Pith. sign in

REVIEW 3 major objections 6 minor 31 references

A diffusion transformer with adaptive routing predicts perineural invasion from 3D MRI at 0.731 AUC while cutting compute.

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 04:55 UTC pith:YLNLMVJW

load-bearing objection Solid engineering paper: DiT + three-way adaptive routing for PNI MRI, clean ablations, small single-center n is the real limit. the 3 major comments →

arxiv 2607.11533 v1 pith:YLNLMVJW submitted 2026-07-13 cs.CV cs.LG

Adaptive Routing for Efficient Diffusion Transformer-Based PNI Prediction

Youngung Han , Dohyun Kweon , Kyeonghun Kim , Hyunsu Go , Jina Jeong , Suah Park , Induk Um , Junga Kim
show 10 more authors
Anna Jung Yului Jeong Sungha Park Jinyong Jun Pa Hong Woo Kyoung Jeong Won Jae Lee Ken Ying-Kai Liao Hyuk-Jae Lee Nam-Joon Kim
This is my paper
classification cs.CV cs.LG
keywords diffusion transformeradaptive routingtoken selectioncomputational efficiency3D MRI analysisperineural invasioncholangiocarcinoma
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 is a hard-to-see prognostic sign in cholangiocarcinoma that often lies outside the tumor itself. The paper shows that casting the prediction as class-conditional denoising inside a transformer, then dynamically routing attention heads, spatial tokens, and MLP width, can recover those subtle volumetric cues more reliably than ordinary CNNs or transformers, especially under noise. On a 155-patient single-center T2-MRI cohort the method reaches an AUC of 0.731 at 257.57 GFLOPs—nearly matching a non-routed version that costs far more. The practical claim is that the same accuracy can be obtained with roughly 40 percent fewer operations, making diffusion-transformer classifiers closer to usable in clinical settings. A sympathetic reader cares because preoperative PNI knowledge can change surgical and neoadjuvant decisions, yet current imaging models either miss the diffuse signal or are too expensive to run.

Core claim

Formulating preoperative PNI prediction as diffusion-based classification with a transformer denoising backbone, and inserting adaptive routing over attention heads, spatial tokens, and MLP width, yields an AUC of 0.731 at 257.57 GFLOPs on tumor-centered 3D T2 MRI—nearly matching the non-routed counterpart while remaining more noise-robust than CNN and standard transformer baselines.

What carries the argument

DiT-RC: a diffusion transformer whose blocks use lightweight routers (conditioned on timestep and global features, plus a local 3D convolution for tokens) to binary-mask attention heads, spatial tokens, and MLP channel groups, so only informative computation is performed under a budget loss.

Load-bearing premise

That results from 155 patients at one hospital, using hand-cropped tumor regions and internal cross-validation alone, are enough to support the claimed accuracy and efficiency for real clinical use.

What would settle it

An external multi-center test set of comparable size, acquired on different scanners and protocols, on which DiT-RC either drops below the CNN/transformer baselines or loses its FLOPs advantage relative to the non-routed model.

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 / 6 minor

Summary. The manuscript proposes DiT-RC, a diffusion-transformer classifier for preoperative perineural invasion (PNI) prediction from tumor-centered 3D T2-weighted MRI. Class scores are obtained by comparing class-conditional noise reconstruction errors under a transformer denoising backbone conditioned on timestep and class hypothesis (Sec. II-A). To reduce the cost of volumetric tokens and iterative denoising, the authors introduce adaptive routing over attention heads, spatial tokens (with a lightweight 3D-convolution importance estimator), and MLP width, trained with a joint diffusion-classification and keep-rate budget loss (Eqs. 3–8). On a single-center cohort of 155 patients with patient-level 5-fold CV, DiT-RC reports AUC 0.731 (bootstrap 95% CI 0.66–0.80) at 257.57 GFLOPs, nearly matching a non-routed DiT-C (0.733 at 418.79 GFLOPs), with improved noise robustness relative to CNN and standard transformer baselines and ablations of routing components and budget targets (Tables I–V, Figs. 2–3).

Significance. Preoperative PNI prediction is clinically relevant for cholangiocarcinoma planning, and combining diffusion-based class scoring with long-range volumetric transformers is a reasonable response to weak, noise-sensitive peritumoral cues. The main technical contribution is the joint adaptive routing design (heads, tokens with local 3D context, MLP width) under an explicit FLOPs-proxy budget, which is well ablated and yields a clear efficiency–accuracy trade-off (Tables III–V). Strengths include patient-level stratified CV, bootstrap CI, multi-noise robustness evaluation, a U-Net diffusion baseline, and transparent reporting that multi-center validation remains necessary. If the efficiency results hold under broader validation, the work is a useful methods contribution for deployable diffusion classifiers in 3D medical imaging; absolute clinical performance claims remain limited by cohort size.

major comments (3)
  1. Sec. III-A and Table I: the central performance claim (AUC 0.731, best across clean/noisy settings) rests on n=155 (61+/94−) single-center patients with a bootstrap CI of 0.66–0.80. That interval substantially overlaps reported baselines (e.g., Swin 0.710, ViT 0.700, Diffusion Classifier 0.692). No paired statistical comparison (e.g., DeLong or patient-level bootstrap tests of AUC differences) is provided. Without such tests—or a clearer framing that absolute superiority is not established—the ranking narrative in Sec. III-B is stronger than the evidence supports. Please add significance testing for key pairwise AUC comparisons and temper abstract/conclusion language accordingly.
  2. Sec. III-A / Table II–III: FLOPs and latency are reported “per patient prediction under the five-timestep and two-class protocol, excluding repeated noise ensembling,” while inference also averages over multiple timesteps and random seeds. For a paper whose primary selling point is efficiency, the evaluation protocol must make clear (i) which operations are included in the 257.57 GFLOPs / 140.47 ms numbers, (ii) the realized keep ratios k_attn, k_tok, k_mlp used for routed FLOPs, and (iii) how weights w_attn, w_mlp in Eq. (7) are set. Please report mean keep ratios per block/timestep and a full end-to-end cost that matches the protocol used to obtain the reported AUC.
  3. Sec. II-C, Eq. (8): L_cls = L_mse^(y) + α L_CE is underspecified for a diffusion classifier. Clarify whether L_mse is computed only under the true class y, under both class hypotheses, or as a contrastive term; how reconstruction errors m_y are turned into CE targets with learnable τ; and the values (or selection) of α, λ_budget, and the Gumbel–Sigmoid temperature. These choices are free parameters of the method and affect reproducibility of both AUC and the routing budget behavior in Tables III–V.
minor comments (6)
  1. Fig. 1 caption states N=6 blocks; confirm consistency with the FLOPs model and whether depth was ablated.
  2. Fig. 2 Grad-CAM comparison of DiT-RC vs DiT-C is qualitative only; a brief note on how gradients are taken through the diffusion scoring path would help.
  3. Table I noise levels NL/NM/NH should explicitly map to σ∈{0.1,0.3,0.5} in the table header or footnote.
  4. Related-work citations include several concurrent/self works on PNI MRI by overlapping authors; a short differentiation paragraph (what DiT-RC adds beyond LoSA-Net / MMA-Former / NeoNet) would improve novelty framing.
  5. Notation: Eq. (2) writes ε̂_θ(x_t,t,y) but the surrounding text sometimes refers to reconstruction error without restating the noise target; keep a single consistent definition of m_y.
  6. Minor prose: “V olumetric” spacing artifact in Sec. II-B.2; standardize “DiT-RC” vs “diffusion transformer” naming in the abstract vs body.

Circularity Check

0 steps flagged

No significant circularity: empirical ML method and CV-reported metrics are self-contained against the paper's own experiments.

full rationale

This is a standard empirical computer-vision paper proposing DiT-RC (diffusion transformer classifier + adaptive head/token/MLP routing) and evaluating it via patient-level 5-fold CV on a 155-patient single-center T2-MRI cohort. Classification scores are obtained by comparing class-conditional noise reconstruction errors (Sec. II-A: my = ||ε̂ heta(xt,t,y)-ε||²₂, ŷ = arg min my); routing decisions are produced by Gumbel–Sigmoid routers and regularized by an explicit budget loss Lbudget = (reff - λ target)² (Eqs. 6–8). Reported AUC 0.731 / 257.57 GFLOPs, ablations (Tables III–V), noise robustness (Table I) and Grad-CAM maps are measured quantities, not algebraic identities of the inputs. Self-citations [2–4] appear only as related prior PNI-MRI work by overlapping authors and are not invoked as uniqueness theorems, uniqueness proofs, or load-bearing premises for the DiT-RC numbers. No fitted constant is renamed a prediction, no ansatz is smuggled via self-citation, and no derivation collapses by construction. The paper itself flags the need for multi-center validation; that is a generalization caveat, not circularity. Score 0 is therefore the correct, proportionate finding.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim is an empirical performance/efficiency result. It rests on standard diffusion and transformer machinery, domain assumptions about MRI-visible PNI cues, a private labeled cohort, and several hand-chosen training/routing hyperparameters (especially λ_target and architecture depth). No new physical entities are postulated; the invented object is the DiT-RC architecture itself.

free parameters (5)
  • λ_target (routing budget target) = 0.5
    Controls keep-rate / FLOPs–AUC trade-off; set to 0.5 as 'Ours' after sweep {0.3,0.5,0.7,1.0} on the same task (Table V).
  • λ_budget / α (loss weights) and learnable temperature τ
    Scale classification vs budget terms and convert reconstruction errors to logits; values are training choices not derived from first principles.
  • Architecture depth N and patch/ROI geometry = N=6; 96×96×48 ROI; 6³ patches
    N=6 blocks, ROI 96×96×48, patches 6×6×6 → 16×16×8 tokens; fixed design choices that define the compute baseline.
  • Inference timestep set and ensembling protocol = five timesteps listed
    Logits averaged over t∈{200,350,500,650,800} and random seeds; chosen protocol that affects reported AUC/latency.
  • AdamW LR / weight decay / batch size = 8e-5 / 0.05 / 4
    8e-5, 0.05, batch 4 on one RTX 3090; standard but free training hyperparameters.
axioms (4)
  • domain assumption Class-conditional noise reconstruction error is a valid scoring rule for binary PNI status on MRI.
    Sec. II-A defines ŷ = arg min_y ||ε̂_θ(x_t,t,y)−ε||²; the method assumes this denoising score separates subtle PNI patterns better than pure discriminative heads.
  • domain assumption Tumor-centered crops from T2-weighted MRI contain sufficient signal for preoperative PNI prediction without using masks as model inputs.
    Dataset section: manual 3D Slicer masks used only for ROI extraction; clinical claim depends on this imaging sufficiency.
  • standard math Standard forward diffusion, transformer self-attention, AdaLN conditioning, and Gumbel–Sigmoid discrete routing are valid building blocks.
    Eq. (1) and routing sections import DDPM-style noise schedules and common differentiable discrete gates without re-deriving them.
  • ad hoc to paper Patient-level 5-fold stratified CV on one center estimates generalization for the reported AUC/FLOPs claims.
    Experimental protocol; the paper later admits multi-center validation is still needed, so this is a paper-specific evaluation axiom.
invented entities (2)
  • DiT-RC (Diffusion Transformer with Routing for Classification) no independent evidence
    purpose: Name the full system: diffusion-transformer classifier plus joint head/token/MLP-width adaptive routing and local-conv token scoring for efficient PNI prediction.
    Architectural composite introduced in this paper; independent evidence is only the internal ablations/tables, not external replications.
  • Local 3D-convolution token importance estimator for routing no independent evidence
    purpose: Score spatial tokens using 3×3×3 context on the token grid so routing focuses near tumor boundaries.
    Design module specific to this work; ablation 'w/o Local conv' drops AUC to 0.700 (Table IV).

pith-pipeline@v1.1.0-grok45 · 13364 in / 3691 out tokens · 37756 ms · 2026-07-14T04:55:15.255661+00:00 · methodology

0 comments
read the original abstract

Perineural invasion (PNI) is a critical prognostic factor in cholangiocarcinoma. However, its preoperative prediction from magnetic resonance imaging (MRI) remains challenging due to subtle imaging features that extend beyond tumor boundaries into surrounding regions. Conventional convolutional neural networks are limited in capturing long-range spatial dependencies. Transformer-based architectures improve global modeling of volumetric MRI by aggregating spatially distributed contextual cues, yet capturing subtle and noise-sensitive patterns in peritumoral regions remains challenging. Diffusion-based classifiers offer an alternative formulation by leveraging denoising-based class scoring to better capture such subtle patterns. However, these approaches introduce substantial computational overhead due to the combination of transformer-based modeling and iterative denoising processes. To address these challenges, we formulate PNI prediction as a diffusion-based classification problem and implement the denoising network using a transformer-based representation. To improve computational efficiency, we introduce adaptive routing across attention heads, spatial tokens, and MLP width. Experimental results demonstrate that the proposed approach achieves an AUC of 0.731 with 257.57 GFLOPs.

Figures

Figures reproduced from arXiv: 2607.11533 by Anna Jung, Dohyun Kweon, Hyuk-jae Lee, Hyunsu Go, Induk Um, Jina Jeong, Jinyong Jun, Junga Kim, Ken Ying-Kai Liao, Kyeonghun Kim, Nam-Joon Kim, Pa Hong, Suah Park, Sungha Park, Won Jae Lee, Woo Kyoung Jeong, Youngung Han, Yului Jeong.

Figure 1
Figure 1. Figure 1: Overview of DiT-RC. The architecture consists of a stack of N transformer blocks (N = 6 in our implementation). [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Grad-CAM visualization for PNI prediction. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Routing analysis. (A) Average head and width routing [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

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

31 extracted references · 6 linked inside Pith

  1. [1]

    Prognostic impact of perineural invasion in intrahepatic cholangiocarcinoma: mul- ticentre study,

    T. Wei, X.-F. Zhang, J. He, I. Popescu, H. P. Marques, L. Aldrighetti, S. K. Maithel, C. Pulitano, T. W. Bauer, F. Shen,et al., “Prognostic impact of perineural invasion in intrahepatic cholangiocarcinoma: mul- ticentre study,”British Journal of Surgery, vol. 109, no. 7, pp. 610–616, 2022

    2022

  2. [2]

    Losa-net: A localized and scale-adaptive network for boundary-sensitive prediction of perineural invasion in 3d mri,

    Y . Han, H. Go, K. Kim, I. Um, J. Kim, J. Jung, N.-J. Kim, W. K. Jeong, W. J. Lee, K. Y .-K. Liao,et al., “Losa-net: A localized and scale-adaptive network for boundary-sensitive prediction of perineural invasion in 3d mri,” in2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI), pp. 1–5, IEEE, 2026

    2026

  3. [3]

    Mma-former: Multi-window mixture- of-head attention transformer for adaptive pni prediction in 3d mri,

    Y . Han, I. Um, K. Kim, J. Kim, H. Go, J. Jung, N.-J. Kim, W. K. Jeong, W. J. Lee, P. Hong,et al., “Mma-former: Multi-window mixture- of-head attention transformer for adaptive pni prediction in 3d mri,” in2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI), pp. 1–5, IEEE, 2026

    2026

  4. [4]

    Neonet: An end-to-end 3d mri-based deep learning framework for non-invasive prediction of perineural invasion via generation-driven classification,

    Y . Han, M. Cha, K. Kim, I. Um, M. Sho, J. Y . Bae, J. Jung, J. H. Park, S. Lee, N.-J. Kim,et al., “Neonet: An end-to-end 3d mri-based deep learning framework for non-invasive prediction of perineural invasion via generation-driven classification,”arXiv preprint arXiv:2603.29449, 2026

    arXiv 2026

  5. [5]

    Z. Liu, C. Luo, X. Chen, Y . Feng, J. Feng, R. Zhang, F. Ouyang, X. Li, Z. Tan, L. Deng,et al., “Noninvasive prediction of perineural inva- sion in intrahepatic cholangiocarcinoma by clinicoradiological features and computed tomography radiomics based on interpretable machine learning: a multicenter cohort study,”International Journal of Surgery, vol. 110...

    2024

  6. [6]

    Perineural invasion in cancer: a review of the literature,

    C. Liebig, G. Ayala, J. A. Wilks, D. H. Berger, and D. Albo, “Perineural invasion in cancer: a review of the literature,”Cancer: Interdisciplinary International Journal of the American Cancer Society, vol. 115, no. 15, pp. 3379–3391, 2009

    2009

  7. [7]

    Conti, N

    S. Conti, N. S. Tissera, F. Castet, M. Basaga ˜na-Farr´es, M. T. Salcedo, E. Pando, C. Dopazo, L. Carit´a, A. Turpin, V . N. Garc´es,et al., “Perineu- ral invasion is a prognostic factor in cholangiocarcinoma, regardless of anatomical location: a systematic review and meta-analysis: Perineural invasion in cholangiocarcinoma prognosis,”JHEP Reports, p. 101...

    2026

  8. [8]

    Perineural invasion of hilar cholangiocarcinoma in chinese population: One center’s experi- ence,

    C.-G. Li, Z.-P. Zhou, X.-L. Tan, and Z.-M. Zhao, “Perineural invasion of hilar cholangiocarcinoma in chinese population: One center’s experi- ence,”World Journal of Gastrointestinal Oncology, vol. 12, no. 4, p. 457, 2020

    2020

  9. [9]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,”arXiv preprint arXiv:1409.1556, 2014

    Pith/arXiv arXiv 2014

  10. [10]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly,et al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  11. [11]

    Unetr: Transformers for 3d medical image segmentation,

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

    2022

  12. [12]

    Transunet: Transformers make strong encoders for medical image segmentation,

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

    Pith/arXiv arXiv 2021

  13. [13]

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

    F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. 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

  14. [14]

    Robust classification via a single diffusion model,

    H. Chen, Y . Dong, Z. Wang, X. Yang, C. Duan, H. Su, and J. Zhu, “Robust classification via a single diffusion model,”arXiv preprint arXiv:2305.15241, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    Text-to-image diffusion models are zero shot classifiers,

    K. Clark and P. Jaini, “Text-to-image diffusion models are zero shot classifiers,”Advances in Neural Information Processing Systems, vol. 36, pp. 58921–58937, 2023

    2023

  16. [16]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations (ICLR), 2021

    2021

  17. [17]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  18. [18]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 4195–4205, 2023

    2023

  19. [19]

    Foscu: Feasibility of synthetic mri generation via duo-diffusion models for enhancement of 3d u-nets in hepatic segmentation,

    Y . Han, K. Kim, S. Ju, Y . Jean, M. Cha, S. Park, H. Jung, N.-J. Kim, W. K. Jeong, K. Y .-K. Liao,et al., “Foscu: Feasibility of synthetic mri generation via duo-diffusion models for enhancement of 3d u-nets in hepatic segmentation,” in2025 IEEE Asia Pacific Conference on Circuits and Systems (APCCAS), pp. 1–5, IEEE, 2025

    2025

  20. [20]

    3d-lldm: Label-guided 3d latent diffusion model for improving high-resolution synthetic mr imaging in hepatic structure segmentation,

    K. Kim, J. Bae, Y . Han, J. Y . Bae, S. Ju, J. Lim, G. Kim, N.-J. Kim, W. K. Jeong, K. Y .-K. Liao,et al., “3d-lldm: Label-guided 3d latent diffusion model for improving high-resolution synthetic mr imaging in hepatic structure segmentation,” in2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI), pp. 1–5, IEEE, 2026

    2026

  21. [21]

    Dynam- icvit: Efficient vision transformers with dynamic token sparsification,

    Y . Rao, W. Zhao, B. Liu, J. Lu, J. Zhou, and C.-J. Hsieh, “Dynam- icvit: Efficient vision transformers with dynamic token sparsification,” Advances in neural information processing systems, vol. 34, pp. 13937– 13949, 2021

    2021

  22. [22]

    Dit: Self-supervised pre-training for document image transformer,

    J. Li, Y . Xu, T. Lv, L. Cui, C. Zhang, and F. Wei, “Dit: Self-supervised pre-training for document image transformer,” inProceedings of the 30th ACM international conference on multimedia, pp. 3530–3539, 2022

    2022

  23. [23]

    Adaln: A vision transformer for multidomain learning and predisaster building information extraction from images,

    Y . Guo, C. Wang, S. X. Yu, F. McKenna, and K. H. Law, “Adaln: A vision transformer for multidomain learning and predisaster building information extraction from images,”Journal of Computing in Civil Engineering, vol. 36, no. 5, p. 04022024, 2022

    2022

  24. [24]

    Film: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, 2018

    2018

  25. [25]

    Categorical reparameterization with gumbel-softmax,

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,”arXiv preprint arXiv:1611.01144, 2016

    Pith/arXiv arXiv 2016

  26. [26]

    Once-for-all: Train one network and specialize it for efficient deployment,

    H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han, “Once-for-all: Train one network and specialize it for efficient deployment,”arXiv preprint arXiv:1908.09791, 2019

    Pith/arXiv arXiv 1908

  27. [27]

    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, pp. 770–778, 2016

    2016

  28. [28]

    Densely connected convolutional networks,

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

    2017

  29. [29]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

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

    2019

  30. [30]

    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 international conference on computer vision, pp. 10012–10022, 2021

    2021

  31. [31]

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

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

    2017