Pith. sign in

REVIEW 3 major objections 49 references

PulmoSight-XAI: An Explainable Multi-View Attention Ensemble with Gradient Boosting Meta-Learning for Multi-Label Chest X-Ray Classification

T0 review · 3 major / 0 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read A hierarchical multi-view CNN ensemble with gradient-boosting meta-learning reaches macro AUROC 0.9319 frontal and 0.9154 lateral on 14 chest X-ray pathologies.

desk verdict Solid within-dataset multi-view CXR ensemble with real engineering detail; the SOTA claim is overstated because Table 6 is cross-dataset and the authors already admit it is uncontrolled. read the letter →

arxiv 2607.04478 v1 pith:LN2PZMQV submitted 2026-07-05 cs.CV cs.AIcs.LG

Moshiur Rahman , Shafqat Alam , Tasnia Binte Mamun This is my paper
classification cs.CVcs.AIcs.LG
keywords chestX-raymulti-labelclassificationmulti-viewensembleCBAMasymmetriclossgradientboostingmeta-learningexplainableAIAUROC
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

Chest X-ray reading is hard for machines because many diseases can appear together, rare findings are scarce, and standard networks throw away the fine local texture radiologists use. This paper claims that training five different CNNs separately on frontal and lateral views, replacing global average pooling with multi-scale features refined by attention, optimizing with a hybrid loss that down-weights easy negatives and focuses hard rare classes, then feeding test-time-augmented predictions and uncertainty into gradient-boosting meta-learners, produces markedly stronger multi-label discrimination. On a large CheXpert-style set the stacked system reports macro-average AUROC of 0.9319 for frontal and 0.9154 for lateral images, above the roughly 0.89–0.91 band of earlier 14-class systems. Seven post-hoc attribution methods are used to show that the decisions land on anatomically plausible regions. A sympathetic reader would care because the design turns architectural diversity and view specialisation into a transparent, high-accuracy computer-aided diagnosis pipeline rather than another single-backbone score.

What carries the argument

Hierarchical multi-view attention ensemble with gradient-boosting meta-learning: view-specific Level-0 multi-scale CBAM fusion heads, hybrid ASL+Adaptive Focal objective, TTA/uncertainty features into Level-1 XGBoost/LightGBM/CatBoost, then Level-2 stacking and optimised alpha blending.

What would settle it

Re-train and evaluate the identical pipeline on the public CheXpert v1.0 official test set (or MIMIC-CXR) under the same 14-label protocol; if macro AUROC falls back into or below the 0.89–0.91 range of prior baselines, or if Level-2 stacking fails to beat simple probability averaging on that set, the central performance claim does not hold.

Watch

Extended reading notes

Core claim

Independently training an ensemble of five complementary CNNs on frontal versus lateral radiographs, fusing multi-scale intermediate maps with CBAM instead of global average pooling, training under a hybrid Asymmetric plus Adaptive Focal loss, and stacking Level-1 gradient-boosting meta-learners with Level-2 alpha blending yields state-of-the-art macro AUROCs of 0.9319 (frontal) and 0.9154 (lateral) across 14 thoracic pathologies, with attributions that concentrate on clinically meaningful anatomy.

Load-bearing premise

That strong scores on one proprietary CheXpert-style competition split can be read as proof of superiority over published systems that were measured on a different public dataset with different prevalence and acquisition conditions.

Editorial extensions

If this is right

  • View-stratified training becomes a practical default for multi-view chest X-ray CAD rather than pooling frontal and lateral images into one network.
  • Replacing global average pooling with multi-scale CBAM fusion can be expected to lift rare-class sensitivity without sacrificing high-contrast findings such as cardiomegaly or support devices.
  • Structured gradient-boosting meta-learners that consume cross-model uncertainty can replace naive probability averaging in multi-architecture medical ensembles.
  • Seven-method post-hoc attribution becomes a standard reporting layer to verify that high AUROC coincides with anatomically coherent localisation.

Reading between the lines

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

  • If the meta-learners mainly exploit disagreement patterns among the five CNNs, similar stacking may transfer to other multi-label medical imaging tasks that already possess diverse pretrained backbones.
  • The residual weakness on long-tail diffuse findings (lung lesion, pleural other, pneumonia) suggests that weakly supervised spatial losses or synthetic rare-class generation would be the next highest-leverage addition.
  • Because the framework is heavy, a distilled single student network that preserves most of the ensemble AUROC would determine whether the approach can leave the research cluster and enter resource-limited clinics.
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

3 major / 0 minor

Summary. The paper proposes PulmoSight-XAI, a hierarchical multi-view ensemble for multi-label classification of 14 thoracic pathologies on a CheXpert-style Grand X-ray Slam Division-B dataset. Five complementary CNNs (InceptionV3, ConvNeXtV2-Tiny, DenseNet201, EfficientNet-B5, ResNeXt-101) are trained separately on frontal and lateral views with multi-scale feature readout, selective CBAM on deep stages, and a hybrid Asymmetric + Adaptive Focal loss. Level-0 outputs are enriched with TTA means/stds and cross-model statistics, then aggregated by Level-1 gradient boosters (XGBoost, LightGBM, CatBoost) and Level-2 stacking/alpha blending. Reported macro AUROCs are 0.9319 (frontal) and 0.9154 (lateral). Seven post-hoc attribution methods are used to argue anatomical plausibility.

Significance. If the within-dataset gains hold under external validation, the work is a solid systems contribution: view-stratified specialization, multi-scale CBAM fusion instead of GAP, a hybrid imbalance-aware loss, and structured gradient-boosting meta-learning with explicit uncertainty features form a coherent, well-engineered pipeline. Tables 3–5 provide thorough per-backbone and ensemble evidence that the meta-learner improves its constituents on both views, and the qualitative XAI suite is broader than typical CXR papers. The main limitation on significance is that the SOTA framing relative to published CheXpert methods is not controlled, so the paper’s primary external claim is weaker than its internal engineering results.

major comments (3)
  1. Abstract, contribution list, §4, Table 6, and Conclusion frame macro-AUROC 0.9319/0.9154 as state-of-the-art relative to prior 14-class thoracic classifiers (range ~0.89–0.91). Table 6 places Grand X-ray Slam numbers next to SSGE, GCF-Net, U-Zeros, cheXGCN, CvTGNet, and MXA numbers from public CheXpert v1.0. §4.1 correctly states that prevalence, annotation protocol, and acquisition differ, that significance testing is impossible, and that baselines are only “contextual references.” The SOTA language should be removed or strictly limited to within-dataset comparisons (Tables 4–5); otherwise the central superiority claim is not supported by matched evidence.
  2. §3.2–3.4 and §4 report ensemble gains over five Level-0 backbones and TTA effects, but there is no controlled ablation of the load-bearing design choices on the same split: multi-scale fusion vs. standard GAP, CBAM vs. no attention, hybrid ASL+Adaptive Focal vs. ASL or BCE alone, and Level-1/2 meta-learning vs. simple probability averaging. Without these, it is unclear which components drive the macro-AUROC lift in Table 5 versus architectural diversity alone.
  3. §4.2 and Figures 2–3 present seven post-hoc attribution methods as demonstrating “strong anatomical consistency and clinically meaningful decision localization,” yet §5 acknowledges the absence of quantitative localization metrics (IoU, pointing game) against ground-truth masks. Qualitative heatmaps alone cannot support the clinical-localization claim at the strength stated in the Abstract and Conclusion; either add quantitative evaluation or soften the claim to qualitative consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical ML pipeline with held-out test metrics; validation-tuned thresholds/alpha and hybrid loss are standard fitting, not definitional reduction of the reported AUROCs.

full rationale

This is an empirical multi-label classification paper whose central claims are measured macro-AUROCs (0.9319 frontal / 0.9154 lateral) on a held-out 10% test split of the Grand X-ray Slam Division-B dataset (Tables 4–5, §4). Level-0 CNNs are trained with a fixed hybrid ASL+Adaptive Focal objective (Eqs. 9–11) whose hyperparameters (γpos=1, γneg=4, Γc=3.5/2.0, α=0.25, ε=0.1) are stated a priori, not fitted to the target AUROC. Level-1 boosters use class-weighting from training counts (Eq. 12). Level-2 alpha blending (Eq. 14) and per-class Youden thresholds (Eq. 19) are optimized by grid search; the paper explicitly places threshold selection on the validation set, and the same practice is the natural reading for α. Reported numbers are therefore ordinary out-of-sample evaluation after validation tuning, not quantities forced by construction from the inputs. There are no self-definitional loops, no uniqueness theorems imported from the authors, no ansatz smuggled via self-citation, and no renaming of a known result as a derivation. Cross-dataset Table 6 comparisons are uncontrolled (as the authors themselves note in §4.1) but that is a validity/overclaim issue, not circularity of the derivation chain. The pipeline is self-contained against its own held-out split; score 0 is the correct non-finding.

Assumptions & free parameters 8 free parameters · 6 assumptions · 3 invented entities

The central performance claim rests on standard CNN transfer learning, a proprietary labeled competition corpus, many hand-chosen loss and ensemble hyperparameters, and the modeling choice that view-stratified multi-scale fusion plus boosting meta-learners is the right inductive bias. No new physical entity is postulated; the 'invented' pieces are the hybrid loss recipe and the engineered meta-feature/stacking pipeline. The largest external dependency is treating CheXpert-style auto-labels and a single stratified split as sufficient for SOTA clinical CAD claims.

free parameters (8)
  • ASL focusing exponents γ_pos, γ_neg = γ_pos=1.0, γ_neg=4.0
    Set to 1.0 and 4.0 without ablation of alternatives; control easy-negative suppression and thus rare-class gradients.
  • Adaptive Focal class exponents Γ_c and α = Γ_hard=3.5, Γ_base=2.0, α=0.25
    Hard classes get Γ_c=3.5, others 2.0, α=0.25; hand-assigned difficulty groups drive the hybrid objective.
  • Hybrid loss mixing weights = 0.5 / 0.5
    Equal 0.5 ASL + 0.5 Adaptive Focal chosen by design; directly shapes the reported AUROCs.
  • Label smoothing ε = 0.1
    ε=0.1 in Adaptive Focal targets; standard but free.
  • Per-class Level-2 alpha blend α = class-specific in [0, 0.55]
    Grid search α∈[0,0.55] step 0.025 maximizing validation macro AUROC per class (Eq. 14).
  • Per-class decision thresholds t_c = class-specific
    Youden-index grid search on validation [0.1,0.9] step 0.02 (Eq. 19); sensitivity/specificity depend on these.
  • MixUp/CutMix batch fraction and temperature T = 40% MixUp/CutMix; T learned
    40% mixed batches; temperature scaled sigmoid optimized jointly—both affect calibration fed to meta-learners.
  • Meta-learner positive class weights w_c,pos = per-class N_neg/N_pos
    Derived from train negative/positive ratios (Eq. 12) but still a design choice that reweights Level-1 objectives.
assumptions (6)
  • domain assumption ImageNet-pretrained CNN stages transfer useful multi-scale features to multi-label CXR after fine-tuning.
    All Level-0 backbones start from ImageNet weights (§3.2); standard medical imaging assumption.
  • ad hoc to paper Independent frontal vs lateral specialization is preferable to pooled multi-view training for all 14 labels.
    Core design choice (§1 contributions, §3 intro); not proven against a strong pooled multi-view baseline in the paper.
  • domain assumption Macro-averaged AUROC with equal class weight is the primary success criterion for clinical multi-label CXR CAD.
    Used throughout evaluation (§3.5, Tables 4–6); can hide prevalence-weighted clinical utility.
  • domain assumption CheXpert-style automated labels on Grand X-ray Slam are sufficiently accurate for ranking models.
    Dataset §2; uncertain-label handling from CheXpert literature is not re-audited here.
  • ad hoc to paper Post-hoc attribution agreement with expected anatomy validates clinical decision localization without ground-truth masks.
    §4.2 uses seven methods qualitatively; authors later note missing IoU/pointing-game metrics (§5).
  • standard math Standard multilabel classification math (sigmoid heads, BCE-style losses, AUROC) applies without formal proof.
    Eqs. 9–19 are conventional empirical risk and metric definitions.
invented entities (3)
  • Hybrid ASL + Adaptive Focal objective with hard-class Γ schedule
    purpose: Jointly address global positive-negative imbalance and rare-class difficulty variance.
    Presented as novel hybrid objective (§3.3.1); combination recipe is paper-specific though both losses exist.
  • Hierarchical TTA/uncertainty meta-feature space for Level-1 boosters + Level-2 stacking/alpha blend
    purpose: Replace probability averaging with structured consensus and per-class blending.
    §3.4 defines the engineered tabular features and two-level ensemble; no external validation of this exact feature schema.
  • PulmoSight-XAI multi-view multi-scale CBAM fusion pipeline
    purpose: End-to-end CAD system for 14-pathology multi-label CXR with explainability.
    System name and full architecture assembly; performance is only evidenced on one competition dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PulmoSight-XAI: An Explainable Multi-View Attention Ensemble with Gradient Boosting Meta-Learning for Multi-Label Chest X-Ray Classification." pith.science (2026). https://pith.science/paper/LN2PZMQV

@misc{pith2026260704478,
  author       = {Pith},
  title        = {Pith review of: PulmoSight-XAI: An Explainable Multi-View Attention Ensemble with Gradient Boosting Meta-Learning for Multi-Label Chest X-Ray Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LN2PZMQV}},
  note         = {Machine review of arXiv:2607.04478}
}
read the original abstract

Automated chest X-ray classification remains challenging due to severe class imbalance, co-occurring pathologies, and the loss of localized features in conventional architectures. To address these, we propose an explainable hierarchical multi-view ensemble framework for the robust classification of 14 thoracic pathologies. The framework employs view-specific training by independently modeling frontal and lateral radiographs using an ensemble of five complementary convolutional neural networks. Replacing global average pooling, a multi-scale feature fusion strategy augmented with Convolutional Block Attention Modules (CBAM) preserves fine-grained intermediate representations while emphasizing high-level pathology-specific semantic features. To mitigate positive-negative imbalance and varying inter-class difficulty, models are optimized using a novel hybrid objective combining Asymmetric Loss with Adaptive Focal Loss. Beyond simple probability averaging, the framework incorporates a hierarchical meta-learning strategy where test-time augmentation (TTA) predictions and cross-model uncertainty measures are integrated into Level-1 gradient-boosting meta-learners (XGBoost, LightGBM, and CatBoost), followed by Level-2 stacking with optimized alpha blending. Evaluated on a large-scale CheXpert-style dataset, the framework achieves state-of-the-art macro-average AUROC scores of 0.9319 for frontal and 0.9154 for lateral radiographs. Furthermore, comprehensive explainability analysis using seven post-hoc attribution techniques demonstrates strong anatomical consistency and clinically meaningful decision localization. By integrating architectural diversity, multi-scale attention, hierarchical meta-learning, and rigorous explainability, the proposed framework provides a transparent, highly accurate, and clinically practical computer-aided diagnosis system for thoracic disease classification.

Figures

Figures reproduced from arXiv: 2607.04478 by the authors.

Figure 1
Figure 1. End-to-end overview of the proposed computational pipeline. The diagram illustrates the data prepara [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. XAI heatmap visualizations of frontal chest X-rays obtained using DenseNet-201. Each row represents a [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. XAI heatmap visualisations of lateral chest X-rays. Each row represents a case with a confirmed positive [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 6 linked inside Pith

  1. [1]

    Global Burden of Dis- ease Study 2021 estimates: implications for health policy and research,

    Z. J. Ward and S. J. Goldie, “Global Burden of Dis- ease Study 2021 estimates: implications for health policy and research,”The Lancet, vol. 403, pp. 1958– 1959, May 2024

    2021

  2. [2]

    ChestX-ray8: Hospital- scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases,

    X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “ChestX-ray8: Hospital- scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases,”arXiv preprint arXiv:1705.02315, 2017

    arXiv 2017

  3. [3]

    Global Strategy on Human Resources for Health: Workforce 2030-A Five-Year Check-In,

    M. McIsaac, J. Buchan, A. Abu-Agla, R. Kawar, and J. Campbell, “Global Strategy on Human Resources for Health: Workforce 2030-A Five-Year Check-In,” Human Resources for Health, vol. 22, p. 68, Oct. 2024

  4. [4]

    Error and discrepancy in radiology: in- evitable or avoidable?,

    A. P. Brady, “Error and discrepancy in radiology: in- evitable or avoidable?,”Insights into Imaging, vol. 8, pp. 171–182, Feb. 2017

    2017

  5. [5]

    A survey on deep learning in medical image analysis,

    G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Se- tio, F. Ciompi, M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, and C. I. Sánchez, “A survey on deep learning in medical image analysis,”Medical Image Analysis, vol. 42, pp. 60–88, Dec. 2017

    2017

  6. [6]

    MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-textreports,

    A. E. W. Johnson, T. J. Pollard, S. J. Berkowitz, N. R. Greenbaum, M. P. Lungren, C.-y. Deng, R. G. Mark, and S. Horng, “MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-textreports,”Scientific Data, vol.6, p.317, Dec. 2019

    2019

  7. [7]

    MSA-Net: multi-scale attention-based DenseNet for multi-label chest X-ray image classification,

    C. Chen, N. A. Mat Isa, X. Liu, J. Ding, and L. Lu, “MSA-Net: multi-scale attention-based DenseNet for multi-label chest X-ray image classification,” Biomedical Signal Processing and Control, vol. 113, p. 109069, Mar. 2026

    2026

  8. [8]

    Style-KD: Class- imbalanced medical image classification via style knowledge distillation,

    I. Park, W. H. Kim, and J. Ryu, “Style-KD: Class- imbalanced medical image classification via style knowledge distillation,”Biomedical Signal Process- ing and Control, vol. 91, p. 105928, May 2024

    2024

Show all 49 references
  1. [9]

    CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison,

    J. Irvin, P. Rajpurkar, M. Ko, Y. Yu, S. Ciurea- Ilcus, C. Chute, H. Marklund, B. Haghgoo, R. Ball, K. Shpanskaya, J. Seekins, D. A. Mong, S. S. Hal- abi, J. K. Sandberg, R. Jones, D. B. Larson, C. P. Langlotz, B. N. Patel, M. P. Lungren, and A. Y. Ng, “CheXpert: A Large Chest...

    2019

  2. [10]

    CheXNet: Radiologist-Level Pneumonia Detection on Chest X- Rays with Deep Learning,

    P. Rajpurkar, J. Irvin, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding, A. Bagul, C. Langlotz, K. Sh- panskaya, M. P. Lungren, and A. Y. Ng, “CheXNet: Radiologist-Level Pneumonia Detection on Chest X- Rays with Deep Learning,” 2017

    2017

  3. [11]

    Comparison of Deep Learning Approaches for Multi-Label Chest X-Ray Classifica- tion,

    I. M. Baltruschat, H. Nickisch, M. Grass, T. Knopp, and A. Saalbach, “Comparison of Deep Learning Approaches for Multi-Label Chest X-Ray Classifica- tion,” 2018

    2018

  4. [12]

    Learning to diagnose from scratch by exploiting dependencies among la- bels,

    L. Yao, E. Poblenz, D. Dagunts, B. Covington, D. Bernard, and K. Lyman, “Learning to diagnose from scratch by exploiting dependencies among la- bels,” 2017

    2017

  5. [13]

    TResNet: High Per- formance GPU-Dedicated Architecture,

    T. Ridnik, H. Lawen, A. Noy, E. B. Baruch, G. Sharir, and I. Friedman, “TResNet: High Per- formance GPU-Dedicated Architecture,” 2020

    2020

  6. [14]

    Multi-label chest X-ray image classification via category-wise residual atten- tion learning,

    Q. Guan and Y. Huang, “Multi-label chest X-ray image classification via category-wise residual atten- tion learning,”Pattern Recognition Letters, vol. 130, pp. 259–266, Feb. 2020

    2020

  7. [15]

    CBAM: Convolutional Block Attention Module,

    S.Woo, J.Park, J.-Y.Lee, andI.S.Kweon, “CBAM: Convolutional Block Attention Module,” 2018

    2018

  8. [16]

    HydraViT: Adaptive multi-branch transformer for multi-label disease classification from Chest X-ray images,

    Ş. Öztürk, M. Y. Turalı, and T. Çukur, “HydraViT: Adaptive multi-branch transformer for multi-label disease classification from Chest X-ray images,” Biomedical Signal Processing and Control, vol. 100, p. 106959, Feb. 2025

    2025

  9. [17]

    Beyond Conven- tional Transformers: The Medical X-ray Atten- tion (MXA) Block for Improved Multi-Label Di- agnosis Using Knowledge Distillation,

    A. Rand and H. Ibrahim, “Beyond Conven- tional Transformers: The Medical X-ray Atten- tion (MXA) Block for Improved Multi-Label Di- agnosis Using Knowledge Distillation,” May 2025. arXiv:2504.02277 [cs]. 15

    2025 arXiv

  10. [18]

    Anoptimizedtransformermodel for efficient detection of thoracic diseases in chest X-rays with multi-scale feature fusion,

    S.YuandP.Zhou, “Anoptimizedtransformermodel for efficient detection of thoracic diseases in chest X-rays with multi-scale feature fusion,”PLOS One, vol. 20, p. e0323239, May 2025

    2025

  11. [19]

    Multi- Label Disease Detection in Chest X-Ray Imaging Us- ing a Fine-Tuned ConvNeXtV2 with a Customized Classifier,

    K. Xiong, Y. Tu, X. Rao, X. Zou, and Y. Du, “Multi- Label Disease Detection in Chest X-Ray Imaging Us- ing a Fine-Tuned ConvNeXtV2 with a Customized Classifier,”Informatics, vol. 12, p. 80, Aug. 2025

    2025

  12. [20]

    Asymmetric Loss For Multi-Label Classification,

    E. Ben-Baruch, T. Ridnik, N. Zamir, A. Noy, I. Friedman, M. Protter, and L. Zelnik-Manor, “Asymmetric Loss For Multi-Label Classification,” 2020

    2020

  13. [21]

    CvTGNet: A Novel Framework for Chest X- Ray Multi-label Classification,

    Y. Lu, Y. Hu, L. Li, Z. Xu, H. Liu, H. Liang, and X. Fu, “CvTGNet: A Novel Framework for Chest X- Ray Multi-label Classification,” inProceedings of the 21st ACM International Conference on Computing Frontiers, (Ischia Italy), pp. 12–20, ACM, May 2024

    2024

  14. [22]

    Clinically-Inspired Hierarchical Multi-Label Classification of Chest X-rays with a Penalty-Based Loss Function,

    M. Asadi, K. Sodoké, I. J. Gerard, and M. Kersten- Oertel, “Clinically-Inspired Hierarchical Multi-Label Classification of Chest X-rays with a Penalty-Based Loss Function,” Feb. 2025. arXiv:2502.03591 [cs]

    2025 arXiv

  15. [23]

    Enhancing Multi-Label Chest X-Ray Clas- sification Using an Improved Ranking Loss,

    M. S. Hanif, M. Bilal, A. H. Alsaggaf, and U. M. Al- Saggaf, “Enhancing Multi-Label Chest X-Ray Clas- sification Using an Improved Ranking Loss,”Bio- engineering, vol. 12, p. 593, May 2025

    2025

  16. [24]

    Aleatoric uncertainty estima- tion with test-time augmentation for medical image segmentation with convolutional neural networks,

    G. Wang, W. Li, M. Aertsen, J. Deprest, S. Ourselin, and T. Vercauteren, “Aleatoric uncertainty estima- tion with test-time augmentation for medical image segmentation with convolutional neural networks,” Neurocomputing, vol. 338, pp. 34–45, Apr. 2019

    2019

  17. [25]

    Label Co-Occurrence Learning With Graph Convolutional Networks for Multi-Label Chest X-Ray Image Clas- sification,

    B. Chen, J. Li, G. Lu, H. Yu, and D. Zhang, “Label Co-Occurrence Learning With Graph Convolutional Networks for Multi-Label Chest X-Ray Image Clas- sification,”IEEE Journal of Biomedical and Health Informatics, vol. 24, pp. 2292–2302, Aug. 2020

    2020

  18. [26]

    Multi-Label Chest X-Ray Image Classification via Semantic Similarity Graph Embedding,

    B. Chen, Z. Zhang, Y. Li, G. Lu, and D. Zhang, “Multi-Label Chest X-Ray Image Classification via Semantic Similarity Graph Embedding,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, pp. 2455–2468, Apr. 2022

    2022

  19. [27]

    Multi-label chest X-ray image classification based on graph convolutional networks and multi-modal fusion,

    J.Sun, J.Hu, X.Wu, Z.Xu, Y.Wang, andY.Zhang, “Multi-label chest X-ray image classification based on graph convolutional networks and multi-modal fusion,”Biomedical Signal Processing and Control, vol. 119, p. 109930, June 2026

    2026

  20. [28]

    BB-GCN: A Bi-modal Bridged Graph Convolu- tional Network for Multi-label Chest X-Ray Recog- nition,

    G. Wang, P. Wang, J. Cong, K. Liu, and B. Wei, “BB-GCN: A Bi-modal Bridged Graph Convolu- tional Network for Multi-label Chest X-Ray Recog- nition,” 2023

    2023

  21. [29]

    Distilling Label Co-Occurrence For Chest X-Ray Image Classification,

    Y. Ding, “Distilling Label Co-Occurrence For Chest X-Ray Image Classification,” in2025 10th Interna- tional Conference on Intelligent Computing and Sig- nal Processing (ICSP), (Xi’an, China), pp. 320–327, IEEE, May 2025

    2025

  22. [30]

    Multi-Label Chest X-Ray Image Classification With Single Positive Labels,

    J. Xiao, S. Li, T. Lin, J. Zhu, X. Yuan, D. D. Feng, and B. Sheng, “Multi-Label Chest X-Ray Image Classification With Single Positive Labels,”IEEE Transactions on Medical Imaging, vol. 43, pp. 4404– 4418, Dec. 2024

    2024

  23. [31]

    Multi-Label General- ized Zero Shot Chest X-Ray Classification by Com- bining Image-Text Information With Feature Disen- tanglement,

    D. Mahapatra, A. Jimeno Yepes, B. Bozorgtabar, S. Roy, Z. Ge, and M. Reyes, “Multi-Label General- ized Zero Shot Chest X-Ray Classification by Com- bining Image-Text Information With Feature Disen- tanglement,”IEEE Transactions on Medical Imag- ing, vol. 44, pp. 31–43, Jan. 2025

    2025

  24. [32]

    Multilabel Chest X-Ray Image Classifica- tion via Category Disentangled Causal Learning,

    Q. Li, M. Liu, R. Chang, W. Nie, S. Bai, and A. Liu, “Multilabel Chest X-Ray Image Classifica- tion via Category Disentangled Causal Learning,” IEEE Transactions on Artificial Intelligence, vol. 7, pp. 1048–1061, Feb. 2026

    2026

  25. [33]

    Grand x-ray slam: Divi- sion b

    G. Dhanjal, S. Sammari, F. Ben Amor, and M. Aissa, “Grand x-ray slam: Divi- sion b.”https://kaggle.com/competitions/ grand-xray-slam-division-b, 2025. Kaggle

    2025

  26. [34]

    mixup: Beyond Empirical Risk Minimization,

    H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez- Paz, “mixup: Beyond Empirical Risk Minimization,” Apr. 2018. arXiv:1710.09412 [cs]

    2018 arXiv

  27. [35]

    CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features,

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo, “CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features,” 2019

    2019

  28. [36]

    Grad-CAM: Visual Expla- nations from Deep Networks via Gradient-based Lo- calization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual Expla- nations from Deep Networks via Gradient-based Lo- calization,”arXiv preprint arXiv:1610.02391, 2016

    2016 arXiv

  29. [37]

    Grad-CAM++: General- ized Gradient-Based Visual Explanations for Deep Convolutional Networks,

    A. Chattopadhay, A. Sarkar, P. Howlader, and V. N. Balasubramanian, “Grad-CAM++: General- ized Gradient-Based Visual Explanations for Deep Convolutional Networks,” in2018 IEEE Winter Conference on Applications of Computer Vision (WACV), (Lake Tahoe, NV), pp. 839–847, IEEE, Mar. 2018

    2018

  30. [38]

    StrivingforSimplicity: TheAllCon- volutional Net,

    J. T. Springenberg, A. Dosovitskiy, T. Brox, and M.Riedmiller, “StrivingforSimplicity: TheAllCon- volutional Net,” Apr. 2015. arXiv:1412.6806 [cs]

    2015 arXiv

  31. [39]

    Axiomatic Attribution for Deep Networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic Attribution for Deep Networks,” 2017

    2017

  32. [40]

    "Why Should I Trust You?

    M. T. Ribeiro, S. Singh, and C. Guestrin, “"Why Should I Trust You?": Explaining the Predictions of Any Classifier,” 2016. 16

    2016

  33. [41]

    Visualizing and Under- standing Convolutional Networks,

    M. D. Zeiler and R. Fergus, “Visualizing and Under- standing Convolutional Networks,” 2013

    2013

  34. [42]

    A Unified Approach to Interpreting Model Predictions,

    S. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” 2017

    2017

  35. [43]

    Measuring Domain Shift for Deep Learning in Histopathology,

    K. Stacke, G. Eilertsen, J. Unger, and C. Lund- strom, “Measuring Domain Shift for Deep Learning in Histopathology,”IEEE Journal of Biomedical and Health Informatics, vol. 25, pp. 325–336, Feb. 2021

    2021

  36. [44]

    PadChest: A large chest x-ray image dataset with multi-label annotated reports,

    A. Bustos, A. Pertusa, J.-M. Salinas, and M. De La Iglesia-Vayá, “PadChest: A large chest x-ray image dataset with multi-label annotated reports,”Medical Image Analysis, vol. 66, p. 101797, Dec. 2020

    2020

  37. [45]

    RoentGen: Vision-Language Foun- dation Model for Chest X-ray Generation,

    P. Chambon, C. Bluethgen, J.-B. Delbrouck, R. Van der Sluijs, M. Połacin, J. M. Z. Chaves, T. M. Abraham, S. Purohit, C. P. Langlotz, and A. Chaudhari, “RoentGen: Vision-Language Foun- dation Model for Chest X-ray Generation,” 2022

    2022

  38. [46]

    Big Self-Supervised Models Advance Medical Image Classification,

    S. Azizi, B. Mustafa, F. Ryan, Z. Beaver, J. Frey- berg, J.Deaton, A.Loh, A.Karthikesalingam, S.Ko- rnblith, T. Chen, V. Natarajan, and M. Norouzi, “Big Self-Supervised Models Advance Medical Image Classification,” 2021

    2021

  39. [47]

    Deep Compres- sion: CompressingDeepNeuralNetworkswithPrun- ing, Trained Quantization and Huffman Coding,

    S. Han, H. Mao, and W. J. Dally, “Deep Compres- sion: CompressingDeepNeuralNetworkswithPrun- ing, Trained Quantization and Huffman Coding,” 2015

    2015

  40. [48]

    Conformal Risk Control,

    A. N. Angelopoulos, S. Bates, A. Fisch, L. Lei, and T. Schuster, “Conformal Risk Control,” 2022

    2022

  41. [49]

    Underdiagno- sis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient popula- tions,

    L. Seyyed-Kalantari, H. Zhang, M. B. A. McDer- mott, I. Y. Chen, and M. Ghassemi, “Underdiagno- sis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient popula- tions,”Nature Medicine, vol. 27, pp. 2176–2182, Dec. 2021. 17

    2021

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.