Pith. sign in

REVIEW 5 major objections 4 minor 71 references

Faithful, Interpretable Chest X-ray Diagnosis with Anti-Aliased B-cos Networks

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper shows that replacing strided convolutions with anti-aliasing pooling (FLCPooling or BlurPool) in B-cos networks removes grid-like artifacts from chest X-ray explanation maps and improves energy-based pointing game scores while…

desk verdict Real artifact fix, but the faithfulness claim is not yet established: the key evidence is a modified EPG that favors B-cos, plus a clear internal contradiction. read the letter →

arxiv 2507.16761 v2 pith:U32QVZ2O submitted 2025-07-22 cs.CV cs.LG

classification cs.CVcs.LG
keywords B-cosnetworkschestX-raydiagnosisanti-aliasingpoolingFLCPoolingBlurPoolenergy-basedpointinggamemulti-labelclassificationinterpretableAI
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

B-cos networks are built to be interpretable: the network itself produces class-specific contribution maps instead of relying on a post-hoc explanation of a black box. Applied to chest X-rays, however, the maps are littered with grid-like aliasing artifacts that come from strided convolutions downsampling without low-pass filtering. This paper argues that replacing those strided convolutions with a stride-1 convolution followed by an anti-aliasing pooling layer—either FLCPooling or BlurPool—removes the artifacts and substantially raises the energy-based pointing game scores of the explanations. The modification also extends to multi-label diagnosis of co-occurring chest conditions, and it keeps classification performance within about one to two percentage points of a standard ResNet-50, so the interpretability gain does not come at a large diagnostic cost. A sympathetic reader would come away with a concrete architectural fix for making inherently interpretable networks clinically usable.

What carries the argument

The machinery is the B-cos contribution map, the class-specific saliency produced by the weight-input alignment operation at every layer, combined with an anti-aliased downsampling path. Where a normal network does feature extraction and resolution reduction in one strided convolution, the modified network applies a stride-1 convolution, a MaxOut activation, and then a pooling layer that low-pass filters before subsampling—FLCPooling in the frequency domain or BlurPool in the spatial domain. The multi-label extension computes contribution maps for every output neuron at once, so co-occurring diseases each receive their own map. The evaluation instrument is the energy-based pointing game borrowed from Score-CAM and adapted by zeroing negative saliency values and adding a recall variant; it measures how much positive explanation energy falls inside the radiologist-annotated bounding box.

What would settle it

Recompute the energy-based pointing game on the released models without the negative-value clipping (the original Score-CAM formula in Eq. 1) and check whether B-cos FLC and B-cos BP still exceed LayerCAM; also run a blinded radiologist preference study on the anti-aliased maps versus LayerCAM. If either check fails, the practical claim that these explanations are more faithful or clinically preferable would not be supported.

Watch

Extended reading notes

Core claim

The central claim is that the artifacts in B-cos chest X-ray explanations are a spectral sampling problem, not a fundamental limit of the B-cos alignment idea. Strided convolutions downsample feature maps by skipping positions, which aliases high-frequency content and propagates grid noise into the contribution maps. Decomposing each strided convolution into a stride-1 convolution followed by FLCPooling or BlurPool separates feature transformation from downsampling and enforces a low-pass filter before resolution is reduced. On the RSNA pneumonia benchmark and the VinBigData multi-label abnormality benchmark, the resulting B-cos FLC and B-cos BP models keep accuracy, recall, precision, F1 and AUC close to the baseline networks while their inherent explanations score higher on the adapted energy-based pointing game than standard B-cos maps and, at their best, than LayerCAM maps.

Load-bearing premise

The load-bearing premise is that the adapted energy-based pointing game defined in Appendix B—which clips all negative saliency values to zero before scoring—measures explanation faithfulness fairly; if negative contributions carry information, the quantitative advantage claimed for the B-cos variants could be partly a metric artifact.

Editorial extensions

If this is right

  • Chest X-ray diagnosis with inherently interpretable B-cos networks can now yield clinically usable explanation maps rather than grid-patterned ones.
  • Anti-aliasing works in both binary pneumonia detection and multi-label abnormality settings, giving one contribution map per possible condition when several pathologies co-occur.
  • Explanation quality, measured by energy-based pointing game, improves by several points over standard B-cos and beats the LayerCAM post-hoc baseline in the configurations reported.
  • The recall-oriented EPG variant reaches 91.56 percent for B-cos BP under augmentation and oversampling, indicating that positive contribution energy is almost entirely inside annotated regions in that setting.
  • Since classification performance stays within roughly one to two percentage points of a standard ResNet-50, adopting the anti-aliased B-cos variant does not force a meaningful diagnostic accuracy trade-off.

Reading between the lines

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

  • The adapted EPG metric is the paper's main quantitative evidence, and it clips negative saliency to zero; a reader should re-run the comparison with the original unclipped Score-CAM pointing game before concluding that B-cos beats LayerCAM on faithfulness, since the clipping may favour maps that contain large negative contributions.
  • The paper reports no human reader study; the ultimate clinical claim—that radiologists find these maps more trustworthy or useful—remains testable but untested.
  • Because the artifact mechanism is architectural rather than specific to X-rays, the same FLCPooling/BlurPool substitution should transfer to other dense medical-imaging tasks, such as mammography or CT, where localized, artifact-free explanations are required.
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

5 major / 4 minor

Summary. The paper proposes two anti-aliasing modifications to B-cos networks, FLCPooling and BlurPool, to remove aliasing artifacts from B-cos explanation maps in chest X-ray diagnosis. The authors evaluate the resulting B-cos_FLC and B-cos_BP models on the RSNA pneumonia binary classification dataset and the VinBigData multi-label dataset, reporting classification metrics and an energy-based pointing game (EPG) metric adapted from Score-CAM. They claim that the anti-aliased variants preserve diagnostic performance while providing artifact-free and more faithful explanations, outperforming LayerCAM and even CheXNet in some comparisons.

Significance. If the central claim holds, the paper addresses a real limitation of B-cos networks in medical imaging: the grid artifacts caused by strided convolutions in explanation maps. The practical evaluation on two medical datasets, including a multi-label setting, is a useful contribution, and the code is promised publicly. However, the load-bearing quantitative evidence for improved faithfulness rests on a modified EPG metric whose validity is not established, and the manuscript contains internal inconsistencies and unsupported comparative claims. The paper's central conclusion is therefore plausible but not yet convincingly demonstrated.

major comments (5)
  1. [Appendix B, Eqs. (2)-(3)] The modified EPG metric is load-bearing for the main faithfulness claim, but its validity is not established. Clipping all negative saliency values to zero (Eq. 2) and the recall variant (Eq. 3) can systematically favor B-cos maps, which contain substantial negative contributions, over LayerCAM maps, which are already ReLU-positive. The paper provides no unclipped EPG comparison, no control experiments (e.g., randomized or Gaussian-smoothed saliency maps), and no justification that this modification preserves the intended notion of localization faithfulness rather than merely a sign-balance property inside the bounding box.
  2. [Section 4.2.1 / Discussion] The sentence 'at threshold 0, all ResNet50 variants achieve recall EPG scores exceeding 99.4%' directly contradicts Table 5, where the t=0 recall EPG values range from 62.70 to 91.56. This inconsistency suggests either a missing table, a mislabeled threshold, or an error in the text. As written, it undermines confidence in the quantitative reporting and must be corrected or reconciled with Table 5.
  3. [Table 7 and Section 4.2.1] The claim that the anti-aliased variants 'consistently achieve the highest EPG scores' is contradicted by Table 7: on the augmented VinBigData setting, standard B-cos achieves EPG 30.62±0.06, while B-cosBP achieves 29.65±0.44 and B-cosFLC is not reported for that configuration. The paper should either qualify the consistency claim, explain why standard B-cos outperforms the anti-aliased variant in this setting, or report the missing FLC result.
  4. [Section 4.2.1 / Table 6] The comparison to CheXNet is cross-dataset and not controlled. CheXNet was trained on ChestX-ray14, not VinBigData, and the paper does not report CheXNet's performance on the same split or with the same preprocessing. The statement that B-cos-based models 'outperform ... CheXNet' on the multi-label task is therefore unsupported and should be removed or replaced with a properly matched baseline.
  5. [Discussion, Section 5] The statement that 'over 50% of high-confidence activations fall within bounding boxes, which only cover 11.65% of the image area' is unsupported: no table or figure reports this statistic, and it is unclear which threshold, dataset, and model variant it refers to. The paper must either provide the supporting result or remove the claim.
minor comments (4)
  1. [Throughout] The notation is inconsistent: the paper uses B-cos_FLC, B-Cos FLC, B-cosFLC, and B-cos FLC interchangeably, and similarly for B-cosBP; please unify the naming.
  2. [Figure 1] The qualitative comparison in Figure 1 is stated as a table with checkmarks and question marks, but the criteria for 'Faithful', 'Interpretable', and 'Artifact-free' are not defined, making the figure potentially misleading; please clarify the basis for these ratings.
  3. [Appendix C.2] There is duplicated text in the preprocessing description: the paragraph starting 'Our primary augmentation pipelines in light augmentation therefore focused' is repeated verbatim from an earlier paragraph; please remove the duplication.
  4. [Section 4.2.1 and Table 2] Table 2 does not include the anti-aliased variants, yet the text refers to 'the anti-aliased variant B-Cos FLC consistently achieves the highest EPG scores' in this section; please restructure the presentation so the referenced tables directly support the claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper combines independently published components and evaluates them with external metrics; no central claim reduces to a fit or to a self-citation chain.

full rationale

I examined the claimed derivation chain: B-cos networks ([8]) provide contribution maps; anti-aliasing replaces strided convolutions with FLC or BlurPool; multi-label support adopts the framework from [46]; explanation quality is measured by the energy-based pointing game (EPG), an external metric whose details are given in Appendix B. None of the central comparisons reduce by construction. FLC and BlurPool are not re-derived here, and the EPG scores are measurements rather than fitted parameters renamed as predictions. The improvements of B-cos_FLC and B-cos_BP over standard B-cos are computed under the same metric, so the anti-aliasing result is not an artifact of the metric's definition. The self-citations ([2], [3], [22], [23]) supply independently published techniques and empirical aliasing analyses that are checkable outside this paper; they are not invoked as an unverified uniqueness theorem and do not bear the load of the central claim alone. The modified 'recall' EPG formula in Eq. 3 is an evaluation choice; whether it is a valid faithfulness proxy is a correctness concern, not a circularity concern. The apparent inconsistency between the sentence 'at threshold 0, all ResNet50 variants achieve recall EPG scores exceeding 99.4%' and the values in Tab. 5 is an internal consistency issue rather than a circular derivation. Overall, the paper is an empirical combination of published components, and the claimed improvements are measured outcomes rather than consequences of how the inputs are defined.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the B-cos faithfulness assumption from prior work and on the adapted EPG metric. The anti-aliasing methods themselves are established, so no new free parameters are introduced by the main mechanism; however, the EPG threshold and augmentation choices are data-dependent.

free parameters (2)
  • EPG saliency threshold t = reported peak around 0.7
    Fig 3 sweeps thresholds and the over 50% claim in Discussion relies on a high threshold; the threshold is not fixed a priori.
  • Augmentation hyperparameters = translation ±32 px, scaling N(0,0.12), rotation ±5°, shearing 2.5°, gamma N(0,0.2)
    Appendix C.2 reports these were selected empirically; they affect the performance tables but not the central anti-aliasing mechanism.
assumptions (3)
  • domain assumption B-cos contribution maps are inherently faithful explanations of the network's decision.
    The paper relies on this to call B-cos maps faithful and to treat EPG scores as measuring faithfulness.
  • domain assumption Aliasing in strided convolutions is the cause of the grid artifacts in B-cos explanation maps.
    Section 3.2 states this via refs [2,3,22] and then assumes anti-aliasing pooling removes the artifacts.
  • ad hoc to paper The customized EPG (positive clipping, Eq. 2) is a valid measure of explanation quality.
    Appendix B introduces modifications to Score-CAM's EPG; the resulting metric is used as the primary evidence of explanation improvement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Faithful, Interpretable Chest X-ray Diagnosis with Anti-Aliased B-cos Networks." pith.science (2026). https://pith.science/paper/U32QVZ2O

@misc{pith2026250716761,
  author       = {Pith},
  title        = {Pith review of: Faithful, Interpretable Chest X-ray Diagnosis with Anti-Aliased B-cos Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U32QVZ2O}},
  note         = {Machine review of arXiv:2507.16761}
}
abstract

Faithfulness and interpretability are essential for deploying deep neural networks (DNNs) in safety-critical domains such as medical imaging. B-cos networks offer a promising solution by replacing standard linear layers with a weight-input alignment mechanism, producing inherently interpretable, class-specific explanations without post-hoc methods. While maintaining diagnostic performance competitive with state-of-the-art DNNs, standard B-cos models suffer from severe aliasing artifacts in their explanation maps, making them unsuitable for clinical use where clarity is essential. In this work, we address these limitations by introducing anti-aliasing strategies using FLCPooling (FLC) and BlurPool (BP) to significantly improve explanation quality. Our experiments on chest X-ray datasets demonstrate that the modified $\text{B-cos}_\text{FLC}$ and $\text{B-cos}_\text{BP}$ preserve strong predictive performance while providing faithful and artifact-free explanations suitable for clinical application in multi-class and multi-label settings. Code available at: GitHub repository (url: https://github.com/mkleinma/B-cos-medical-paper).

Figures

Figures reproduced from arXiv: 2507.16761 by the authors.

Figure 1
Figure 1. An illustration comparing our proposed B-cos [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the anti-aliasing downsampling strategy [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Energy-based pointing game result for precision with [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison of B-cos explanations when using different [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: B-cos Multi-label explanations of a healthy patient for all fourteen abnormalities using standard B-cos models. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Multi-label explanations of an unhealthy patient who has the aortic enlargement and pleural thickening condition. B-cos [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison between GradCAM and B-cosBP network explanations. From left to the right, we can see the original image, B-cosBP explanation, and the GradCAM heatmap over the original image for interpretability purposes. Images taken from the RSNA Pneumonia Detection Datase…
Figure 8
Figure 8. Figure 8: Multi-label explanations of a healthy patient for all fourteen abnormalities with the normal B-cos explanation. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Multi-label explanations of an unhealthy patient that has the aortic enlargement and pleural thickening condition. B-cos [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Multi-label explanations of an unhealthy patient that has the aortic enlargement and pleural thickening condition. B-cos [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 56 canonical work pages

  1. [1]

    Ravi Aggarwal, Viknesh Sounderajah, Guy Martin, Daniel S. W. Ting, Alan Karthikesalingam, Dominic King, Hutan Ashrafian, and Ara Darzi. Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta- analysis. NPJ Digital Medicine, 4:65, Apr. 2021. 2

  2. [2]

    Beware of Aliases -- Signal Preservation is Crucial for Robust Image Restoration

    Shashank Agnihotri, Julia Grabinski, Janis Keuper, and Margret Keuper. Beware of aliases–signal preservation is crucial for robust image restoration. arXiv preprint arXiv:2406.07435, 2024. 4

  3. [3]

    Improving feature stability during upsampling–spectral ar- tifacts and the importance of spatial context

    Shashank Agnihotri, Julia Grabinski, and Margret Keuper. Improving feature stability during upsampling–spectral ar- tifacts and the importance of spatial context. In European Conference on Computer Vision , pages 357–376. Springer,

  4. [4]

    DEHB: Evolutionary Hyberband for Scalable, Robust and Efficient Hyperparameter Optimization

    Noor Awad, Neeratyoy Mallik, and Frank Hutter. DEHB: Evolutionary Hyberband for Scalable, Robust and Efficient Hyperparameter Optimization. In Proceedings of the Thirti- eth International Joint Conference on Artificial Intelligence, pages 2147–2153, Montreal, Canada, Aug. 2021. Interna- tional Joint Conferences on Artificial Intelligence Organiza- tion. 4

  5. [5]

    Interpretation of Chest X-ray: An Illus- trated Companion

    G Balachandran. Interpretation of Chest X-ray: An Illus- trated Companion . Jaypee Brothers Medical Publishers, New Delhi, 1 edition, 2014. 4

  6. [6]

    Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI

    Alejandro Barredo Arrieta, Natalia D ´ıaz-Rodr´ıguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garcia, et al. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion, 58:82–115, June

  7. [7]

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

    Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria de la Iglesia-Vay ´a. PadChest: A large chest x-ray image dataset with multi-label annotated reports. Medical Image Analysis, 66:101797, Dec. 2020. arXiv:1901.07441 [eess]. 3

  8. [8]

    B-cos Net- works: Alignment is All We Need for Interpretability, May

    Moritz B ¨ohle, Mario Fritz, and Bernt Schiele. B-cos Net- works: Alignment is All We Need for Interpretability, May

Show all 71 references
  1. [9]

    Chawla, Aleksandar Lazarevic, Lawrence O

    Nitesh V . Chawla, Aleksandar Lazarevic, Lawrence O. Hall, and Kevin W. Bowyer. SMOTEBoost: Improving Prediction of the Minority Class in Boosting. In Gerhard Goos, Juris Hartmanis, Jan Van Leeuwen, Nada Lavra ˇc, Dragan Gam- berger, Ljup ˇco Todorovski, and Hendrik Blockeel, ...

  2. [10]

    Explainable medical imaging AI needs human-centered design: guidelines and evidence from a sys- tematic review

    Haomin Chen, Catalina Gomez, Chien-Ming Huang, and Mathias Unberath. Explainable medical imaging AI needs human-centered design: guidelines and evidence from a sys- tematic review. npj Digital Medicine, 5(1):1–15, Oct. 2022. Publisher: Nature Publishing Group. 3, 7

  3. [11]

    Receiver op- erating characteristic curve analysis in diagnostic accuracy studies: A guide to interpreting the area under the curve value

    S ¸eref Kerem C ¸ orbacıo˘glu and G ¨okhan Aksel. Receiver op- erating characteristic curve analysis in diagnostic accuracy studies: A guide to interpreting the area under the curve value. Turkish Journal of Emergency Medicine, 23(4):195– 198, Oct. 2023. 13

  4. [12]

    Augmenting Medical Imaging: A Com- prehensive Catalogue of 65 Techniques for Enhanced Data Analysis, Mar

    Manuel Cossio. Augmenting Medical Imaging: A Com- prehensive Catalogue of 65 Techniques for Enhanced Data Analysis, Mar. 2023. arXiv:2303.01178 [eess]. 15

  5. [13]

    Brindle, Lisa M

    David Crosby, Sangeeta Bhatia, Kevin M. Brindle, Lisa M. Coussens, Caroline Dive, Mark Emberton, Sadik Esener, et al. Early detection of cancer. Science (New York, N.Y.), 375(6586):eaay9040, Mar. 2022. 5

  6. [14]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, June 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, June 2021. arXiv:2010.11929 [cs]. 3

  7. [15]

    Deep Learning for Automatic Pneumonia Detec- tion

    Tatiana Gabruseva, Dmytro Poplavskiy, and Alexandr Kalinin. Deep Learning for Automatic Pneumonia Detec- tion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1436– 1443, Seattle, W A, USA, June 2020. IEEE. 3, 5, 14, 15

  8. [16]

    Diagnostic accuracy of point- of-care tests in acute community-acquired lower respira- tory tract infections

    Elisa Gentilotti, Pasquale De Nardo, Eleonora Cremonini, Anna G ´orska, Fulvia Mazzaferri, Lorenzo Maria Canziani, Mona Mustafa Hellou, et al. Diagnostic accuracy of point- of-care tests in acute community-acquired lower respira- tory tract infections. A systematic review and ...

  9. [17]

    Bridging the gap: From post hoc explana- tions to inherently interpretable models for medical imaging

    Shantanu Ghosh, Ke Yu, Forough Arabshahi, and Kayhan Batmanghelich. Bridging the gap: From post hoc explana- tions to inherently interpretable models for medical imaging. In Proceedings of the 3rd Workshop on Interpretable Ma- chine Learning in Healthcare (IMLH) at ICML 2023 ,...

  10. [18]

    Gilpin, David Bau, Ben Z

    Leilani H. Gilpin, David Bau, Ben Z. Yuan, Ayesha Bajwa, Michael Specter, and Lalana Kagal. Explaining Explana- tions: An Overview of Interpretability of Machine Learning, Feb. 2019. arXiv:1806.00069 [cs]. 3

  11. [19]

    Deep Learning: A Visual Approach

    Andrew Glassner. Deep Learning: A Visual Approach . No Starch Press, San Francisco, 2021. 13

  12. [20]

    Gonzalez, Richard E

    Rafael C. Gonzalez, Richard E. Woods, and Barry R. Mas- ters. Digital Image Processing, Third Edition. Journal of Biomedical Optics, 14(2):029901, 2009. 15

  13. [21]

    Deep Learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning . MIT Press, 2016. http : / / www . deeplearningbook.org. 13

  14. [22]

    FrequencyLowCut Pooling - Plug and Play Against Catastrophic Overfitting

    Julia Grabinski, Steffen Jung, Janis Keuper, and Margret Ke- uper. FrequencyLowCut Pooling - Plug and Play Against Catastrophic Overfitting. In Shai Avidan, Gabriel Brostow, Moustapha Ciss ´e, Giovanni Maria Farinella, and Tal Hass- ner, editors, Computer Vision – ECCV 2022 , ...

  15. [23]

    Fix your downsampling ASAP! Be natively more robust via Aliasing and Spectral Artifact free Pooling, July 2023

    Julia Grabinski, Janis Keuper, and Margret Keuper. Fix your downsampling ASAP! Be natively more robust via Aliasing and Spectral Artifact free Pooling, July 2023. arXiv:2307.09804 [cs]. 4

  16. [24]

    Deep Residual Learning for Image Recognition, Dec

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition, Dec. 2015. arXiv:1512.03385 [cs]. 2, 3, 4 9

  17. [25]

    Belal Hossain, S

    Md. Belal Hossain, S. M. Hasan Sazzad Iqbal, Md. Monirul Islam, Md. Nasim Akhtar, and Iqbal H. Sarker. Transfer learning with fine-tuned deep CNN ResNet50 model for clas- sifying COVID-19 from chest X-ray images. Informatics in Medicine Unlocked, 30:100916, Jan. 2022. 2, 4

  18. [26]

    Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco An- dreetto, and Hartwig Adam

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco An- dreetto, and Hartwig Adam. MobileNets: Efficient Convolu- tional Neural Networks for Mobile Vision Applications, Apr

  19. [27]

    Weinberger

    Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kil- ian Q. Weinberger. Densely Connected Convolutional Net- works, Jan. 2018. arXiv:1608.06993 [cs]. 2

  20. [28]

    CheXpert: A Large Chest Radiograph Dataset with Uncer- tainty Labels and Expert Comparison, Jan

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Sil- viana Ciurea-Ilcus, Chris Chute, Henrik Marklund, et al. CheXpert: A Large Chest Radiograph Dataset with Uncer- tainty Labels and Expert Comparison, Jan. 2019. 2, 3

  21. [29]

    LayerCAM: Exploring Hierarchi- cal Class Activation Maps for Localization

    Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. LayerCAM: Exploring Hierarchi- cal Class Activation Maps for Localization. IEEE Transac- tions on Image Processing, 30:5875–5888, 2021. 1, 2, 5

  22. [30]

    Alistair E. W. Johnson, Tom J. Pollard, Seth J. Berkowitz, Nathaniel R. Greenbaum, Matthew P. Lungren, Chih-ying Deng, Roger G. Mark, and Steven Horng. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Scientific Data, 6(1):317...

  23. [31]

    Computer-assisted diagnosis for an early iden- tification of lung cancer in chest X rays

    Judith Juan, Eduard Mons ´o, Carme Lozano, Marta Cuf ´ı, Paula Sub ´ıas-Beltr´an, Laura Ruiz-Dern, Xavier Rafael- Palou, et al. Computer-assisted diagnosis for an early iden- tification of lung cancer in chest X rays. Scientific Reports, 13(1):7720, May 2023. Publisher: Nature...

  24. [32]

    Solving min- imum cost lifted multicut problems by node agglomera- tion

    Amirhossein Kardoost and Margret Keuper. Solving min- imum cost lifted multicut problems by node agglomera- tion. In Asian Conference on Computer Vision, pages 74–89. Springer, 2018. 2

  25. [33]

    Hierarchical markov random fields for mast cell segmentation in electron microscopic record- ings

    Margret Keuper, Thorsten Schmidt, Marta Rodriguez- Franco, Wolfgang Schamel, Thomas Brox, Hans Burkhardt, and Olaf Ronneberger. Hierarchical markov random fields for mast cell segmentation in electron microscopic record- ings. In 2011 IEEE International Symposium on Biomedical...

  26. [34]

    Blind deconvolution of widefield fluorescence micro- scopic data by regularization of the optical transfer function (otf)

    Margret Keuper, Thorsten Schmidt, Maja Temerinac-Ott, Jan Padeken, Patrick Heun, Olaf Ronneberger, and Thomas Brox. Blind deconvolution of widefield fluorescence micro- scopic data by regularization of the optical transfer function (otf). In Proceedings of the IEEE Conference ...

  27. [35]

    A study of cross-validation and bootstrap for accuracy estimation and model selection

    Ron Kohavi. A study of cross-validation and bootstrap for accuracy estimation and model selection. In Proceedings of the 14th international joint conference on Artificial intelli- gence - Volume 2 , IJCAI’95, pages 1137–1143, San Fran- cisco, CA, USA, Aug. 1995. Morgan Kaufman...

  28. [36]

    Pneumonia detection in chest X-ray images using an ensemble of deep learning models

    Rohit Kundu, Ritacheta Das, Zong Woo Geem, Gi-Tae Han, and Ram Sarkar. Pneumonia detection in chest X-ray images using an ensemble of deep learning models. PLOS ONE , 16(9):e0256630, Sept. 2021. Publisher: Public Library of Science. 2, 3, 4, 5, 13

  29. [37]

    Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Ar- naud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen A. W. M. van der Laak, et al. A survey on deep learning in medical image analysis. Medical Image Analysis, 42:60–88, Dec. 2017. 2

  30. [38]

    A ConvNet for the 2020s, Mar

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s, Mar. 2022. arXiv:2201.03545 [cs]. 4

  31. [39]

    To- wards a Guideline for Evaluation Metrics in Medical Image Segmentation, Feb

    Dominik M ¨uller, I ˜naki Soto-Rey, and Frank Kramer. To- wards a Guideline for Evaluation Metrics in Medical Image Segmentation, Feb. 2022. arXiv:2202.05273 [eess]. 2

  32. [40]

    The power of data mining in diagnosis of childhood pneumonia.Journal of The Royal Society Interface , 13(120):20160266, July 2016

    Elina Naydenova, Athanasios Tsanas, Stephen Howie, Cli- ment Casals-Pascual, and Maarten De V os. The power of data mining in diagnosis of childhood pneumonia.Journal of The Royal Society Interface , 13(120):20160266, July 2016. Publisher: Royal Society. 2, 5

  33. [41]

    Nguyen, Khanh Lam, Linh T

    Ha Q. Nguyen, Khanh Lam, Linh T. Le, Hieu H. Pham, Dat Q. Tran, Dung B. Nguyen, Dung D. Le, et al. Vindr- cxr: An open dataset of chest x-rays with radiologist’s anno- tations, 2020. 3, 4, 7

  34. [42]

    Exploring large scale public medical image datasets, July 2019

    Luke Oakden-Rayner. Exploring large scale public medical image datasets, July 2019. arXiv:1907.12720 [eess]. 1

  35. [43]

    IEViT: An enhanced vision transformer architec- ture for chest X-ray image classification

    Gabriel Iluebe Okolo, Stamos Katsigiannis, and Naeem Ramzan. IEViT: An enhanced vision transformer architec- ture for chest X-ray image classification. Computer Methods and Programs in Biomedicine, 226:107141, Nov. 2022. 3

  36. [44]

    Py- Torch: An Imperative Style, High-Performance Deep Learn- ing Library, Dec

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, et al. Py- Torch: An Imperative Style, High-Performance Deep Learn- ing Library, Dec. 2019. arXiv:1912.01703 [cs]. 14, 15

  37. [45]

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

    Pranav Rajpurkar, Jeremy Irvin, Kaylie Zhu, Brandon Yang, Hershel Mehta, Tony Duan, Daisy Ding, et al. CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning, Dec. 2017. arXiv:1711.05225 [cs]. 2, 6, 7, 13, 15

  38. [46]

    Studying how to efficiently and effectively guide models with explanations

    Sukrut Rao, Moritz B ¨ohle, Amin Parchami-Araghi, and Bernt Schiele. Studying how to efficiently and effectively guide models with explanations. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 1922–1933, 2023. 2, 4

  39. [47]

    A multichannel EfficientNet deep learning-based stacking ensemble approach for lung disease detection using chest X-ray images

    Vinayakumar Ravi, Vasundhara Acharya, and Mamoun Alazab. A multichannel EfficientNet deep learning-based stacking ensemble approach for lung disease detection using chest X-ray images. Cluster Computing, 26(2):1181–1203,

  40. [48]

    Detection of Pneumonia from Chest X-ray Images Utilizing MobileNet Model

    Mana Saleh Al Reshan, Kanwarpartap Singh Gill, Vatsala Anand, Sheifali Gupta, Hani Alshahrani, Adel Sulaiman, and Asadullah Shaikh. Detection of Pneumonia from Chest X-ray Images Utilizing MobileNet Model. Healthcare, 11(11):1561, Jan. 2023. Number: 11 Publisher: Multidis- cip...

  41. [49]

    Stop explaining black box machine learn- ing models for high stakes decisions and use interpretable models instead

    Cynthia Rudin. Stop explaining black box machine learn- ing models for high stakes decisions and use interpretable models instead. Nature machine intelligence, 1(5):206–215,

  42. [50]

    Current challenges of implementing artificial intelligence in medical imaging

    Shier Nee Saw and Kwan Hoong Ng. Current challenges of implementing artificial intelligence in medical imaging. Physica Medica: European Journal of Medical Physics , 100:12–17, Aug. 2022. Publisher: Elsevier. 1

  43. [51]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Com- puter Vision, 128(2):336–359, Feb. 2020. arXiv:1610.023...

  44. [52]

    Wu, Safwan S

    George Shih, Carol C. Wu, Safwan S. Halabi, Marc D. Kohli, Luciano M. Prevedello, Tessa S. Cook, Arjun Sharma, et al. Augmenting the National Institutes of Health Chest Radio- graph Dataset with Expert Annotations of Possible Pneumo- nia. Radiology: Artificial Intelligence , 1...

  45. [53]

    Explainable Deep Learning Models in Medi- cal Image Analysis

    Amitojdeep Singh, Sourya Sengupta, and Vasudevan Laksh- minarayanan. Explainable Deep Learning Models in Medi- cal Image Analysis. Journal of Imaging, 6(6):52, June 2020. 3

  46. [54]

    Efficient pneumonia detection using Vision Transform- ers on chest X-rays

    Sukhendra Singh, Manoj Kumar, Abhay Kumar, Biren- dra Kumar Verma, Kumar Abhishek, and Shitharth Selvara- jan. Efficient pneumonia detection using Vision Transform- ers on chest X-rays. Scientific Reports , 14(1):2487, Jan

  47. [55]

    Task driven sensor layouts-joint optimiza- tion of pixel layout and network parameters

    Hendrik Sommerhoff, Shashank Agnihotri, Mohamed Saleh, Michael Moeller, Margret Keuper, Bhaskar Choubey, and Andreas Kolb. Task driven sensor layouts-joint optimiza- tion of pixel layout and network parameters. In 2024 IEEE International Conference on Computational Photography...

  48. [56]

    Attri-net: A globally and lo- cally inherently interpretable model for multi-label classifi- cation using class-specific counterfactuals

    Susu Sun, Stefano Woerner, Andreas Maier, Lisa M Koch, and Christian F Baumgartner. Attri-net: A globally and lo- cally inherently interpretable model for multi-label classifi- cation using class-specific counterfactuals. arXiv preprint arXiv:2406.05477, 2024. 2

  49. [57]

    B-cos aligned transformers learn human-interpretable fea- tures

    Manuel Tran, Amal Lahiani, Yashin Dicente Cid, Melanie Boxberg, Peter Lienemann, Christian Matek, Sophia J Wag- ner, Fabian J Theis, Eldad Klaiman, and Tingying Peng. B-cos aligned transformers learn human-interpretable fea- tures. In International Conference on Medical Image ...

  50. [58]

    Evaluation of post-hoc interpretability methods in time-series classification

    Hugues Turb ´e, Mina Bjelogrlic, Christian Lovis, and Gi- anmarco Mengaldo. Evaluation of post-hoc interpretability methods in time-series classification. Nature Machine Intel- ligence, 5(3):250–260, Mar. 2023. Publisher: Nature Pub- lishing Group. 3

  51. [59]

    Score-CAM: Score-Weighted Visual Explanations for Convolutional Neu- ral Networks

    Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-CAM: Score-Weighted Visual Explanations for Convolutional Neu- ral Networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pag...

  52. [60]

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Moham- madhadi Bagheri, and Ronald M. Summers. ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Com- mon Thorax Diseases. In 2017 IEEE Conference on Com- puter ...

  53. [61]

    Early Convolutions Help Trans- formers See Better, Oct

    Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Doll´ar, and Ross Girshick. Early Convolutions Help Trans- formers See Better, Oct. 2021. arXiv:2106.14881 [cs]. 4, 7

  54. [62]

    Making Convolutional Networks Shift- Invariant Again, June 2019

    Richard Zhang. Making Convolutional Networks Shift- Invariant Again, June 2019. arXiv:1904.11486 [cs]. 2, 3, 4, 7

  55. [63]

    A Comprehensive Survey on Transfer Learning, June 2020

    Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A Comprehensive Survey on Transfer Learning, June 2020. arXiv:1911.02685 [cs]. 3

  56. [64]

    van Leeuwen, and Keelin Murphy

    Erdi C ¸ allı, Ecem Sogancioglu, Bram van Ginneken, Kicky G. van Leeuwen, and Keelin Murphy. Deep learning for chest X-ray analysis: A survey. Medical Image Analysis, 72:102125, Aug. 2021. 1, 2, 8 11 Faithful, Interpretable Chest X-ray Diagnosis with Anti-Aliased B-cos Network...

  57. [69]

    In the formulas this is displayed by only con- sidering Lp which are the positive contributions

    setting all negative values to zero which leads to the fo- cus of the activations which contribute to the class pre- diction. In the formulas this is displayed by only con- sidering Lp which are the positive contributions. This allows for comparability of the B-cos network ene...

  58. [70]

    Conse- quently, this metric should display whether the model sees the values inside bounding box area as primarily contributing to the class prediction

    an adjustment of the metric which divides the sum of the positive values through the sum of the positive and absolute values of the negative contributions. Conse- quently, this metric should display whether the model sees the values inside bounding box area as primarily contri...

  59. [71]

    using a threshold t to the existing metrics above which makes it so only values above t · max(Lc p) are consid- ered. The adjustments of the first two extensions lead to the following formulas where Lp references to the positive val- ues within the contribution map and Ln to t...

  60. [2017]

    arXiv:1704.04861 [cs]. 2

  61. [2019]

    3, 4, 7, 8

    Publisher: Radiological Society of North America. 3, 4, 7, 8

  62. [2022]

    2, 3, 4, 5, 7

    arXiv:2205.10268 [cs]. 2, 3, 4, 5, 7

  63. [2024]

    Publisher: Nature Publishing Group. 3

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.