Pith. sign in

REVIEW 5 major objections 4 minor 42 references

From Classification to Localization and Clinical Validation: Large-Scale Development of a Deep Learning System for Thoracic Disease Detection on Chest Radiographs in Thailand

T0 review · 5 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a chest X-ray AI trained on 874,858 Thai images reaches near-ceiling accuracy at its home hospital, keeps most of that performance across 13 other hospitals, and localizes lesions so well that radiologists agree with

desk verdict A substantial applied CXR validation with real multi-site numbers, but the missing patient-level split definition and proprietary labeling pipeline mean the headline AUROC should be treated as unverified until the authors close those gaps. read the letter →

arxiv 2607.09305 v2 pith:3EWITLMD submitted 2026-07-10 cs.CV cs.LG

Isarun Chamveha , Tretap Promwiset , Napat Wanchaitanawong , Trongtum Tongdee , Pairash Saiviroonporn , Warasinee Chaisangmongkon This is my paper · ORCID
classification cs.CVcs.LG
keywords chestradiographthoracicdiseasedetectiondeeplearningweaklysupervisedlocalizationclinicalvalidationgeneralizationsecondreaderThailand
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 reports a deep learning system for reading chest X-rays, built and tested entirely on data from Thailand. The authors claim that a single model can both classify nine thoracic conditions and produce heatmaps showing where findings are, and that on a large held-out test set it reaches a mean AUROC of 0.994, with 0.970 across 13 other hospitals. They further report that five thoracic radiologists agreed with the model's classification in 93.6% of cases and with its localization in 94.7%, and rated it highly usable. The point, if true, is that locally trained AI can serve as a trustworthy second reader in settings with few radiologists, and that localization matters as much as classification for clinical acceptance.

What carries the argument

The load-bearing mechanism is a single multi-task network that outputs, for each of nine conditions, both a classification score and a heatmap. It uses a densely connected convolutional encoder with attention layers that compare different image regions, a decoder that restores spatial detail via skip connections, and a probabilistic class-activation-map pooling layer that produces sharper heatmaps while preserving classification. Because the same representation drives both outputs, localization is learned from image-level labels alone, with no pixel annotations during training; pixel annotations are used only to evaluate the resulting heatmaps. This coupling is what lets the paper report cla

What would settle it

Compare patient identifiers between the training and test sets: if any patient appears in both, recompute the in-domain AUROC on a disjoint subset. Separately, take a random sample of training images, have radiologists re-label them from scratch, and compare with the natural-language-derived labels; a high disagreement rate would show the reported metrics inherit label noise.

Watch

Extended reading notes

Core claim

The paper's central claim is that a locally developed chest X-ray model can combine high in-domain discrimination with cross-site generalization and clinically acceptable localization. Trained on 874,858 frontal radiographs with paired reports from one large Bangkok hospital, the system achieved a mean AUROC of 0.994 on 19,871 radiologist-verified held-out cases, and 0.970 on 5,992 independent cases from 13 hospitals. On 4,549 annotated cases, it localized lesions with a mean lesion-localization fraction of 77.9% at 0.59 false localizations per image. In a usability study, five thoracic radiologists showed 93.6% classification concordance and 94.7% localization concordance, with a standard u

Load-bearing premise

The central claims rest on the held-out test sets being free of patients who also appear in training; the paper does not state whether the split was by patient or by image, and if patients overlap, the headline 0.994 AUROC would be inflated.

Editorial extensions

If this is right

  • If the model is deployed as a second reader, it can flag abnormal studies and point radiologists to the region of interest, potentially shortening reading time.
  • The small drop from 0.994 to 0.970 AUROC across 13 hospitals indicates that a model trained at one site can retain ranking ability elsewhere without retraining.
  • Fixed operating points chosen in-domain lose sensitivity on external data, especially for diffuse findings; per-site threshold adjustment may be needed.
  • A concordance rate near 94% with radiologists suggests the system's outputs are clinically actionable rather than merely statistically discriminative.
  • The combination of classification and localization in a single model avoids the need for separate detection pipelines, simplifying deployment.

Reading between the lines

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

  • Because the paper does not state whether the held-out split was at patient level, a reader should treat the in-domain 0.994 as an upper bound until a patient-disjoint split is confirmed; this is a testable extension, not a claim the paper makes.
  • The training labels come from a proprietary natural-language model reviewed by radiologists; the error rate of that labeler is not quantified, so the true ceiling of the system is unknown. An audit that compares fully manual labels to the pipeline's labels on a random sample would bound this.
  • The usability numbers rest on five radiologists; a larger multi-reader study across more sites would give a more stable estimate of the 89 score and concordance rates.
  • The fixed-threshold sensitivity drop on external data suggests that deployment would benefit from site-specific threshold calibration, which could raise sensitivity without sacrificing AUROC.
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 presents Inspectra CXR version 5, a deep learning system for multi-label thoracic disease classification and weakly supervised lesion localization on frontal chest radiographs. The model combines a DenseNet-121 backbone with Attend-and-Compare Modules and a PCAM aggregation layer, and was trained on 874,858 frontal CXRs from Siriraj Hospital. The authors report a held-out in-domain mean AUROC of 0.994 on 19,871 radiologist-verified cases, a multi-site generalization mean AUROC of 0.970 on 5,992 cases from 13 Thai hospitals, a mean lesion-localization fraction of 77.9% on 4,549 annotated cases, radiologist concordance of 93.6% (classification) and 94.7% (localization), and a System Usability Scale score of 89.

Significance. If the results hold, this would be a practically valuable validation of a locally developed CXR AI system for a Southeast Asian population, combining in-domain accuracy, cross-site transfer, localization, and usability in a single deployment-oriented study. The scale of the training cohort and the inclusion of an external 13-hospital test set are notable strengths, as is the pixel-annotated localization evaluation. However, the central validity of the headline accuracy numbers is not yet established: the split unit used for the held-out sets is not specified, the training-label generation pipeline is proprietary and unquantified, and the concordance metric counts a substantial number of misses as concordant. These issues are load-bearing for the paper's main claims.

major comments (5)
  1. [§3.1, §3.3] The held-out test set is not demonstrated to be patient-disjoint from training. Section 3.1 excludes only "duplicate examinations from the same patient on the same date," and Section 3.3 states the 19,871 images were "held out" without specifying the split unit. If the same patient appears in training and test on different dates, patient-specific anatomy or follow-up patterns can inflate the in-domain AUROC. Please report the number of overlapping patients (or perform a patient-disjoint split) for Dataset-A, and similarly for Dataset-B and Dataset-C.
  2. [§3.2] All training labels originate from a proprietary natural-language labeling model whose accuracy is never reported. The statement that radiologists "reviewed and corrected" these labels does not quantify the fraction audited, the correction rate, or inter-reader agreement. Because the model is trained on these labels, every downstream metric is conditional on an unmeasured label-error process. Please report a random audit of the training labels, the NLP model's performance, or an equivalent assessment.
  3. [§3.7, Tables 4–5] The concordance metric counts "Add" as concordant, yet for localization "Add" is defined as "The heatmap misses some lesions" and for classification as the system under-calling relative to the radiologist. In Table 9, 753 of 4,549 localization cases (16.6%) and 856 of 19,871 classification cases are "Add." Treating misses as concordant systematically inflates the 93.60%/94.72% headline concordance. Please report the four categories separately and recompute concordance with Add excluded or as a sensitivity analysis.
  4. [Tables 6–8] All metrics are point estimates without confidence intervals, despite small positive counts (e.g., pneumothorax n=64 in Dataset-A; atelectasis n=118 in Dataset-B) and near-ceiling AUROCs. Report 95% CIs (e.g., bootstrap or DeLong) for AUROC, sensitivity, specificity, and LLF. Without intervals, the 0.994 vs 0.970 gap and per-condition differences cannot be properly assessed.
  5. [§3.3, §4.4] The relationship among Dataset-A, Dataset-B, and Dataset-C is not stated. If Dataset-C is a subset of Dataset-A, the localization and classification results are not independent; if disjoint, the total case counts need clarification. In addition, the five radiologists in the usability study may include the same readers who verified labels or created annotations; if so, the concordance analysis is partially a consistency check of the labelers. Please clarify reader overlap and case-set overlap.
minor comments (4)
  1. [§5 (Discussion)] The sentence "Per-condition ROC analysis from case-level prediction scores is planned as future work" conflicts with Tables 6 and 7, which already report per-condition AUROC. Please clarify what additional ROC analysis is intended (e.g., ROC curves with confidence intervals).
  2. [§3.4] "Cut-Noise" and the "proprietary image-processing quality algorithms" are not described beyond a single sentence. Please provide a specification or a citation so that the method is reproducible.
  3. [§3.5, §4.2] The operating point selection is described only as balancing sensitivity and specificity at a "clinically appropriate PPV." Please state the actual thresholds applied to Dataset-B and whether these are the same thresholds selected on the validation set.
  4. [Table 3] The label set is the nine conditions, but Table 3 also lists "Normal" counts. Please state whether Normal is a separate class used in the test evaluation and how normal-only cases are defined relative to the nine-condition labels.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported AUROC, localization, and usability results are empirical measurements against external radiologist-verified ground truth and reader ratings, not derivations from the model's own outputs.

full rationale

The paper claims no first-principles derivation; its central results are empirical evaluations against external benchmarks. Training labels come from radiologist reports with radiologist review; test labels and localization annotations are explicitly radiologist-verified and are not produced by the evaluated model. AUROC on Datasets A and B is computed against these external labels, LLF is computed against radiologist pixel annotations, and SUS comes from independent radiologist ratings, so the headline numbers are not fitted parameters renamed as predictions. The self-citation [14] motivating local adaptation is background context, not a load-bearing step: the current claims stand on the reported held-out and multi-site test sets. No equation or definition reduces a reported quantity to the training labels or to another claimed result. Remaining concerns such as the unspecified patient-level train/test split, the unquantified proprietary NLP labeler, and the decision to count 'Add' as concordant are external-validity limitations rather than circularity, and do not warrant raising the circularity score.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The central claims depend on at least five under-specified settings (validation-selected thresholds, undisclosed per-class weights, undefined Cut-Noise parameters, the IoU 0.5 convention, and a hand-chosen training schedule) and four domain assumptions, three of which involve proprietary, undescribed components (NLP labeler, quality filters, Cut-Noise). No confidence intervals, ablation studies, or inter-reader statistics are provided, and the patient-level independence of the test sets is asserted but not demonstrated. The reported numbers are internally consistent point estimates whose uncertainty and external validity cannot be assessed from the paper alone.

free parameters (5)
  • Per-condition operating points = not reported
    Sensitivity/specificity figures in Tables 6-7 are evaluated at thresholds selected on the validation set (Section 3.6); the reported operating performance is conditional on these choices.
  • Per-class loss weights = not reported
    Training uses 'per-class weighting to address class imbalance' (Section 3.5) but the weights are never given; they shape scores, thresholds, and therefore all reported metrics.
  • Cut-Noise hyperparameters = not reported
    Cut-Noise is credited for heatmap quality (Section 3.4) with no definition of its parameters or an ablation, so localization metrics are conditional on an unspecified setting.
  • IoU threshold for localization metrics = 0.5
    LLF and NLF are reported at an IoU threshold of 0.5 (Section 3.7); LLF would change with the threshold.
  • Training schedule = 10 epochs, LR 1e-4, decay 0.1 at epochs 7/9, batch 10
    Hand-chosen hyperparameters (Section 3.5); the reported performance depends on them and no sensitivity analysis is provided.
assumptions (5)
  • domain assumption ImageNet-pretrained DenseNet-121 features transfer to frontal chest radiographs
    Model is initialized with ImageNet weights (Sections 3.4-3.5); no evidence is given that this pretraining is suitable for CXR, though it is standard practice.
  • domain assumption Free-text radiologist reports, processed by an unspecified proprietary NLP model, yield sufficiently accurate training labels
    Section 3.2: initial labels come from a proprietary natural-language labeling model; its accuracy is never quantified, yet it defines the training targets for all 874,858 images.
  • domain assumption The proprietary image-quality filters correctly identify diagnostic-quality images and do not bias disease prevalence
    Sections 3.1-3.2: images are admitted/excluded by proprietary algorithms; exclusion rates and their correlation with disease are unreported.
  • domain assumption Radiologist-verified labels on the test sets are definitive ground truth
    Section 3.3: Dataset-A/B labels were 'comprehensively reviewed and confirmed' by chest radiologists; no inter-reader agreement or adjudication statistics are reported.
  • domain assumption The development and test partitions do not share patients
    Section 3.3 states the test set was 'held out' but never states the split unit was the patient; Section 3.1 only excludes same-date duplicates, so cross-date patient overlap between partitions is not ruled out.
invented entities (3)
  • Cut-Noise technique
    purpose: Attenuate spurious activations outside lesion regions to improve heatmap precision
    Section 3.4 describes it qualitatively; no formal definition, parameters, or ablation, and no external benchmark demonstrating its effect.
  • Proprietary natural-language labeling model
    purpose: Extract initial disease labels from free-text radiologist reports
    Section 3.2: architecture, training data, and accuracy never specified; all downstream claims inherit its unknown error rate.
  • Proprietary image-processing quality algorithms
    purpose: Assess positioning and diagnostic quality to include or exclude images
    Section 3.2: no description or validation; their decisions determine the composition of the training and test sets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Classification to Localization and Clinical Validation: Large-Scale Development of a Deep Learning System for Thoracic Disease Detection on Chest Radiographs in Thailand." pith.science (2026). https://pith.science/paper/3EWITLMD

@misc{pith2026260709305,
  author       = {Pith},
  title        = {Pith review of: From Classification to Localization and Clinical Validation: Large-Scale Development of a Deep Learning System for Thoracic Disease Detection on Chest Radiographs in Thailand},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3EWITLMD}},
  note         = {Machine review of arXiv:2607.09305}
}
read the original abstract

Chest radiography (CXR) remains the most widely used thoracic imaging modality, yet expert interpretation is constrained by a severe shortage of radiologists in Thailand and across Southeast Asia. Local adaptation of deep learning models to Thai data has been shown to substantially improve accuracy on Thai populations. Here we present the development and comprehensive validation of the chest radiograph analysis model in Inspectra CXR version 5, a deep learning system that performs multi-label thoracic disease classification and weakly supervised lesion localization within a single model. The architecture couples a DenseNet-121 backbone with Attend-and-Compare Modules (ACM) and a Probabilistic Class Activation Map (PCAM) aggregation layer, producing a per-condition classification score and heatmap simultaneously. The model was developed on 874,858 frontal chest radiographs with paired radiologist reports from Siriraj Hospital, Bangkok. On a held-out, radiologist-verified in-domain test set of 19,871 cases, it achieved a mean AUROC of 0.994 (mean sensitivity 92.4%, specificity 98.6%) across nine clinically important conditions. On an independent generalization set of 5,992 cases from 13 hospitals across Thailand, the mean AUROC was 0.970, indicating robust transfer across sites. For localization, evaluated on 4,549 radiologist-annotated cases, the model attained a mean lesion-localization fraction (LLF) of 77.9% at 0.59 non-lesion localizations per image. In a usability evaluation with five thoracic radiologists, the system reached a classification concordance of 93.6%, a localization concordance of 94.7%, and a mean System Usability Scale (SUS) score of 89. These results indicate that a locally developed, localization-capable CXR system can deliver high accuracy, generalize across heterogeneous Thai hospitals, and earn the trust of practicing radiologists.

Figures

Figures reproduced from arXiv: 2607.09305 by the authors.

Figure 1
Figure 1. Overall structure of the Inspectra CXR version 5 analysis model: a DenseNet-121 + ACM encoder, an [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Per-condition AUROC on the in-domain test set (Dataset-A) and the multi-site generalization set (Dataset-B). [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 7 linked inside Pith

  1. [8]

    Deep learning in chest radiography: detection of findings and presence of change.PloS one, 13(10), 2018

    Ramandeep Singh, Mannudeep K Kalra, Chayanin Nitiwarangkul, John A Patti, Fatemeh Homayounieh, Atul Padole, Pooja Rao, Preetham Putha, Victorine V Muse, Amita Sharma, et al. Deep learning in chest radiography: detection of findings and presence of change.PloS one, 13(10), 2018

    2018

  2. [37]

    Kalra, Chayanin Nitiwarangkul, John A

    Ramandeep Singh, Mannudeep K. Kalra, Chayanin Nitiwarangkul, John A. Patti, Fatemeh Homayounieh, Atul Padole, Pooja Rao, Preetham Putha, Victorine V . Muse, Amita Sharma, and Subba R. Digumarthy. Deep learning in chest radiography: Detection of findings and presence of change.PLOS ONE, 13(10):e0204155, 2018. 11 APREPRINT- JULY23, 2026

    2018

  3. [1]

    Chest radiography in general practice: indications, diagnostic yield and consequences for patient management.Br J Gen Pract, 56(529):574–578, 2006

    Anouk M Speets, Yolanda van der Graaf, Arno W Hoes, Sandra Kalmijn, Alfred PE Sachs, Matthieu JCM Rutten, Jan Willem C Gratama, Alexander D Montauban van Swijndregt, and Willem PThM Mali. Chest radiography in general practice: indications, diagnostic yield and consequences for patient management.Br J Gen Pract, 56(529):574–578, 2006

    2006

  4. [2]

    Mortality estimates for south east asia, and indepth mortality surveillance: necessary but not sufficient?International journal of epidemiology, 42(4):1196–1199, 2013

    Chalapati Rao. Mortality estimates for south east asia, and indepth mortality surveillance: necessary but not sufficient?International journal of epidemiology, 42(4):1196–1199, 2013

    2013

  5. [3]

    Global tuberculosis report 2019, 2019

    World Health Organization. Global tuberculosis report 2019, 2019. Last accessed 13 April 2020

    2019

  6. [4]

    Assessing the accuracy and certainty in interpreting chest x-rays in the medical division.Clinical medicine, 13(4):349, 2013

    I Satia, S Bashagha, A Bibi, R Ahmed, S Mellor, and F Zaman. Assessing the accuracy and certainty in interpreting chest x-rays in the medical division.Clinical medicine, 13(4):349, 2013

    2013

  7. [5]

    Thai specialist statistics 2019, 2019

    The Medical Council of Thailand. Thai specialist statistics 2019, 2019. Last accessed 13 April 2020

    2019

  8. [6]

    Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning.arXiv preprint arXiv:1711.05225, 2017

    Pranav Rajpurkar, Jeremy Irvin, Kaylie Zhu, Brandon Yang, Hershel Mehta, Tony Duan, Daisy Ding, Aarti Bagul, Curtis Langlotz, Katie Shpanskaya, et al. Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning.arXiv preprint arXiv:1711.05225, 2017

    arXiv 2017

Show all 42 references
  1. [7]

    Chexpert: A large chest radiograph dataset with uncer- tainty labels and expert comparison.Proceedings of the AAAI Conference on Artificial Intelligence, 33:590–597, Jul 2019

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, and et al. Chexpert: A large chest radiograph dataset with uncer- tainty labels and expert comparison.Proceedings of the AAA...

    2019

  2. [9]

    Diagnose like a radiologist: Attention guided convolutional neural network for thorax disease classification.arXiv preprint arXiv:1801.09927, 2018

    Qingji Guan, Yaping Huang, Zhun Zhong, Zhedong Zheng, Liang Zheng, and Yi Yang. Diagnose like a radiologist: Attention guided convolutional neural network for thorax disease classification.arXiv preprint arXiv:1801.09927, 2018

    2018 arXiv

  3. [10]

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. Chestx- ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases.2017 IEEE Conference on Computer Vision an...

    2017

  4. [11]

    Mimic-cxr: A large publicly available database of labeled chest radiographs.arXiv preprint arXiv:1901.07042, 1(2), 2019

    Alistair EW Johnson, Tom J Pollard, Seth Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr: A large publicly available database of labeled chest radiographs.arXiv preprint arXiv:1901.07042, 1(2), 2019

    1901 arXiv

  5. [12]

    Padchest: A large chest x-ray image dataset with multi-label annotated reports.arXiv preprint arXiv:1901.07441, 2019

    Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria de la Iglesia-Vayá. Padchest: A large chest x-ray image dataset with multi-label annotated reports.arXiv preprint arXiv:1901.07441, 2019

    1901 arXiv

  6. [13]

    Seelwan Sathitratanacheewin and Krit Pongpirul. Deep learning for automated classification of tuberculosis-related chest x-ray: Dataset specificity limits diagnostic performance generalizability.arXiv preprint arXiv:1811.07985, 2018

    2018

  7. [14]

    Local adaptation improves accuracy of deep learning model for automated x-ray thoracic disease detection: A thai study.arXiv preprint arXiv:2004.10975, 2020

    Isarun Chamveha, Trongtum Tongdee, Pairash Saiviroonporn, and Warasinee Chaisangmongkon. Local adaptation improves accuracy of deep learning model for automated x-ray thoracic disease detection: A thai study.arXiv preprint arXiv:2004.10975, 2020

    2004 arXiv

  8. [15]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017

    2017

  9. [16]

    Learning visual context by comparison

    Minchul Kim, Jongchan Park, Seil Na, Chang Min Park, and Donggeun Yoo. Learning visual context by comparison. InEuropean Conference on Computer Vision (ECCV), pages 576–592. Springer, 2020

    2020

  10. [17]

    Weakly supervised lesion localization with probabilistic-cam pooling

    Wenwu Ye, Jin Yao, Hui Xue, and Yi Li. Weakly supervised lesion localization with probabilistic-cam pooling. arXiv preprint arXiv:2005.14480, 2020

    2005 arXiv

  11. [18]

    Learning to recognize abnormalities in chest x-rays with location-aware dense networks.arXiv preprint arXiv:1803.04565, 2018

    Sebastian Guendel, Sasa Grbic, Bogdan Georgescu, Kevin Zhou, Ludwig Ritschl, Andreas Meier, and Dorin Comaniciu. Learning to recognize abnormalities in chest x-rays with location-aware dense networks.arXiv preprint arXiv:1803.04565, 2018

    2018 arXiv

  12. [19]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  13. [20]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 10 APREPRINT- JULY23, 2026

    2009

  14. [21]

    Learning deep features for discriminative localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2921–2929, 2016

    2016

  15. [22]

    Jarrel CY Seah, Cyril HM Tang, Quinlan D Buchlak, et al. Effect of a comprehensive deep-learning model on the accuracy of chest x-ray interpretation by radiologists: A retrospective, multireader multicase study.The Lancet Digital Health, 3(8):e496–e506, 2021

    2021

  16. [23]

    Di Capua, Markus Y

    Jong Seok Ahn, Shadi Ebrahimian, Shaunagh McDermott, Sanghyup Lee, Laura Naccarato, John F. Di Capua, Markus Y . Wu, Eric W. Zhang, Victorine Muse, Benjamin Miller, Farid Sabzalipour, Bernardo C. Bizzo, Keith J. Dreyer, Parisa Kaviani, Subba R. Digumarthy, and Mannudeep K. Kal...

    2022

  17. [24]

    Al-antari, Pavlína Polášková, and Sergei Strukov

    Daniel Kvak, Anna Chromcová, Robert Hrubý, Eva Jan˚ u, Marek Biroš, Marija Pajdakovi´c, Karolína Kvaková, Mugahed A. Al-antari, Pavlína Polášková, and Sergei Strukov. Leveraging deep learning decision-support system in specialized oncology center: A multi-reader retrospective ...

    2023

  18. [25]

    Ai-assisted detection for chest x-rays (aid-cxr): a multi-reader multi-case study protocol.BMJ Open, 14:e080554, 2024

    Farhaan Khan, Indrajeet Das, Marusa Kotnik, Louise Wing, Edwin Van Beek, John Murchison, Jong Seok Ahn, Sang Hyup Lee, Ambika Seth, Abdala Trinidad Espinosa Morgado, Howell Fu, Alex Novak, Nabeeha Salik, Alan Campbell, Ruchir Shah, Fergus Gleeson, and Sarim Ather. Ai-assisted ...

    2024

  19. [26]

    Development and validation of deep learning–based automatic detection algorithm for malignant pulmonary nodules on chest radiographs.Radiology, 290(1):218–228, 2019

    Ju Gang Nam, Sunggyun Park, Eui Jin Hwang, et al. Development and validation of deep learning–based automatic detection algorithm for malignant pulmonary nodules on chest radiographs.Radiology, 290(1):218–228, 2019

    2019

  20. [27]

    Digumarthy, and Mannudeep K

    Hyunsuk Yoo, Sang Hyup Lee, Chiara Daniela Arru, Ruhani Doda Khera, Ramandeep Singh, Sean Siebert, Dohoon Kim, Yuna Lee, Ju Hyun Park, Hye Joung Eom, Subba R. Digumarthy, and Mannudeep K. Kalra. Ai-based improvement in lung cancer detection on chest radiographs: results of a m...

    2021

  21. [28]

    Jong Hyuk Lee, Hye Young Sun, Sunggyun Park, Hyungjin Kim, Eui Jin Hwang, Jin Mo Goo, and Chang Min Park. Performance of a deep learning algorithm compared with radiologic interpretation for lung cancer detection on chest radiographs in a health screening population.Radiology,...

    2020

  22. [29]

    Real-world evaluation of an ai triaging system for chest x-rays: A prospective clinical study.European Journal of Radiology, 181:111783, 2024

    Srinath Sridharan, Alicia Seah Xin Hui, Narayan Venkataraman, Prasanna Sivanath Tirukonda, Ram Pratab Jeyaratnam, Sindhu John, Saraswathy Suresh Babu, Perry Liew, Joe Francis, Tsai Koh Tzan, Wong Kang Min, Goh Min Liong, and Charlene Liew Jin Yee. Real-world evaluation of an a...

    2024

  23. [30]

    Kiraly, Sreenivasa Raju Kalidindi, Monde Muyoyeta, Jameson Malemela, Ting Shih, Greg S

    Sahar Kazemzadeh, Jin Yu, Shahar Jamshy, Rory Pilgrim, Zaid Nabulsi, Christina Chen, Neeral Beladia, Charles Lau, Scott Mayer McKinney, Thad Hughes, Atilla P. Kiraly, Sreenivasa Raju Kalidindi, Monde Muyoyeta, Jameson Malemela, Ting Shih, Greg S. Corrado, Lily Peng, Katherine ...

    2023

  24. [31]

    Zhi Zhen Qin, Shahriar Ahmed, Mohammad Shahnewaz Sarker, Kishor Paul, Ahammad Shafiq Sikder Adel, Tasneem Naheyan, Rachael Barrett, Sayera Banu, and Jacob Creswell. Tuberculosis detection from chest x-rays for triaging in a high tuberculosis-burden setting: an evaluation of fi...

    2021

  25. [32]

    ClinicalTrials.gov, 2025

    Ai assisted detection of chest x-rays (aid-cxr). ClinicalTrials.gov, 2025. NCT06075836

    2025

  26. [33]

    ClinicalTrials.gov, 2025

    Radiograph accelerated detection and identification of cancer in the lung (radical). ClinicalTrials.gov, 2025. NCT06044454

    2025

  27. [34]

    ClinicalTrials.gov, 2024

    Validity of pleural effusion detection software. ClinicalTrials.gov, 2024. NCT05903287

    2024

  28. [35]

    ClinicalTrials.gov, 2023

    Multi-reader retrospective study examining carebot ai cxr 2.0.21-v2.01 implementation in everyday radiology clinical practice. ClinicalTrials.gov, 2023. NCT05963945

    2023

  29. [36]

    Can deep learning reliably recognize abnormality patterns on chest x-rays? a multi-reader study examining one month of ai implementation in everyday radiology clinical practice

    Daniel Kvak, Anna Chromcová, Petra Ovesná, Jakub Dandár, Marek Biroš, Robert Hrubý, Daniel Dufek, and Marija Pajdakovi´c. Can deep learning reliably recognize abnormality patterns on chest x-rays? a multi-reader study examining one month of ai implementation in everyday radiol...

    2023

  30. [38]

    Sus: A retrospective.Journal of Usability Studies, 8(2):29–40, 2013

    John Brooke. Sus: A retrospective.Journal of Usability Studies, 8(2):29–40, 2013

    2013

  31. [39]

    An empirical evaluation of the system usability scale

    Aaron Bangor, Philip T Kortum, and James T Miller. An empirical evaluation of the system usability scale. International Journal of Human–Computer Interaction, 24(6):574–594, 2008

    2008

  32. [40]

    Hyunsuk Yoo, Eun Young Kim, Hyungjin Kim, Ye Ra Choi, Moon Young Kim, Sung Ho Hwang, Young Joong Kim, Young Jun Cho, and Kwang Nam Jin. Artificial intelligence-based identification of normal chest radiographs: A simulation study in a multicenter health screening cohort.Korean ...

    2022

  33. [41]

    Performance of a deep-learning algorithm for referable thoracic abnormalities on chest radiographs: A multicenter study of a health screening cohort.PLOS ONE, 16(2):e0246472, 2021

    Eun Young Kim, Young Jae Kim, Won-Jun Choi, Gi Pyo Lee, Ye Ra Choi, Kwang Nam Jin, and Young Jun Cho. Performance of a deep-learning algorithm for referable thoracic abnormalities on chest radiographs: A multicenter study of a health screening cohort.PLOS ONE, 16(2):e0246472, 2021

    2021

  34. [42]

    Deep learning-based detection system for multiclass lesions on chest radiographs: comparison with observer readings.European Radiology, 30:1359–1368, 2020

    Sohee Park, Sang Min Lee, Kyung Hee Lee, Kyu-Hwan Jung, Woong Bae, Jooae Choe, and Joon Beom Seo. Deep learning-based detection system for multiclass lesions on chest radiographs: comparison with observer readings.European Radiology, 30:1359–1368, 2020. 12

    2020

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.