Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

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

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Chest X-ray model with hierarchy reaches 0.903 AUROC on CheXpert

desk verdict The central penalty in Eq. (2) has zero gradient, so HBCE cannot enforce parent-child consistency as written, and the headline AUROC comes from a test-set sweep. read the letter →

arxiv 2502.03591 v1 pith:ESJHR6Y6 submitted 2025-02-05 cs.CV cs.AIcs.LG

Mehrdad Asadi , Komi Sodoké , Ian J. Gerard , Marta Kersten-Oertel This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords chestX-raymulti-labelclassificationhierarchicallossfunctionCheXpertinterpretabilityAUROCmedicalimaging
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

This paper aims to show that a single chest X-ray classifier, trained once with a hierarchy of clinically grouped diagnoses and a penalty-based loss, can match or beat more complex multi-stage pipelines on the CheXpert benchmark. The authors report a mean AUROC of 0.903 on the official test set (weighted 0.9034), while adding parent-child consistency penalties and uncertainty estimates for interpretability. If the claim holds, it suggests that clinical structure can be injected directly into a loss function without sacrificing accuracy or requiring model ensembles.

What carries the argument

The mechanism is the HBCE loss: standard binary cross-entropy plus a term $\lambda \sum_{p,c} P_{p,c}$, where each penalty is $\text{Penalty}(p,c) \cdot \mathbf{1}\{ y_{\text{pred},p} < 0.5 \text{ and } y_{\text{pred},c} > 0.5 \}$. The penalty is a fixed value $\beta$ or a data-driven ratio $(N_{\text{parent}=0,\text{child}=1}+\epsilon)/(N_{\text{parent}=0}+2\epsilon)$, so the loss is meant to raise the cost of clinically implausible parent-negative/child-positive predictions. The taxonomy (e.g., Pleural Effusion and Edema under 'Fluid Accumulation') defines the parent-child pairs that the penalty acts on, and the 'Uncertain' label captures ambiguous cases.

What would settle it

Run the same training setup with the HBCE penalty term removed (i.e., plain BCE with the same hierarchy and 'Uncertain' label) and compare AUROC and parent-child inconsistency rates; if the numbers match, the penalty plays no role. Additionally, compute the gradient of Eq. (5) with respect to the network outputs and show it is zero for the penalty term outside a measure-zero set, confirming that no learning signal comes from the penalty.

Watch

Extended reading notes

Core claim

The central claim is that the proposed hierarchical binary cross-entropy (HBCE) loss, which adds a penalty when a child label is predicted positive while its parent is predicted negative, yields a model that reaches a mean AUROC of 0.903 on CheXpert. Using a DenseNet121 trained from scratch with fixed or data-driven penalties, the authors report that data-driven penalties at a scale factor of 0.5 give the best high-level performance, and that the hierarchy produces clinically consistent predictions with visual explanations and Monte Carlo dropout uncertainty. On the five common CheXpert pathologies the model achieves 0.892 mean AUROC, with the highest reported scores for Atelectasis (0.879) and Pleural Effusion (0.945) among the compared methods.

Load-bearing premise

The paper assumes the penalty term in the loss actually influences training, but the hard indicator it contains has zero gradient almost everywhere, so under standard backpropagation the penalty cannot enforce parent-child consistency.

Editorial extensions

If this is right

  • A single DenseNet121 trained from scratch with the HBCE loss reaches a mean AUROC of 0.903 on the CheXpert test set, matching or exceeding several more complex multi-stage or graph-based baselines.
  • Data-driven penalties at scale 0.5 produce the best high-level performance (0.904 mean AUROC on the hierarchy), suggesting the conditional-probability weighting is a useful tuning knob.
  • Adding an 'Uncertain' label does not significantly change overall AUROC (p=0.098 for five labels) but yields marginally better scores, which may be clinically useful for flagging ambiguous cases.
  • The single-model, single-run pipeline simplifies deployment relative to multi-stage hierarchical methods, and the Grad-CAM maps provide spatial explanations aligned with clinically expected regions.

Reading between the lines

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

  • Because the indicator function in Eq. (2) has zero gradient almost everywhere, the penalty term as written cannot change network weights during standard backpropagation; any observed AUROC difference likely comes from the hierarchy structure, the extra 'Uncertain' label, or other training changes, not from the penalty's enforcement.
  • A testable extension would be to replace the hard indicator with a soft, differentiable penalty (e.g., based on predicted probabilities) or a straight-through estimator; the authors do not provide code, so the exact implementation is unverified.
  • The data-driven penalty formula resembles a conditional empirical probability of child positives given parent negatives; if instead used as a sample reweighting scheme, it might produce similar gains without the zero-gradient issue.
  • The reported AUROC on high-level categories (0.942 for Abnormal, 0.922 for Fluid Accumulation) suggests the hierarchy itself may be the main contributor; ablating the penalty entirely while keeping the same taxonomy would isolate its contribution.
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 / 4 minor

Summary. The paper proposes a clinically-inspired hierarchical multi-label classification framework for chest X-rays using the CheXpert dataset with VisualCheXbert-derived labels. The method groups 14 original labels into six high-level parent categories (plus a manually defined 'Uncertain' label) and introduces a Hierarchical Binary Cross-Entropy (HBCE) loss that adds a penalty when a child label is predicted positive while its parent is predicted negative. Two penalty schemes are studied: a fixed constant penalty and a data-driven penalty based on empirical label counts, with a scale factor controlling penalty strength. The authors report a mean AUROC of 0.903 on the test set and a 'weighted AUROC of 0.9034' in the conclusion, and claim that the custom loss 'enforces label dependencies' and improves interpretability. The paper also includes Monte Carlo dropout uncertainty estimates, Grad-CAM visualizations, and a comparison with prior CheXpert classification methods.

Significance. If the hierarchical penalty mechanism worked as described, the contribution would be a straightforward, single-model way to inject clinical label hierarchies into multi-label CXR classification, with interpretability aids. The authors also provide measured AUROC results on the CheXpert test set and a comparison table against prior work. However, the central mechanism is invalid as written: the penalty term in Eq. (2) uses a hard indicator whose gradient is zero almost everywhere, so it cannot influence network weights during backpropagation. In addition, the headline AUROC is selected from a test-set sweep over multiple penalty configurations, which contradicts the 'single-model, single-run' claim. Because the core claim of the paper depends on an inoperative loss term, the current manuscript does not support its own conclusions; the idea may be salvageable with a differentiable penalty formulation and a corrected experimental protocol, but the present evidence is not adequate.

major comments (3)
  1. [III-D.2, Eq. (2) and III-E] The hierarchical penalty term as written cannot provide a learning signal during training. In Eq. (2), Pp,c = Penalty(p,c) · 1{ypred,p < 0.5 and ypred,c > 0.5}. The factor Penalty(p,c) is independent of the model outputs: Eq. (3) sets it to a constant β and Eq. (4) computes it from training-set label counts. The only output-dependent part is the indicator function, which is piecewise constant and has zero derivative almost everywhere. Under the standard TensorFlow/Keras backpropagation described in Section III-E, this term contributes exactly zero gradient to the DenseNet121 weights. Consequently, the HBCE loss cannot 'enforce label dependencies' during training as claimed in the abstract and Section III-D. The hierarchy could at most influence checkpoint selection via the validation loss, not the learned model. The authors must either replace the hard indicator with a differentiable surrogate (e.g., a sigmoid relaxation) and state this explicitly in Eq. (2), or provide code demonstrating a straight-through estimator or equivalent gradient-providing implementation. As published, the mechanism is inoperative.
  2. [IV-B, Table II, IV-C] The reported AUROC of 0.903/0.904 is selected as the maximum over no fewer than eight penalty configurations (four data-driven and four fixed scale factors) evaluated on the official test set. The text in Section IV-B states that the data-driven penalty at scale factor 0.5 'achieved the highest mean of AUROC (0.903)' and was then followed. Selecting the best test-set configuration from a sweep inflates the expected performance and makes the 'single-model, single-run training pipeline' claim in the abstract and conclusion misleading. The authors should select hyperparameters on a validation split and report the corresponding test result, or report all configurations with an appropriate multiple-comparison correction.
  3. [Abstract, Conclusion, and Table II] The primary reported metric is inconsistent and undefined. The abstract reports a 'mean AUROC of 0.903,' the conclusion reports a 'weighted AUROC of 0.9034,' and Table II reports the highest mean AUROC over six high-level categories as 0.904 (data-driven, scale 0.5). No definition of 'weighted AUROC' is given anywhere in the paper. The authors should specify exactly which metric is computed (e.g., unweighted mean over the six parent categories), use a single consistent number in the abstract and conclusion, and ensure that the reported figure matches the definition.
minor comments (4)
  1. [III-C] The statement that the model 'was initialized with random weights and trained from scratch, as the domain-specific nature of CXR images often benefits from task-specific training [23]' is not supported by reference [23]: Raghu et al. actually recommend ImageNet pretraining for medical imaging tasks and report that it improves accuracy and reduces training time. The citation should be replaced or the claim revised.
  2. [I and V] The paper claims that 'all code, model configurations, and experiment details are made available,' but no repository URL is provided; the text in Section I only gives the placeholder name 'CIHMLC.' Without a working link or an explicit identifier (e.g., a DOI or GitHub URL), the reproducibility claim cannot be verified.
  3. [IV-B] The paired t-tests comparing data-driven and fixed penalties across multiple labels and scale factors are numerous (at least nine comparisons are described) and are reported without any correction for multiple testing. At the 0.05 significance level, several of the significant p-values could be false positives by chance; the authors should apply a correction or clearly label the comparisons as exploratory.
  4. [Fig. 3 caption] The ground-truth list in the Figure 3 caption includes 'Enlarged Cardiomegaly,' but the corresponding CheXpert label is 'Enlarged Cardiomediastinum.' The caption should be corrected, and the header of the paper also contains a typo ('CLINICALL Y' in the running title).

Circularity Check

1 steps flagged · score 2.0 of 10

Mild circularity: the scale factor λ is chosen by test-set AUROC and the same test-set AUROC is reported as the headline result; the loss derivation itself is not circular.

  1. fitted input called prediction [Section IV-B (Penalty) and Section VI (Conclusion)]
    "Based on these results, we followed the study with the data-driven penalty approach at a scale factor of 0.5 that achieved the highest mean of AUROC (0.903), suggesting a favorable balance of penalties for boosting classifier performance on the high-level categories compared to other configurations."

    The final AUROC is not independent of the hyperparameter search: the paper evaluates data-driven penalties at scale factors 0.3, 0.5, 0.7, and 1.0 on the official test set, selects λ=0.5 because it 'achieved the highest mean of AUROC (0.903)' on that same test set, and then reports 0.903/0.9034 as the model's performance. The headline number is therefore the maximum of a small grid search over the test labels rather than a held-out prediction for the chosen configuration; the fitted scale factor and the reported AUROC come from the same test-set evaluations, so the 'prediction' is statistically forced by the selection rule.

full rationale

The loss derivation (Eqs. 1-5) is self-contained: the penalties are constants computed from training-label counts, and the reported test AUROC is a measured quantity, not a value reconstructed from those counts. The hierarchy is externally motivated by clinical references and clinician feedback rather than by a self-citation chain, and no 'uniqueness' theorem or ansatz is imported from the authors' prior work. The only circularity-like step is hyperparameter selection: the authors evaluate several scale factors on the official CheXpert test set and adopt the best one (λ=0.5), then report that same test-set AUROC (0.903 / 0.9034) as the headline performance, making the final number a test-set-selected maximum rather than an unbiased prediction. This is a mild fitted-input-called-prediction issue and accounts for the low score. Separately, Eq. (2) multiplies the penalty by a hard indicator whose derivative vanishes almost everywhere, which undermines the stated training mechanism; however, that is a correctness/mechanism concern, not a circularity. The Discussion's stated limitations (batch-size sensitivity and label-structure generalizability) are external correctness caveats and do not create a circular derivation. The central AUROC result still has independent empirical content, so the circularity score remains low.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The core method adds no physical entity, but it introduces a custom label hierarchy and an 'Uncertain' class, and it depends on data-driven penalty counts and three tuning constants. Because the hard-indicator penalty has zero gradient, the method's central assumption about learning is unsupported.

free parameters (4)
  • Scale factor lambda (penalty weight) = 0.5 (best of 0.3, 0.5, 0.7, 1.0)
    Controls the strength of the hierarchical penalty in Eq. (5). The paper reports AUROC for all four values on the official test set and then adopts 0.5 because it has the highest mean, so the headline result is the winner of a test-set sweep.
  • Fixed penalty beta = 1
    Constant penalty in Eq. (3) for all fixed-penalty runs. The authors state a value of 1 was selected to 'ensure a fair comparison' of penalty strategies, but no sweep or justification is given in Section IV.
  • Laplace smoothing epsilon = Not reported
    Appears in Eq. (4) to avoid division by zero in the data-driven penalty. The value is never stated, so the exact penalty numbers cannot be reproduced.
  • Data-driven penalty counts, N_parent0_child1 / N_parent0 = Computed from training labels
    These empirical conditional-frequency penalties are inputs to the loss; they are data-derived rather than tuned, but they still determine the penalty magnitudes and are not provided as a table or file.
assumptions (4)
  • domain assumption VisualCheXbert labels are valid ground truth for the CheXpert images and are comparable to official CheXpert U-Zeros labels.
    Section III-B changes the labeler from the standard CheXpert rules to VisualCheXbert, yet Table III compares against methods trained on official U-Zeros labels; no analysis shows the labelers produce equivalent test targets.
  • domain assumption The parent-child taxonomy in Fig. 1 is a clinically meaningful and complete organization of the 14 labels.
    Section III-B constructs parent categories from prior works, Web sources, and one clinician's feedback; the hierarchy itself is not validated externally, and if it is wrong the penalties enforce false dependencies.
  • ad hoc to paper The hard-indicator penalty term in Eq. (2) participates in gradient-based training.
    Equation (5) adds the indicator-multiplied penalty to the loss, but the indicator has zero derivative almost everywhere; the paper relies on this assumption when claiming the loss enforces parent-child consistency.
  • domain assumption Parent label ground truth can be obtained by combining child labels.
    High-level AUROC in Table II evaluates parent classes derived from children; this makes high-level scores partly dependent on the child labels rather than independent clinical outcomes.
invented entities (2)
  • High-level parent labels (Abnormal, Cardiac, Fluid Accumulation, Missing Lung Tissue, Opacity, Other)
    purpose: New output classes for hierarchical supervision and for the high-level AUROC in Table II.
    These labels are constructed by the authors from the 14 CheXpert diagnoses; no external annotations validate the grouping, and the parent ground truth is derived from child labels.
  • Uncertain label
    purpose: Class for images with neither positive findings nor No Finding, to capture ambiguous cases.
    The label is defined from VisualCheXbert outputs in Section III-B, and Section IV-A's t-tests show no significant performance effect, so its clinical value is asserted rather than demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Clinically-Inspired Hierarchical Multi-Label Classification of Chest X-rays with a Penalty-Based Loss Function." pith.science (2026). https://pith.science/paper/ESJHR6Y6

@misc{pith2026250203591,
  author       = {Pith},
  title        = {Pith review of: Clinically-Inspired Hierarchical Multi-Label Classification of Chest X-rays with a Penalty-Based Loss Function},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ESJHR6Y6}},
  note         = {Machine review of arXiv:2502.03591}
}
read the original abstract

In this work, we present a novel approach to multi-label chest X-ray (CXR) image classification that enhances clinical interpretability while maintaining a streamlined, single-model, single-run training pipeline. Leveraging the CheXpert dataset and VisualCheXbert-derived labels, we incorporate hierarchical label groupings to capture clinically meaningful relationships between diagnoses. To achieve this, we designed a custom hierarchical binary cross-entropy (HBCE) loss function that enforces label dependencies using either fixed or data-driven penalty types. Our model achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.903 on the test set. Additionally, we provide visual explanations and uncertainty estimations to further enhance model interpretability. All code, model configurations, and experiment details are made available.

Figures

Figures reproduced from arXiv: 2502.03591 by the authors.

Figure 1
Figure 1. Clinically-Inspired Taxonomy • Improved Interpretability and Clinical Decision Mak￾ing: The hierarchical approach makes the model’s predic￾tions easier to interpret by clinicians [26] as it categorizes pathologies into broader groups (e.g., lung opacities vs. cardiac abnormalities), then refining the diagnosis based on specific child labels (e.g., Cardiomegaly under Car￾diac). Similar to clinical workflows, broad di… view at source ↗
Figure 2
Figure 2. AUROC curves for the Hierarchical training strategy with a Data￾Driven Penalty and a scale factor of 0.5 on the CheXpert dataset. Together, clipping the activations and segmented colormap approach aims to highlight key areas on the CXR where the model “looks” to predict each pathology, making the activation maps more clinically meaningful by limiting visual noise and emphasizing the most diagnostically relevant area… view at source ↗
Figure 3
Figure 3. Activations on the sample test image. Top row: defined high-level and “Uncertain” labels. Middle and bottom rows: original labels in the CheXpert dataset. Activations are normalized and clipped with a 0.5 threshold, and the color map is segmented to enhance both localization and interpretation. In this example, Fluid Accumulation shows high activation over potential fluid-affected areas, consistent with findings in … view at source ↗

Discussion (0). Continue with ORCID 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. PulmoSight-XAI: An Explainable Multi-View Attention Ensemble with Gradient Boosting Meta-Learning for Multi-Label Chest X-Ray Classification

    cs.CV 2026-07 conditional novelty 4.0 of 10

    View-specific multi-scale CBAM CNN ensembles plus hybrid ASL/focal loss and two-level gradient-boosting stacking reach ~0.93/0.92 macro AUROC on a CheXpert-style multi-label CXR dataset.

Reference graph

Works this paper leans on

37 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    Clinical Relevance of the Groupings: The hierarchy’s structure in Fig 1 reflects the relationships between different conditions, making it possible for the model to capture clinical dependencies and associations more effectively. The structure aligns with clinical reasoning as follows: • Grouping Related Conditions: By categorizing labels into clinically ...

    work page

  2. [2]

    BCE measures the difference between the predicted probabilities ypred and the true binary labels ytrue for each label

    Binary Cross-Entropy: At the core of our loss function is the Binary Cross-Entropy (BCE) Loss, which is traditionally used in multi-label classification. BCE measures the difference between the predicted probabilities ypred and the true binary labels ytrue for each label. For L labels and a batch size of B, the BCE loss is given by: LBCE(ytrue, ypred) = −...

    work page

  3. [3]

    Fluid Accumulation

    Hierarchical Penalty for Parent-Child Relationships: Incor- porating a hierarchical structure requires addressing parent- child dependencies in label predictions. For example, if a par- ent label (e.g., “Fluid Accumulation”) is predicted as negative, it is inconsistent for a child label (e.g., “Pleural Effusion”) to be predicted as positive. To enforce su...

    work page

  4. [4]

    Uncertain

    Fixed vs. Data-Driven Penalties: The hierarchical loss function operates in two modes depending on the source of the penalties: • Fixed Penalty: A constant penalty value is assigned to all parent-child inconsistencies. This penalty mode is straightforward and computationally efficient, but it lacks adaptability to data: Penalty(p, c) =β (3) • Data-Driven ...

    work page 2017

  5. [5]

    Causes of Death,

    S. Dattani, F. Spooner, H. Ritchie, and M. Roser, “Causes of Death,” Our World in Data , 2023. Available: https://ourworldindata.org/causes- of-death

    work page 2023

  6. [6]

    Work overload and diagnostic errors in radiol- ogy,

    ¨O. Kasalak, H. Alnahwi, R. Toxopeus, J. P. Pennings, D. Yakar, and T. C. Kwee, “Work overload and diagnostic errors in radiol- ogy,” European Journal of Radiology , vol. 167, p. 111032, 2023. doi: 10.1016/j.ejrad.2023.111032

    work page arXiv 2023

  7. [7]

    Workload for radiologists during on- call hours: dramatic increase in the past 15 years,

    R. J. M. Bruls and R. M. Kwee, “Workload for radiologists during on- call hours: dramatic increase in the past 15 years,” Insights into Imaging, vol. 11, p. 121, 2020. doi: 10.1186/s13244-020-00925-z

    work page doi:10.1186/s13244-020-00925-z 2020

  8. [8]

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

    Irvin, Jeremy, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea- Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison . In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence, 2019

    work page 2019

Show all 37 references
  1. [9]

    VisualCheXbert: Addressing the Discrepancy Between Radiology Report Labels and Image Labels

    Jain, Saahil, Akshay Smit, Steven QH Truong, Chanh DT Nguyen, Minh-Thanh Huynh, Mudit Jain, Victoria A Young, Andrew Y Ng, Matthew P Lungren, and Pranav Rajpurkar. VisualCheXbert: Addressing the Discrepancy Between Radiology Report Labels and Image Labels . arXiv preprint arXi...

    2021 arXiv

  2. [10]

    Ng, and Matthew P

    Smit, Akshay, Saahil Jain, Pranav Rajpurkar, Anuj Pareek, Andrew Y . Ng, and Matthew P. Lungren. CheXbert: Combining Automatic Labelers and Expert Annotations for Accurate Radiology Report Labeling Using BERT. arXiv preprint arXiv:2004.09167, 2020

    2004 arXiv

  3. [11]

    Weinberger.Densely connected convolutional networks

    Huang, Gao, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger.Densely connected convolutional networks. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition*, 2017

    2017

  4. [12]

    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 Proceedings of the IEEE Interna- tional Conference on Computer Vision (ICCV) , 2017, pp. 618–626. doi10.1109/ICCV .2017.74

    2017

  5. [13]

    G. Wang, P. Wang, and B. Wei, ”Multi-label local awareness and global co-occurrence priori learning improve chest X-ray classification,” Mul- timedia Systems, vol. 30, no. 132, pp. 1-12, 2024. doi: 10.1007/s00530- 024-01321-z

    2024 doi

  6. [14]

    Zhang, W

    K. Zhang, W. Liang, P. Cao, X. Liu, J. Yang, and O. Zaiane, ”Label cor- relation guided discriminative label feature learning for multi-label chest image classification,” Computer Methods and Programs in Biomedicine, vol. 245, Article 108032, 2024. doi: 10.1016/j.cmpb.2024.108032

    2024

  7. [15]

    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,” Proceed- ings of the 21st ACM International Conference on Computing Frontiers (CF ’24), May 2024, pp. 1-9. doi: 10.1145/3649153.3649216

    2024

  8. [16]

    Z. Liu, Y . Cheng, and S. Tamura, ”Multi-Label Local to Global Learning: A Novel Learning Paradigm for Chest X-Ray Abnormality Classifica- tion,” IEEE Journal of Biomedical and Health Informatics , vol. 27, no. 9, pp. 4409-4420, Sep. 2023. doi: 10.1109/JBHI.2023.3281466

    2023

  9. [17]

    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, no. 4, pp. 2455-2467, Apr. 2022. doi: 10.1109/TCSVT.2021.3079900

    2022

  10. [18]

    A. Ke, W. Ellsworth, O. Banerjee, A. Y . Ng, and P. Rajpurkar, ”CheX- transfer: Performance and Parameter Efficiency of ImageNet Models for Chest X-Ray Interpretation,” in Proceedings of the ACM Conference on Health, Inference, and Learning (CHIL) , Virtual Event, 2021, pp. 2097-

    2021

  11. [19]

    H. H. Pham, T. T. Le, D. Q. Tran, D. T. Ngo, and H. Q. Nguyen, ”Interpreting chest X-rays via CNNs that exploit hierarchical disease dependencies and uncertainty labels,” Neurocomputing, vol. 437, pp. 186–194, 2021

    2021

  12. [20]

    Huang, Q.-J

    G.-H. Huang, Q.-J. Fu, M.-Z. Gu, N.-H. Lu, K.-Y . Liu, and T.-B. Chen, ”Deep Transfer Learning for the Multilabel Classification of Chest X-ray Images,” Diagnostics, vol. 12, no. 6, p. 1457, 2022. doi: 10.3390/diagnostics12061457

    2022 doi

  13. [21]

    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,” Proceedings of the IEEE Conference on Computer Vision and Pattern R...

    2017

  14. [22]

    Hierarchical Multi-Label Classification Networks,

    J. Wehrmann, R. Cerri, and R. C. Barros, “Hierarchical Multi-Label Classification Networks,” in Proceedings of the 35th International Conference on Machine Learning , Stockholm, Sweden: PMLR, vol. 80, 2018, pp. 5075–5084

    2018

  15. [23]

    H. Chen, S. Miao, D. Xu, G. D. Hager, and A. P. Harrison, ”Deep hierarchical multi-label classification applied to chest X-ray abnormality taxonomies,” Medical Image Analysis , vol. 66, 2020, Art. no. 101811

    2020

  16. [24]

    N., ”A disciplined approach to neural network hyper- parameters: Part 1–learning rate, batch size, momentum, and weight decay,” *arXiv preprint arXiv:1803.09820*, 2018

    Smith, L. N., ”A disciplined approach to neural network hyper- parameters: Part 1–learning rate, batch size, momentum, and weight decay,” *arXiv preprint arXiv:1803.09820*, 2018

    2018 arXiv

  17. [25]

    A. E. W. Johnson, T. J. Pollard, S. J. Berkowitz, et al., ”MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports,” Scientific Data , vol. 6, Article 317, 2019. [Online]. Available: https://doi.org/10.1038/s41597-019-0322-0

    2019 doi

  18. [26]

    A. P. Aberle, M. D. Adams, C. J. Berg, et al., ”The Prostate, Lung, Col- orectal and Ovarian (PLCO) Cancer Screening Randomized Controlled Trial,” Journal of the National Cancer Institute , vol. 92, no. 18, pp. 1536-1544, 2000. [Online]. Available: https://cdas.cancer.gov/plco/

    2000

  19. [27]

    Bustos, A

    A. Bustos, A. Pertusa, J. Salinas, and M. de la Iglesia-Vay ´a, ”PadChest: A large chest x-ray image dataset with multi-label annotated reports,” Medical Image Analysis , vol. 66, 101797, 2020. [Online]. Available: https://doi.org/10.1016/j.media.2020.101797

    2020

  20. [28]

    Transfusion: Understanding transfer learning for medical imaging,

    M. Raghu, C. Zhang, J. Kleinberg, and S. Bengio, “Transfusion: Understanding transfer learning for medical imaging,” in Proceedings of the 33rd International Conference on Neural Information Processing Systems (NeurIPS), 2019, pp. 3342–3352

    2019

  21. [29]

    Pulmonary Edema,

    Cleveland Clinic, “Pulmonary Edema,” [Online]. Available: https://my.clevelandclinic.org/health/diseases/ 17841-pulmonary-edema

  22. [30]

    AI lifts nonradiologists’ x-ray interpre- tation skill levels,

    W. Morton, “AI lifts nonradiologists’ x-ray interpre- tation skill levels,” AuntMinnie.com, Oct. 25, 2024. [Online]. Available: https://www.auntminnie.com/ clinical-news/digital-x-ray/article/15706835/ ai-lifts-nonradiologists-xray-interpretation-skill-levels

    2024

  23. [31]

    From extreme multi-label to multi-class: A hierarchical approach for automated ICD-10 coding using phrase-level attention,

    C. Sen, B. Ye, J. Aslam, and A. Tahmasebi, “From extreme multi-label to multi-class: A hierarchical approach for automated ICD-10 coding using phrase-level attention,” arXiv preprint arXiv:2102.09136 , 2021. Available: https://arxiv.org/abs/2102.09136

    2021 arXiv

  24. [32]

    Pleural Effusion: Symptoms, Causes, Treatments,

    WebMD, “Pleural Effusion: Symptoms, Causes, Treatments,” [Online]. Available: https://www.webmd.com/lung/ pleural-effusion-symptoms-causes-treatments

  25. [33]

    Pneumonia,

    Mayo Clinic, “Pneumonia,” [Online]. Available: https://www. mayoclinic.org/diseases-conditions/pneumonia/ symptoms-causes/syc-20354204

  26. [35]

    Consolidation,

    Radiopaedia.org, “Consolidation,” [Online]. Available: https:// radiopaedia.org/articles/consolidation

  27. [36]

    Chest X-Ray - Lung Disease,

    R. Smithuis, “Chest X-Ray - Lung Disease,” The Radiology Assistant,

  28. [2014]

    Available: https://radiologyassistant.nl/ chest/chest-x-ray/lung-disease

    [Online]. Available: https://radiologyassistant.nl/ chest/chest-x-ray/lung-disease

  29. [2106]

    doi: 10.1145/3450439.3451867

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.