Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

EVM-Fusion: An Explainable Vision Mamba Architecture with Neural Algorithmic Fusion

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A three-path Mamba architecture with algorithm-learned fusion reaches 94.79% accuracy on a 9-class medical triage benchmark while explaining its own decisions.

desk verdict An interesting multi-path fusion architecture undermined by an unspecified dataset split and an internal accuracy contradiction; worth peer review, not acceptance. read the letter →

arxiv 2505.17367 v4 pith:MF5CP76A submitted 2025-05-23 cs.CV cs.AI

Zichuan Yang , Yongzhi Wang This is my paper
classification cs.CVcs.AI
keywords explainableAIVisionMambaneuralalgorithmicfusionmedicalimagetriagemulti-pathclassificationcross-modalattentionablationstudy
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

The paper tries to establish that holistic medical image triage is best done by combining three complementary feature extractors—a DenseNet branch, a U-Net branch, and a handcrafted texture-feature branch—each augmented with attention and, for the deep branches, a Vision Mamba module—and then fusing them with a two-stage, learnable process rather than a fixed rule. It reports that this system, EVM-Fusion, reaches 94.79% test accuracy on a pooled 9-class dataset spanning brain MRI, breast X-ray, and chest X-ray images, beating strong single-paradigm baselines. The paper further claims that every component earns its place: removing the U-Net branch drops accuracy to 88.20%, and replacing the two-stage fusion with concatenation or mean pooling loses several points. The broader significance is that the same architecture provides intrinsic explanations—spatial attention maps, Mamba delta-value maps, texture-feature importance scores, and cross-modal attention weights—so a clinician could see why a triage decision was made.

What carries the argument

The load-bearing mechanism is the two-stage Core Fusion Module. Stage A is cross-modal attention: the three path vectors are stacked and passed through multi-head self-attention so each path can re-weight the others before fusion. Stage B is the Neural Algorithmic Fusion (NAF) block, a controller-driven iterative refinement: a GRU observes the current fused state, emits softmax mixing weights, and those weights combine the outputs of a fixed bank of small MLP primitives; the mixture is added to the state through a residual connection and layer normalization, then repeated for a number of iterations before a final linear projection. The Vision Mamba modules supply the long-range modeling; the paper stores their data-dependent delta values as explanations, and path-specific spatial attention and SE attention on the texture features complete the intrinsic XAI picture.

What would settle it

Re-train under a patient-level, source-stratified split that keeps each patient's images in one fold and report per-source accuracy; the central claim is falsified if the margin over the baselines shrinks to noise or if a simple classifier can predict the dataset source of a test image from the model's features. The abstract's 99.75% and Section V's 94.79% are both printed in the paper, so a first audit is simply determining which number the evaluation actually produced.

Watch

Extended reading notes

Core claim

EVM-Fusion is a three-path network. The DenseNet path extracts reusable semantic features, the U-Net path preserves multi-scale local detail, and the traditional path supplies GLCM and LBP texture statistics; the two deep paths are each refined by a bidirectional Vision Mamba module with input-dependent gating. The three path vectors are contextualized by cross-modal attention and then refined by a Neural Algorithmic Fusion block, where a GRU controller iteratively mixes a bank of MLP primitives to compose the final fused vector for classification. The paper's central discovery is that this division of labor plus learnable iterative fusion is what produces the reported 94.79% accuracy, and that the model's internal attention maps and delta-value maps reveal interpretable, input-dependent decision behavior. On the authors' own account, the best fusion is not a single weighted sum but a short learned algorithm: contextualize first, then iterate the state through a controller-chosen mixture of primitive transformations.

Load-bearing premise

The reported accuracy rests on the assumption that a random split of the pooled brain, breast, and chest X-ray images measures genuine multi-organ triage ability; if the split is not patient-level and source-stratified, the model may be exploiting source-specific appearance rather than learning the diagnostic patterns the paper claims.

Editorial extensions

If this is right

  • Removing the U-Net path drops accuracy from 94.79% to 88.20%, so the multi-scale detail branch is not redundant.
  • Using only one fusion stage (CMA-Only 89.34% or NAF-Only 86.81%) underperforms the full two-stage pipeline, so both stages earn their place.
  • Simple concatenation (87.21%) and mean pooling (88.80%) are clearly below the full method, supporting the claim that fixed fusion rules fail on heterogeneous features.
  • Because spatial attention maps, delta-value maps, SE scores, and cross-modal weights are produced intrinsically, the same forward pass yields both a prediction and a multi-view explanation.
  • The model outperforms strong single-paradigm baselines from CNN, transformer, and hybrid families on the same split, indicating that synergy among paths is the source of the gain.

Reading between the lines

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

  • A natural test the paper leaves implicit: re-split the pooled dataset by patient and by source; if accuracy holds, EVM-Fusion's triage claim is much stronger than if the split is random at the image level.
  • The NAF controller-primitive design could be lifted out and applied to other multi-stream fusion problems—combining clinical metadata with images or fusing multiple imaging modalities—since it does not depend on the specific paths used here.
  • The manuscript prints two different headline accuracies (99.75% in the abstract, 94.79% in Table III), so any external replication should first confirm which number corresponds to the released evaluation.
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 / 5 minor

Summary. The paper proposes EVM-Fusion, a three-path architecture for multi-organ medical image triage. A DenseNet-Mamba path and a U-Net-Mamba path provide deep semantic and multi-scale features, while a third path supplies GLCM/LBP statistical features; all three are combined by a cross-modal attention stage and an iterative Neural Algorithmic Fusion (NAF) block. The authors claim state-of-the-art accuracy on a composite 9-class dataset assembled from brain MRI, breast X-ray, and chest X-ray sources, and they provide qualitative intrinsic explanations via spatial attention maps, Mamba delta-value maps, SE-attention scores, and cross-modal attention weights.

Significance. If the empirical claims are valid, the paper would make a useful contribution to multi-organ medical image triage by showing that fusing diverse architectural families (CNN, U-Net, Mamba, handcrafted features) with a learned iterative fusion rule can outperform strong single-paradigm baselines, while providing multiple intrinsic explanation channels. The use of three public datasets, comparisons against six named baselines, and detailed per-path and per-fusion ablations are commendable strengths. However, the current evidence is conditional on resolving a direct contradiction in the reported accuracy and on demonstrating that the data split does not permit source or patient leakage; until these are addressed, the claimed superiority is not established.

major comments (4)
  1. [Abstract vs. Section V, Table III] The abstract reports 99.75% test accuracy, while Section V and Table III report 94.79% for the same EVM-Fusion model. This is a direct internal contradiction and the headline number is therefore ambiguous. The Discussion text and the ablation narrative both depend on 94.79%, so the authors must decide which number is correct and reconcile the abstract accordingly.
  2. [Section IV, Table I] The paper never states whether the train/validation/test split is performed at the patient level or at the image level, nor whether it is stratified by source. Because each of the nine class labels is perfectly confounded with one of the three imaging modalities (brain MRI, breast X-ray, chest X-ray), an image-level random split would allow the model to exploit source-specific appearance and any patient-level duplication; this would inflate every accuracy and per-class score in Tables II and III. A patient-level, source-stratified split with per-source and per-class results is required to support the stated multi-organ triage claims.
  3. [Section IV and Tables II-III] All reported results are single runs with no confidence intervals, no number of seeds, and no description of hyperparameter ranges (K_NAF, N_primitives, d_model_fusion, d_state, d_ctrl, learning rate, epochs, batch size, image resolution). The paper also does not provide code, a model checkpoint, or an evaluation protocol. Consequently, the claimed state-of-the-art margin over baselines and the ablation conclusions are not independently reproducible or statistically assessable.
  4. [Section II-B] The NAF mechanism is cited as "NAF [ ?]" with a missing reference. Since the paper's central novelty is the Neural Algorithmic Fusion block, the authors must provide the correct citation or clearly state that NAF is introduced here; otherwise the relationship to prior neural-algorithmic work is unverifiable.
minor comments (5)
  1. [Section III-B2] The U-Net-Mamba path is described as producing "the feature vector vD ∈ R^{d_model_fusion}", which is a typo; this should be vU to match the notation introduced in Section III-A.
  2. [Table III] The baseline is labeled "Swim Transformer" but the cited model is the Swin Transformer; the table header should be corrected.
  3. [Section IV-A heading] The heading "subsectionAblation Studies: Validating the Synergistic Framework" appears as a malformed LaTeX command in the rendered text and should be formatted as a proper subsection.
  4. [Section V, Fig. 2 and Fig. 3] The discussion of Fig. 3 would benefit from explicitly stating that the misclassified examples are selected from which classes and how representative they are of the overall error pattern; the current narrative is anecdotal.
  5. [Throughout] The paper would benefit from a short statement on clinical limitations and intended use, since the task is framed as a triage system; this is a presentation issue rather than a technical flaw.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: EVM-Fusion's accuracy claims are empirical results from external public datasets and baseline comparisons, not reductions to fitted inputs or self-citations.

full rationale

The paper's central numbers (94.79% accuracy in Section V/Table III, and the ablation drops in Table II) are obtained by training and testing EVM-Fusion on three public datasets: a brain tumor MRI dataset from Kaggle, a breast cancer X-ray dataset from Kaggle, and a chest X-ray pneumonia dataset from Mendeley Data. No parameter is fitted to a subset of data and then renamed as a prediction; no claimed result is defined in terms of the quantity it is supposed to predict; and no load-bearing premise is justified by a self-citation. The NAF block is presented as a novel architectural proposal (Algorithm 1) without invoking any uniqueness theorem or importing a forced choice from the authors' prior work. The related-work section contains a placeholder citation 'NAF [ ?]' in Section II-B, but this is a missing-reference and completeness issue, not a circularity issue. Likewise, the discrepancy between the abstract's 99.75% test accuracy and Section V's 94.79% is an internal-consistency/reproducibility concern, not evidence that the result reduces to its inputs. The reviewer's concern about random versus patient-level splitting affects external validity and generalization claims, but it does not make the derivation circular. Under the stated rules, a paper that is self-contained against external benchmarks and contains no parametric self-derivation should receive a low score; here the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The ledger shows that the architecture's success depends on many undisclosed hyperparameters and on assumptions about dataset splits, pretraining, and the transfer of specific feature extractors. No new physical entities are introduced; the NAF block is a learned mechanism rather than an invented entity.

free parameters (5)
  • NAF iteration count K_NAF = not reported
    Controls the number of fusion refinement steps; no sensitivity analysis is provided, so the optimality of NAF over simpler fusion is not isolated.
  • Number of primitive operations N_primitives = not reported
    Size of the primitive bank in the NAF block; chosen by hand and unspecified.
  • Feature dimensions d_model_fusion, d_state, d_ctrl = not reported
    Dimensions of the fused representation and NAF controller; no ablation or specification is given.
  • GLCM and LBP parameters = not reported
    Distance, angles, LBP radius, and histogram bins affect the traditional feature path; the paper does not state the values.
  • Training hyperparameters = not reported
    Epochs, learning rate, batch size, optimizer, data augmentation, and random seeds are absent; these can change the ranking in Table III.
assumptions (5)
  • domain assumption A random split of the concatenated public datasets yields a valid estimate of triage generalizability
    Section IV constructs the train/validation/test split but does not report patient-level separation or stratification; all accuracy claims inherit this assumption.
  • domain assumption ImageNet-pretrained weights provide a fair common starting point and fine-tuning baselines without specialized tuning is comparable
    Section IV states all models were trained under identical conditions, but no hyperparameters are given, so fairness rests on this assumption.
  • domain assumption Vision Mamba modules capture long-range dependencies useful for these medical images
    The S6 state-space update in Eq. (1) is adopted from [32]; the paper does not independently validate this beyond the reported accuracy.
  • standard math Standard neural network and attention equations are valid as implemented
    Equations (1)-(10) use standard state-space, attention, and GRU machinery from the literature; no formal verification is provided.
  • domain assumption GLCM and LBP texture features carry complementary diagnostic information for these classes
    Motivates the traditional feature path, but the specific GLCM/LBP parameter choices are not disclosed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EVM-Fusion: An Explainable Vision Mamba Architecture with Neural Algorithmic Fusion." pith.science (2026). https://pith.science/paper/MF5CP76A

@misc{pith2026250517367,
  author       = {Pith},
  title        = {Pith review of: EVM-Fusion: An Explainable Vision Mamba Architecture with Neural Algorithmic Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MF5CP76A}},
  note         = {Machine review of arXiv:2505.17367}
}
read the original abstract

Medical image classification is critical for clinical decision-making, yet demands for accuracy, interpretability, and generalizability remain challenging. This paper introduces EVM-Fusion, an Explainable Vision Mamba architecture featuring a novel Neural Algorithmic Fusion (NAF) mechanism for multi-organ medical image classification. EVM-Fusion leverages a multipath design, where DenseNet and U-Net based pathways, enhanced by Vision Mamba (Vim) modules, operate in parallel with a traditional feature pathway. These diverse features are dynamically integrated via a two-stage fusion process: cross-modal attention followed by the iterative NAF block, which learns an adaptive fusion algorithm. Intrinsic explainability is embedded through path-specific spatial attention, Vim {\Delta}-value maps, traditional feature SE-attention, and cross-modal attention weights. Experiments on a diverse 9-class multi-organ medical image dataset demonstrate EVM-Fusion's strong classification performance, achieving 99.75% test accuracy and provide multi-faceted insights into its decision-making process, highlighting its potential for trustworthy AI in medical diagnostics.

Figures

Figures reproduced from arXiv: 2505.17367 by the authors.

Figure 1
Figure 1. The proposed EVM-Fusion architecture. The model takes a single image as input and processes it through three parallel feature extraction and enhancement paths: (1) a DenseNet-Mamba path leveraging a DenseNet121 backbone followed by a Vision Mamba module and path-specific spatial attention; (2) a U-Net-Mamba path employing a full U-Net encoder-decoder structure, with a Vision Mamba module applied to the final decoder… view at source ↗
Figure 2
Figure 2. Explainable AI (XAI) Analysis of Representative Correctly Classified Test Samples by EVM-DUHF Across All Classes. This figure showcases the intrinsic explainability outputs for nine correctly classified samples, one from each of the distinct medical image classes in the test set. Noted from (a) to (i). Similar to [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Explainable AI (XAI) Analysis of Representative Misclassified Test Samples by EVM-Fusion. This figure presents a consolidated view of intrinsic explainability outputs for five misclassified samples from the 9-class test set. Noted from (a) to (e). Each sample is analyzed using eight distinct XAI visualizations. From top to bottom, from left to right, they are DenseNet Vision Mamba Forward ∆ Map, DenseNet Vision Mamb… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. User-Centric Modeling of Transactional Sequences with Explainable State Space Models

    cs.LG 2026-07 conditional novelty 4.0 of 10

    Injecting a pretrained CoLES user embedding into Mamba as an initial hidden state or prefix token improves accuracy by up to 3.2 pp on three transaction benchmarks and speeds convergence 2–3x.

Reference graph

Works this paper leans on

53 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    Inter-Observer and Intra-Observer Variability in Gross Tumor Volume Delineation of Primary Esophageal Carcinomas Based on Different Combinations of Diagnostic Multimodal Images,

    F. Li, Y . Li, X. Wang, Y . Zhang, X. Liu, S. Liu, W. Wang, J. Wang, Y . Guo, M. Xu, and J. Li, “Inter-Observer and Intra-Observer Variability in Gross Tumor Volume Delineation of Primary Esophageal Carcinomas Based on Different Combinations of Diagnostic Multimodal Images,” Front. Oncol., vol. 12, p. 817413, Apr 2022, PMID: 35433413; PMCID: PMC9010659. A...

    work page 2022

  2. [2]

    Improving SVM classification accuracy using a hierarchical approach for hyperspectral images,

    B. Demir and S. Ert ¨urk, “Improving SVM classification accuracy using a hierarchical approach for hyperspectral images,” in 2009 16th IEEE International Conference on Image Processing (ICIP) . Cairo, Egypt: IEEE, 2009, pp. 2849–2852

    work page 2009

  3. [3]

    A Survey of Convolu- tional Neural Networks: Analysis, Applications, and Prospects,

    Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A Survey of Convolu- tional Neural Networks: Analysis, Applications, and Prospects,” IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 12, pp. 6999–7019, 2022

    work page 2022

  4. [4]

    A simple and intelligent approach for brain MRI classification,

    M. Nazir, F. Wahid, and S. A. Khan, “A simple and intelligent approach for brain MRI classification,” Journal of Intelligent & Fuzzy Systems , vol. 28, no. 3, pp. 1127–1135, 2015, https://journals.sagepub.com/doi/ abs/10.3233/IFS-141396

    work page doi:10.3233/ifs-141396 2015

  5. [5]

    Normalized cuts and image segmentation,

    J. Shi and J. Malik, “Normalized cuts and image segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 22, no. 8, pp. 888–905, 2000

    2000

  6. [6]

    LV-Transformer: Language Meets Vision Transformer in Medical Image Segmentation,

    Z. Li, Y . Li, Q. Li, P. Wang, D. Guo, L. Lu, D. Jin, Y . Zhang, and Q. Hong, “LV-Transformer: Language Meets Vision Transformer in Medical Image Segmentation,” IEEE Transactions on Medical Imaging , vol. 43, no. 1, pp. 96–107, 2024

    work page 2024

  7. [7]

    MTANet: Multi-Task Attention Network for Automatic Medical Image Segmenta- tion and Classification,

    Y . Ling, Y . Wang, W. Dai, J. Yu, P. Liang, and D. Kong, “MTANet: Multi-Task Attention Network for Automatic Medical Image Segmenta- tion and Classification,” IEEE Transactions on Medical Imaging, vol. 43, no. 2, pp. 674–685, 2024

    work page 2024

  8. [8]

    Segmentation Aided Multiclass Tumor Classification in Ultrasound Images using Graph Neural Net- work,

    I. I. Shovon, I. Ahmad, and S. Shin, “Segmentation Aided Multiclass Tumor Classification in Ultrasound Images using Graph Neural Net- work,” in 2025 International Conference on Artificial Intelligence in Information and Communication (ICAIIC) . Fukuoka, Japan: IEEE, 2025, pp. 1012–1015

    work page 2025

Show all 53 references
  1. [9]

    ImageNet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,”Commun. ACM, vol. 60, no. 6, pp. 84–90, Jun 2017, https://doi.org/10.1145/3065386

    2017 doi

  2. [10]

    Very Deep Convolutional Networks for Large-Scale Image Recognition,

    K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” 2014, preprint at https://arxiv.org/ abs/1409.1556

    2014 arXiv

  3. [11]

    Multi-scale feature based medical image classification,

    B. Li, W. Li, and D. Zhao, “Multi-scale feature based medical image classification,” in Proceedings of 2013 3rd International Conference on Computer Science and Network Technology . Dalian: IEEE, 2013, pp. 1182–1186

    2013

  4. [12]

    Medical image classification with convolutional neural network,

    Q. Li, W. Cai, X. Wang, Y . Zhou, D. D. Feng, and M. Chen, “Medical image classification with convolutional neural network,” in 2014 13th International Conference on Control Automation Robotics and Vision (ICARCV). Singapore: IEEE, 2014, pp. 844–848

    2014

  5. [13]

    Clas- sification of X-Ray Images Using Grid Approach,

    F. Bertalya, B. Prihandoko, D. Kerami, and T. M. Kusuma, “Clas- sification of X-Ray Images Using Grid Approach,” in 2008 IEEE International Conference on Signal Image Technology and Internet Based Systems (SITIS) . Bali, Indonesia: IEEE, 2008, pp. 314–319

    2008

  6. [14]

    Hyperspectral Image Classification Using Random Forest and Deep Learning Algorithms,

    J. V . Rissati, P. C. Molina, and C. S. Anjos, “Hyperspectral Image Classification Using Random Forest and Deep Learning Algorithms,” in 2020 IEEE Latin American GRSS & ISPRS Remote Sensing Conference (LAGIRS). Santiago, Chile: IEEE, 2020, pp. 132–132

    2020

  7. [15]

    Medical CT Images Classification Model Based on BBO-HS Algorithm Optimized SVM,

    S. Rui, “Medical CT Images Classification Model Based on BBO-HS Algorithm Optimized SVM,” in 2020 IEEE International Conference on Power, Intelligent Computing and Systems (ICPICS). Shenyang, China: IEEE, 2020, pp. 1053–1057

    2020

  8. [16]

    Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model,” in Proceedings of the 41st International Conference on Machine Learning (ICML’24) . Vienna, Austria: JMLR.org, 2024, pp. 1–14

    2024

  9. [17]

    A Novel Attention-based Explainable Deep Learning Framework Towards Medical Image Classification,

    G. W. Muoka, D. Yi, C. C. Ukwuoma, M. D. Martin, A. A. Aydin, and M. A. Al-Antari, “A Novel Attention-based Explainable Deep Learning Framework Towards Medical Image Classification,” in 2023 7th Inter- national Symposium on Innovative Approaches in Smart Technologies (ISAS). A...

    2023

  10. [18]

    U-Net: Convolutional Net- works for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Net- works for Biomedical Image Segmentation,” in Medical Image Com- puting and Computer-Assisted Intervention – MICCAI 2015 , N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer Internationa...

    2015

  11. [19]

    U-DENSENET Deep Learning Model for Medical Image Segmentation,

    A. G. Sarparrah, N. Ahmed, M. Dawood, S. Husain, S. Mohammad, and M. A. Dehwar, “U-DENSENET Deep Learning Model for Medical Image Segmentation,” VAWKUM Transactions on Computer Sciences , vol. 12, no. 2, pp. 112–122, Nov 2024, https://vfast.org/journals/index. php/VTCS/article...

    2024

  12. [20]

    Segmentation of Breast Ultrasound Images using Densely Connected Deep Convolutional Neural Network and Attention Gates,

    N. Thirusangu and M. Almekkawy, “Segmentation of Breast Ultrasound Images using Densely Connected Deep Convolutional Neural Network and Attention Gates,” in 2021 IEEE UFFC Latin America Ultrasonics Symposium (LAUS). Virtual Event: IEEE, 2021, pp. 1–4

    2021

  13. [21]

    Segmentation of Lesions from Breast Ultrasound Images using Deep Convolutional Neural Network,

    N. Thirusangu, “Segmentation of Lesions from Breast Ultrasound Images using Deep Convolutional Neural Network,” 2021, Electronic Thesis/Dissertation, Pennsylvania State University. Available: https:// etda.libraries.psu.edu/catalog/26490nxt359

    2021

  14. [22]

    Deep learning- based automatic pipeline for quantitative assessment of thigh muscle morphology and fatty infiltration,

    S. Gaj, B. L. Eck, D. Xie, R. Lartey, C. Lo, W. Zaylor, M. Yang, K. Nakamura, C. S. Winalski, K. P. Spindler, and X. Li, “Deep learning- based automatic pipeline for quantitative assessment of thigh muscle morphology and fatty infiltration,” Magn. Reson. Med. , vol. 89, no. 6,...

    2023

  15. [23]

    Breast cancer histopatholog- ical image classification using convolutional neural networks with small SE-ResNet module,

    Y . Jiang, L. Chen, H. Zhang, and X. Xiao, “Breast cancer histopatholog- ical image classification using convolutional neural networks with small SE-ResNet module,” PLoS ONE, vol. 14, no. 3, p. e0214587, 2019

    2019

  16. [24]

    Residual and plain convolutional neural networks for 3D brain MRI classification,

    S. Korolev, A. Safiullin, M. Belyaev, and Y . Dodonova, “Residual and plain convolutional neural networks for 3D brain MRI classification,” in 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017). Melbourne, Australia: IEEE, 2017, pp. 835–838

    2017

  17. [25]

    Anatomy-specific classification of medical im- ages using deep convolutional nets,

    H. R. Roth, C. T. Lee, H.-C. Shin, A. Seff, L. Kim, J. Yao, L. Lu, and R. M. Summers, “Anatomy-specific classification of medical im- ages using deep convolutional nets,” in 2015 IEEE 12th International Symposium on Biomedical Imaging (ISBI). New York, NY , USA: IEEE, 2015, pp...

    2015

  18. [26]

    Image Classification Algorithm Based on Improved AlexNet in Cloud Computing Environment,

    Y . Lu, “Image Classification Algorithm Based on Improved AlexNet in Cloud Computing Environment,” in2020 IEEE International Conference on Industrial Application of Artificial Intelligence (IAAI) . Harbin, China: IEEE, 2020, pp. 250–253

    2020

  19. [27]

    The Combine of GLCM and Group, Focuses on the Grayscale of Medical Images,

    Z. Zheng, C. Fan, C. Wang, M. Wang, L. Zhang, X. He, and X. He, “The Combine of GLCM and Group, Focuses on the Grayscale of Medical Images,” in Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) , vol. 2024. Orlando, FL, USA: IEEE, J...

    2024

  20. [28]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . Las Vegas, NV , USA: IEEE, 2016, pp. 770–778

    2016

  21. [29]

    Densely Connected Convolutional Networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely Connected Convolutional Networks,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA, USA: IEEE Computer Society, Jul 2017, pp. 2261–2269, ISSN: 1063-

    2017

  22. [30]

    An Image is Worth 16x16 Words: Trans- formers 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, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale,” in International Conference on Learni...

    2021

  23. [31]

    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,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) . Montreal, QC, Canada: IEEE, Oct 2021, pp. 10 0...

    2021

  24. [32]

    Mamba: Linear-Time Sequence Modeling with Se- lective State Spaces,

    A. Gu and T. Dao, “Mamba: Linear-Time Sequence Modeling with Se- lective State Spaces,” 2023, preprint at https://arxiv.org/abs/2312.00752

    2023 arXiv

  25. [33]

    VMamba: Visual State Space Model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, J. Jiao, and Y . Liu, “VMamba: Visual State Space Model,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS), 2024, https://openreview.net/forum?id=ZgtLQQR1K7

    2024

  26. [34]

    Dual path net- works,

    Y . Chen, J. Li, H. Xiao, X. Jin, S. Yan, and J. Feng, “Dual path net- works,” in Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17) . Red Hook, NY, USA: Curran Associates Inc., 2017, pp. 4470–4478, ISBN: 9781510860964

    2017

  27. [35]

    Attentional Feature Fusion,

    Y . Dai, F. Gieseke, S. Oehmcke, Y . Wu, and K. Barnard, “Attentional Feature Fusion,” in 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). Waikoloa, HI, USA: IEEE, 2021, pp. 3559– 3568

    2021

  28. [36]

    Explainability and Artificial Intelligence in Medicine,

    S. Reddy, “Explainability and Artificial Intelligence in Medicine,” The Lancet Digital Health , vol. 4, no. 4, pp. e214–e215, 2022

    2022

  29. [37]

    Why Should I Trust You?’: Explaining the Predictions of Any Classifier,

    M. T. Ribeiro, S. Singh, and C. Guestrin, “Why Should I Trust You?’: Explaining the Predictions of Any Classifier,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’16) . New York, NY, USA: Association for Computing Ma...

    2016

  30. [38]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Proceedings of the 31st International Confer- 8 IEEE TRANSACTIONS AND JOURNALS TEMPLATE ence on Neural Information Processing Systems (NIPS’17) . Red Hook, NY, USA: Curran Associates Inc.,...

    2017

  31. [39]

    Learn- ing Deep Features for Discriminative Localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learn- ing Deep Features for Discriminative Localization,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . Los Alamitos, CA, USA: IEEE Computer Society, Jun 2016, pp. 2921–2929, ISSN: 106...

    2016

  32. [40]

    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,” in 2017 IEEE International Conference on Computer Vision (ICCV) . Venice, Italy: IEEE, 2017, pp. 618–626

    2017

  33. [41]

    Implementing Explainable AI in Healthcare: Techniques for Interpretable Machine Learning Models in Clinical Decision- Making,

    G. Arjunan, “Implementing Explainable AI in Healthcare: Techniques for Interpretable Machine Learning Models in Clinical Decision- Making,” International Journal of Scientific Research and Management (IJSRM), vol. 9, no. 5, pp. 597–603, May 2021

    2021

  34. [42]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17) . Red Hook, NY, USA: Curran Associates Inc....

    2017

  35. [43]

    Squeeze-and-Excitation Networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-Excitation Networks,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Salt Lake City, UT, USA: IEEE, 2018, pp. 7132–7141

    2018

  36. [44]

    Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation,

    K. Cho, B. van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation,” inProceed- ings of the 2014 Conference on Empirical Methods in Natural Language Pro...

    2014

  37. [45]

    Internimage: Exploring large-scale vision foundation models with deformable convolutions,

    W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Li et al., “Internimage: Exploring large-scale vision foundation models with deformable convolutions,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 14...

    2023

  38. [46]

    Deit iii: Revenge of the vit,

    H. Touvron, M. Cord, and H. Jegou, “Deit iii: Revenge of the vit,” arXiv preprint arXiv:2204.07118, 2022

    2022 arXiv

  39. [47]

    MambaVision: A hybrid mamba- transformer vision backbone,

    A. Hatamizadeh and J. Kautz, “MambaVision: A hybrid mamba- transformer vision backbone,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2025

    2025

  40. [48]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . New Orleans, LA, USA: IEEE, 2022, pp. 11 966–11 976

    2022

  41. [49]

    Brain Tumor (MRI Scans),

    rm1000, “Brain Tumor (MRI Scans),” 2024, Kaggle. Available at https://www.kaggle.com/datasets/rm1000/brain-tumor-mri-scans. Ac- cessed: Oct. 26, 2024

    2024

  42. [50]

    Breast Cancer,

    HAYDER, “Breast Cancer,” 2024, Kaggle. Available at https://www. kaggle.com/datasets/hayder17/breast-cancer-detection. Accessed: Oct. 26, 2024

    2024

  43. [51]

    Labeled Optical Coherence Tomography (OCT) and Chest X-Ray Images for Classification,

    D. Kermany, K. Zhang, and M. Goldbaum, “Labeled Optical Coherence Tomography (OCT) and Chest X-Ray Images for Classification,” 2018, Mendeley Data, V2, https://doi.org/10.17632/rscbjbr9sj.2

    2018 doi

  44. [52]

    Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning,

    D. S. Kermany, M. Goldbaum, W. Cai, C. C. S. Valentim, H. Liang, S. L. Baxter, A. McKeown, G. Yang, X. Wu, F. Yan, J. Dong, M. K. Prasadha, J. Pei, M. Y . L. Ting, J. Zhu, C. Li, S. Hewett, J. Dong, I. Ziyar, A. Shi, R. Zhang, L. Zheng, R. Hou, W. Shi, X. Fu, Y . Duan, V . A. ...

    2018

  45. [6919]

    https://doi.ieeecomputersociety.org/10.1109/CVPR.2017.243

    2017 doi

Pith tools

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