to specialized attention heads predicts perineural invasion better than standard CNNs and Swin.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 07:49 UTC pith:3GVCOX27

load-bearing objection Clean window-level MoH adaptation for 3D PNI MRI; solid internal ablations, but the 0.044 AUC edge sits on 168 single-center crops with no error bars or external check. the 3 major comments →

arxiv 2607.10988 v1 pith:3GVCOX27 submitted 2026-07-13 cs.CV cs.AI

MMA-Former: Multi-Window Mixture-of-Head Attention Transformer for Adaptive PNI Prediction in 3D MRI

Youngung Han , Induk Um , Kyeonghun Kim , Junga Kim , Hyunsu Go , Jaewon Jung , Woo Kyoung Jeong , Won Jae Lee
show 4 more authors
Pa Hong Ken Ying-Kai Liao Hyuk-Jae Lee Nam-Joon Kim
This is my paper
classification cs.CV cs.AI
keywords Vision TransformerMixture-of-Head attentionAdaptive Feature ExtractionWindow-level RoutingPerineural invasion3D MRICholangiocarcinoma
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Perineural invasion (PNI) is a high-stakes prognostic marker in cholangiocarcinoma, yet it is hard to see on preoperative MRI. This paper claims that a hierarchical 3D transformer can make that call more accurately by letting different spatial windows use different attention heads. The architecture runs coarse and fine windows in parallel (the Coarse-Fine Transformer) and replaces ordinary multi-head attention with Window-Specific Mixture-of-Head attention: each 3D window is summarized, then routed to a shared set of always-on heads plus a top-k subset of specialized heads. On 168 tumor-centered T1 MRI volumes the model reaches a mean AUC of 0.752, beating the best 3D CNN (0.708) and 3D Swin (0.681). Ablations show both the parallel windows and the window-level routing contribute to the gain. The practical claim is that spatially adaptive head selection, without extra parameters, improves non-invasive PNI prediction from routine MRI.

Core claim

On a single-institution set of 168 cropped T1-weighted MRI volumes, the MMA-Former architecture—parallel coarse-fine windows plus Window-Specific Mixture-of-Head attention and hierarchical Xpace fusion—achieves a mean 5-fold AUC of 0.752 for binary PNI prediction, outperforming the strongest 3D CNN and Transformer baselines tested, with the performance lift attributable to adaptive, window-level head routing rather than uniform multi-head attention.

What carries the argument

Window-Specific Mixture-of-Head attention (WS-MoH): each 3D window is reduced to one representative vector that routes the entire window to always-active shared heads plus a top-k subset of specialized heads, enabling spatially adaptive feature extraction without adding parameters.

Load-bearing premise

That a single-hospital collection of 168 tumor-centered MRI crops, labeled after surgery, is representative enough for the reported accuracy ranking to hold on new scanners and institutions.

What would settle it

An external multi-center cohort of cholangiocarcinoma MRI cases, processed with the same tumor-centered crop protocol, on which MMA-Former no longer outperforms the same 3D ResNet and Swin baselines under identical 5-fold evaluation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces MMA-Former, a hierarchical 3D Vision Transformer for binary perineural invasion (PNI) prediction from tumor-centered T1-weighted MRI volumes of cholangiocarcinoma. It combines a Coarse-Fine Transformer (CFT) structure that processes parallel fine (e.g., 3×3×3) and coarse (e.g., 6×6×6) windows with a Window-Specific Mixture-of-Head (WS-MoH) attention mechanism that average-pools each window to produce a representative feature, then routes the entire window to a mixture of always-on shared heads and Top-K routed heads (Eqs. 6–8). An Xpace block performs hierarchical cross-attention fusion across stages. On a single-institution retrospective cohort of 168 cropped 96×96×48 volumes (67 positive / 101 negative), stratified 5-fold CV yields mean AUC 0.752, exceeding 3D ResNet (0.708) and 3D Swin Transformer (0.681). Component ablations (Table 2) attribute gains to CFT parallelism, WS-MoH routing (S=2 shared heads, Top-K 75 %), and Xpace, with Grad-CAM localizing attention to the tumor–peritumoral interface.

Significance. If the reported ranking and ablation gains hold under external validation, the work supplies a concrete, parameter-efficient adaptation of Mixture-of-Head attention to 3D medical volumes that enables spatially adaptive specialization without token-level routing cost. Non-invasive PNI prediction remains clinically relevant for surgical planning in cholangiocarcinoma; the architecture’s multi-scale + window-level routing design, systematic ablations, load-balance loss, and qualitative Grad-CAM evidence constitute a useful methodological contribution to hierarchical 3D transformers even if absolute AUC remains modest. The explicit localization necessity (Table 2A) and shared/routed head analysis are particularly clear strengths.

major comments (3)
  1. [Table 1, §3.2] Table 1 and §3.2 report only mean 5-fold AUCs (0.752 vs 0.708 vs 0.681) with no standard deviations, confidence intervals, or paired statistical tests across folds. With N=168 and class imbalance, fold variance can readily exceed the 0.044–0.071 absolute gaps; without uncertainty quantification the central superiority claim for MMA-Former (and therefore for WS-MoH and CFT) is not statistically supported.
  2. [§4, Table 2A] The entire ranking of architectures and components rests on a single-institution retrospective set of 168 tumor-centered crops (explicitly flagged in §4). Table 2A shows that removing the crop drops AUC by 0.047, confirming heavy dependence on localization and scanner/protocol homogeneity. No multi-center or multi-scanner hold-out is provided, so transportability of the WS-MoH / CFT gains remains untested.
  3. [§2.3.2, Eqs. (6)–(8)] Eqs. (6)–(8) define the router exclusively from an average-pooled window representative X_rep. No ablation compares alternative aggregators (max-pool, attention-weighted, CLS token, etc.). Given that the paper’s core claim is spatially adaptive head selection, the adequacy of simple average pooling is a load-bearing modeling assumption that should be stress-tested.
minor comments (4)
  1. [Fig. 2, §2.3] Several typographical and notational inconsistencies appear: “hierachical” (Fig. 2 caption), “Xpace” vs “Cross-Spatial Attention”, “SWS-MoH” introduced without expansion in the figure legend, and inconsistent subscripts (P^c / P^f vs P_c / P_f).
  2. [§2.3, §3.1] Implementation details needed for reproducibility are missing: total number of attention heads h, exact channel dimensions after each patch-merging stage, and whether the same Top-K / S settings are used in every MMA block.
  3. [Fig. 1, §3.4] Fig. 1 Grad-CAM examples are informative but lack quantitative localization metrics (e.g., overlap with annotated tumor or nerve-adjacent regions) that would strengthen the qualitative claim.
  4. [§3.1, Eq. (10)] The load-balance loss weight β=0.01 and positive class weight 1.5 are stated without sensitivity analysis; a brief note on stability under modest changes would help.

Circularity Check

0 steps flagged

No circularity: purely empirical architecture proposal whose AUC claims are measured on held-out 5-fold CV, not algebraically forced by fitted constants or self-referential definitions.

full rationale

MMA-Former is a standard supervised 3D vision-transformer paper. The central claims (AUC 0.752 vs. ResNet 0.708 / Swin 0.681, plus component ablations) are obtained by training end-to-end models and evaluating mean 5-fold cross-validation performance on a fixed retrospective cohort of 168 cropped volumes; the numbers are therefore empirical measurements, not quantities derived by construction from the same inputs. Hyper-parameters (S=2 shared heads, Top-K 75 %, eta=0.01 load-balance weight, window sizes, crop size) are selected by internal ablation (Table 2) and then frozen for the final comparison (Table 1); this is ordinary model selection, not a fitted-input-called-prediction loop that forces the reported ranking. The WS-MoH equations (Eqs. 5–8) simply define a routing mechanism that is subsequently trained; they do not embed the target AUC or any uniqueness theorem. Citations (MoH [16], Swin [13], etc.) are to external prior work with non-overlapping authors and are used only for architectural building blocks, not as load-bearing uniqueness results. No self-definitional identities, self-citation chains, or renamed known results appear. The paper is therefore free of the circularity patterns enumerated in the analyzer guidelines.

Axiom & Free-Parameter Ledger

6 free parameters · 3 axioms · 3 invented entities

The claim rests on standard deep-learning practice plus several hand-chosen architectural and training knobs tuned on the same small cohort. No new physical entities are postulated; the invented components are architectural modules whose value is measured only by internal AUC.

free parameters (6)
  • Top-K routing ratio = 75%
    Fraction of routed heads activated per window; default 75 % selected by ablation Table 2E.
  • Number of shared heads S = 2
    Always-active heads; default S=2 chosen by ablation Table 2D.
  • Load-balance loss weight β = 0.01
    Multiplier on L_LB in total loss; set to 0.01.
  • Positive class weight in BCE = 1.5
    Addresses 67:101 imbalance; set to 1.5.
  • Learning rate / batch size = 8e-5 / 4
    AdamW 8e-5, batch 4; standard but still free choices affecting the reported AUC.
  • Crop size and window sizes = 96×96×48; 3³ / 6³
    96×96×48 tumor crop; fine 3×3×3 and coarse 6×6×6 windows chosen by design.
axioms (3)
  • domain assumption PNI status is sufficiently encoded in local T1 hepatobiliary-phase MRI intensity patterns around the tumor that a 3D network can recover it above chance.
    Foundational premise of the entire prediction task; never independently proven, only assumed from clinical literature.
  • ad hoc to paper Average-pooled window features are adequate router inputs for selecting specialized attention heads.
    Core design choice of WS-MoH (Eq. 6–8); alternative aggregators not compared.
  • domain assumption Standard multi-head attention and MoE load-balancing theory transfer to 3D medical volumes.
    Inherited from Vaswani, Shazeer, Jin et al. without re-derivation.
invented entities (3)
  • WS-MoH (Window-Specific Mixture-of-Head attention) no independent evidence
    purpose: Route entire 3D windows to shared + top-k specialized heads for spatially adaptive feature extraction.
    Primary architectural novelty; independent evidence limited to internal ablations on this dataset.
  • CFT (Coarse-Fine Transformer) structure no independent evidence
    purpose: Parallel multi-scale window pathways inside consecutive MMA blocks.
    Enables simultaneous fine and coarse context; value shown only by ablation Table 2C.
  • Xpace block no independent evidence
    purpose: Hierarchical cross-stage fusion via bidirectional multi-head cross-attention and spatial attention.
    Auxiliary fusion module; ablation shows modest contribution.

pith-pipeline@v1.1.0-grok45 · 13003 in / 3031 out tokens · 26869 ms · 2026-07-14T07:49:42.494428+00:00 · methodology

0 comments
read the original abstract

Perineural invasion (PNI) is a critical prognostic factor in cholangiocarcinoma. Non-invasive prediction from 3D MRI is challenging, demanding models that efficiently capture both fine-grained details and global context. We propose the Multi-window Mixture-of-Head Attention Transformer (MMA-Former), a novel end-to-end 3D architecture featuring a Coarse-Fine Transformer (CFT) structure for parallel multi-scale feature extraction. We advance this structure by integrating a novel Window-Specific Mixture-of-Head attention (WS-MoH) mechanism. Unlike standard Multi-Head Self Attention (MSA), WS-MoH generates a representation for each 3D window and dynamically routes the entire window to specialized or common attention heads. This enables spatially adaptive feature extraction tailored to the local context of each window, enhancing specialization and reducing redundancy without increasing parameters. Evaluated on a retrospective dataset of 168 T1-weighted MRI scans, MMA-Former achieved an AUC of 0.752, outperforming other 3D architectures, including the best CNN (AUC of 0.708) and Transformer baselines (AUC of 0.681).

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 4 linked inside Pith

  1. [1]

    Its presence significantly escalates the risk of recurrence, correlates with poor survival, and dictates surgical planning [1, 2, 3]

    INTRODUCTION Perineural invasion (PNI), the insidious infiltration of cancer cells along nerve sheaths, is a critical route of metastasis in cholangiocar- cinoma. Its presence significantly escalates the risk of recurrence, correlates with poor survival, and dictates surgical planning [1, 2, 3]. Accurate preoperative identification of PNI is therefore par...

    Pith/arXiv arXiv 2026

  2. [2]

    METHODOLOGY 2.1. Datasets and Preprocessing We utilized a retrospective dataset of anonymized T1-weighted, contrast-enhanced MR images (hepatobiliary phase) acquired from multiple MRI scanners at Samsung Medical Center over a decade. Images were provided in NIfTI format with ground truth annotations for the liver and tumor. After quality control, the fina...

  3. [3]

    Experimental Setup All experiments were performed on NVIDIA A100 GPUs

    EXPERIMENTS AND RESULTS 3.1. Experimental Setup All experiments were performed on NVIDIA A100 GPUs. We em- ployed stratified 5-fold cross-validation across the 168 cases. Mod- Table 2. Comprehensive ablation study of MMA-Former components and configurations on the PNI dataset. Category Configuration Description AUC BaselineMMA-Former (Full Model) Cropped ...

  4. [4]

    The effectiveness of the approach is strongly supported by the necessity of input localization

    DISCUSSION AND CONCLUSION The MMA-Former introduces a novel approach to 3D medical image analysis by enabling spatially adaptive feature extraction within a hi- erarchical transformer framework. The effectiveness of the approach is strongly supported by the necessity of input localization. The core innovation is the WS-MoH within the CFT framework. The CF...

  5. [5]

    2020-0-01305)

    ACKNOWLEDGMENTS This work was supported by the Institute of Information & Commu- nications Technology Planning & Evaluation (IITP), funded by the Korea government (MSIT), under the Artificial Intelligence Semi- conductor Support Program to nurture the best talents (IITP-2023- RS-2023-00256081) and the grant for the Development of an AI Deep Learning Proce...

    2023

  6. [6]

    Perineural invasion con- fers poorer clinical outcomes in patients with t1/t2 intrahep- atic cholangiocarcinoma: a single center, retrospective cohort study,

    M. Zou, J. Sheng, M. Ruan, W. Zhou, F. Ye, G. Yang, Y . Qian, J. Wang, R. Wang, S. Liuet al., “Perineural invasion con- fers poorer clinical outcomes in patients with t1/t2 intrahep- atic cholangiocarcinoma: a single center, retrospective cohort study,”Journal of Gastrointestinal Oncology, vol. 14, no. 6, p. 2500, 2023

    2023

  7. [7]

    Zhang, Y

    Z. Zhang, Y . Zhou, K. Hu, D. Wang, Z. Wang, and Y . Huang, “Perineural invasion as a prognostic factor for intrahepatic cholangiocarcinoma after curative resection and a potential in- dication for postoperative chemotherapy: a retrospective co- hort study,”Bmc Cancer, vol. 20, no. 1, p. 270, 2020

    2020

  8. [8]

    Z. Liu, C. Luo, X. Chen, Y . Feng, J. Feng, R. Zhang, F. Ouyang, X. Li, Z. Tan, L. Denget al., “Noninvasive predic- tion of perineural invasion in intrahepatic cholangiocarcinoma by clinicoradiological features and computed tomography ra- diomics based on interpretable machine learning: a multicenter cohort study,”International Journal of Surgery, vol. 11...

    2024

  9. [9]

    Perineural tumour spread in head and neck cancer: a picto- rial review,

    S. Doran, R. Whiriskey, N. Sheehy, C. Johnston, and D. Byrne, “Perineural tumour spread in head and neck cancer: a picto- rial review,”Clinical Radiology, vol. 79, no. 10, pp. 749–756, 2024

    2024

  10. [10]

    An mri-based fusion model for preoperative prediction of perineural invasion status in patients with intrahepatic cholangiocarcinoma,

    Z. Qi, H. Yuan, Q. Li, P. Chen, D. Li, K. Chen, B. Meng, P. Ning, H. Yu, and D. Li, “An mri-based fusion model for preoperative prediction of perineural invasion status in patients with intrahepatic cholangiocarcinoma,”World Journal of Sur- gical Oncology, vol. 23, no. 1, p. 164, 2025

    2025

  11. [11]

    Ct-based ra- diomics analysis for noninvasive prediction of perineural inva- sion of perihilar cholangiocarcinoma,

    P.-C. Zhan, P.-j. Lyu, Z. Li, X. Liu, H.-X. Wang, N.-N. Liu, Y . Zhang, W. Huang, Y . Chen, and J.-b. Gao, “Ct-based ra- diomics analysis for noninvasive prediction of perineural inva- sion of perihilar cholangiocarcinoma,”Frontiers in Oncology, vol. 12, p. 900478, 2022

    2022

  12. [12]

    Feasibility of magnetic resonance imaging-based ra- diomics features for preoperative prediction of extrahepatic cholangiocarcinoma stage,

    X. Huang, J. Shu, Y . Yan, X. Chen, C. Yang, T. Zhou, and M. Li, “Feasibility of magnetic resonance imaging-based ra- diomics features for preoperative prediction of extrahepatic cholangiocarcinoma stage,”European Journal of Cancer, vol. 155, pp. 227–235, 2021

    2021

  13. [13]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770– 778

    2016

  14. [14]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” inProceedings of the IEEE conference on computer vision and pattern recog- nition, 2017, pp. 4700–4708

    2017

  15. [15]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Trans- formers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  16. [16]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Sys- tems, vol. 30, 2017, pp. 6000–6010

    2017

  17. [17]

    Unetr: Transformers for 3d medical image segmentation,

    A. Hatamizadeh, Y . Tang, V . Nath, D. Yang, A. Myronenko, B. Landman, H. R. Roth, and D. Xu, “Unetr: Transformers for 3d medical image segmentation,” inProceedings of the IEEE/CVF winter conference on applications of computer vi- sion, 2022, pp. 574–584

    2022

  18. [18]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF in- ternational conference on computer vision, 2021, pp. 10 012– 10 022

    2021

  19. [19]

    Multiscale vision transformers,

    H. Fan, B. Xiong, K. Mangalam, Y . Li, Z. Yan, J. Malik, and C. Feichtenhofer, “Multiscale vision transformers,” inPro- ceedings of the IEEE/CVF international conference on com- puter vision, 2021, pp. 6824–6835

    2021

  20. [20]

    Focal self-attention for local-global interactions in vi- sion transformers,

    J. Yang, C. Li, P. Zhang, X. Dai, B. Xiao, L. Yuan, and J. Gao, “Focal self-attention for local-global interactions in vi- sion transformers,”arXiv preprint arXiv:2107.00641, 2021

    Pith/arXiv arXiv 2021

  21. [21]

    Moh: Multi- head attention as mixture-of-head attention,

    P. Jin, B. Zhu, L. Yuan, and S. Yan, “Moh: Multi- head attention as mixture-of-head attention,”arXiv preprint arXiv:2410.11842, 2024

    arXiv 2024

  22. [22]

    Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned,

    E. V oita, D. Talbot, F. Moiseev, R. Sennrich, and I. Titov, “Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned,” inProceedings of the 57th Annual Meeting of the Association for Computational Lin- guistics, 2019, pp. 5797–5808

    2019

  23. [23]

    Are sixteen heads really better than one?

    P. Michel, O. Levy, and G. Neubig, “Are sixteen heads really better than one?” inAdvances in Neural Information Process- ing Systems, vol. 32, 2019, pp. 14 014–14 024

    2019

  24. [24]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” inProceedings of the International Conference on Learning Representations, 2017

    2017

  25. [25]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,

    W. Fedus, B. Zoph, and N. Shazeer, “Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,”Journal of Machine Learning Research, vol. 23, no. 120, pp. 1–39, 2022

    2022

  26. [26]

    Efficientnet: Rethinking model scaling for convolutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 6105–6114

    2019

  27. [27]

    Decoupled weight decay regular- ization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regular- ization,”arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  28. [28]

    Grad-cam: Visual explanations from deep net- works via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep net- works via gradient-based localization,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 618–626

    2017