Pith. sign in

REVIEW 3 major objections 5 minor 43 references

Once radiology reports are structured once, the label schema becomes a dictionary edit, not a corpus to relabel.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 03:40 UTC pith:SRS2A2AF

load-bearing objection Solid systems paper: one-time SRA cache plus alias recompile is real, costed, and useful for regrouping and long-tail coverage; open-ended schema expansion is weaker than the intro pitch. the 3 major comments →

arxiv 2607.06597 v1 pith:SRS2A2AF submitted 2026-07-06 eess.IV cs.CLcs.CV

Reconfigurable Radiology Labels Without Relabeling

Jean-Benoit Delbrouck , Dave Van Veen , Akash Pattnaik , Kalina Slavkova , Javid Abderezaei , Harris Bergman , Khan Siddiqui This is my paper
classification eess.IV cs.CLcs.CV
keywords radiology report labelingchest radiographsreconfigurable schemasstructured report annotationalias dictionaryCheXpert-14long-tail findingsmulti-label classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Public chest X-ray datasets ship with small fixed label lists such as CheXpert-14, yet free-text reports describe many more findings whose importance changes with task, site, and reader. This paper releases a pipeline that runs one structured annotation pass over the reports, caches the result, and then compiles any desired multi-label matrix by matching a radiologist-edited phrase dictionary against that cache. Schema changes therefore require only a local recompile: reconfiguring 223,000 MIMIC-CXR reports takes about three minutes and zero API cost, versus roughly $6,600 for an equivalent frontier-model relabeling pass. With a 58-label taxonomy the authors show that 43 percent of studies contain at least one finding outside CheXpert-14, and that image probes trained on the expanded labels match coarse baselines on shared targets while reaching 0.78 AUROC on expert-reviewed long-tail findings the fixed schema cannot express. The central claim is that the right unit of work for radiology labeling is no longer the corpus but the configuration of aliases once structure exists.

Core claim

After free-text radiology reports are converted once into cached structured annotations, the label schema becomes a configuration that can be edited through a radiologist-defined alias dictionary rather than a corpus that must be re-parsed or re-inferred. Reconfiguration of MIMIC-CXR therefore costs minutes and no API dollars, a 58-label taxonomy captures findings outside CheXpert-14 in 43 percent of studies, and image models trained on those labels preserve overall performance while unlocking measurable long-tail tasks.

What carries the argument

Radiological Aliases: a small radiologist-edited dictionary that maps each target label to report phrases (plus optional parents and exclude lists); matching those phrases against a one-time Structured Report Annotator graph produces auditable multi-label matrices without re-running inference.

Load-bearing premise

The one-time structured annotator must extract entities, statuses, and relations completely enough that closed-vocabulary phrase matching on the cache can produce accurate labels for any new schema; if the annotator misses something, the dictionary cannot recover it.

What would settle it

On a held-out set of expert-adjudicated native reports spanning the full 58-label taxonomy, measure whether dictionary recompilation after a single structured pass still yields high-precision positives for newly added long-tail labels, or whether many clinically present findings remain invisible because the structured annotator never recorded them.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Institutions can maintain one cached structured corpus and ship many task-specific label matrices by editing a shared alias file.
  • Long-tail findings that fixed 14-label schemas hide become first-class training and evaluation targets without new full-corpus labeling jobs.
  • Schema adaptation across hospitals or languages reduces to vocabulary synonym edits rather than model retraining or API re-labeling.
  • Reviewers can audit every positive label down to the exact report tokens and matched phrase that produced it.
  • The marginal cost of a schema change becomes minutes of local compute instead of thousands of dollars and privacy-constrained API calls.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same one-pass-plus-dictionary pattern could apply to other free-text clinical domains (pathology, discharge summaries) where fixed taxonomies currently force expensive re-labeling.
  • If structured-report quality continues to improve, the limiting factor for flexible medical labels may shift from inference cost to the clinical labor of writing good aliases.
  • Public dataset releases could ship the structured cache and the alias compiler, letting downstream users define their own heads without ever seeing raw report text again.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-stage radiology labeling pipeline: a one-time Structured Report Annotator (SRA; RadGraph-XL) pass that caches entity–relation graphs with present/uncertain/absent status, followed by a radiologist-edited alias dictionary (“Radiological Aliases”) that compiles multi-label matrices by closed-vocabulary matching in relational neighborhoods. After the SRA pass, schema changes are local dictionary edits and recompiles (196 s for MIMIC-CXR, no API cost) rather than full-corpus LLM or distillation relabeling (~$6.6K for Claude Opus 4.7). On a 58-label taxonomy the authors report that 43% of HOPPR CXR studies have at least one positive outside CheXpert-14; report-label precision is competitive with CheXbert on shared labels and non-zero on long-tail labels CheXbert cannot emit; frozen-feature image probes match CheXpert-14 aggregate AUROC while reaching ~0.78 macro-AUROC on expert CXR-LT long-tail labels. Code and the full alias dictionary are released.

Significance. If the result holds, the practical unit of work for large-scale radiology labeling shifts from “relabel the corpus” to “edit a configuration,” which is a genuine systems contribution for multi-site, task-dependent, and long-tail CXR research. Strengths that should be credited: (i) transparent, reproducible cost and recompile measurements (Table 1) with an explicit token model; (ii) open release of code and the full alias dictionary; (iii) precision-first evaluation on three gold sets with CheXbert and raw-text ablations (Tables 2–4); (iv) honest limitations on SRA error propagation, report-vs-image labels, and non-external CXR-LT transfer. The work is more engineering/systems than a new learning method, but the reconfiguration framing and measured marginal cost of schema change are useful and under-served in public CXR labeling practice.

major comments (3)
  1. [Abstract; §1; §3.1–3.2; §5.2; App. C] Introduction and abstract frame the method as supporting open-ended, task-/site-/reader-dependent schemas “without relabeling.” That guarantee is load-bearing and is only fully secure for regrouping, synonym alignment, and findings already surfaced as SRA observation seeds (§3.1–3.2; Appendix A). Limitations §5.2 correctly states that missed entities/relations cannot be recovered by aliases. Evidence is strong for within-vocabulary and synonym edits (CheXpert 5x200; PadChest synonyms raising macro-F1 0.50→0.62 in Table 3; CXR-LT coverage of 12 non-CheXbert labels) but thin for clinically novel findings outside the SRA’s training distribution; the LVAD appendix demo (12 aliases, 7 hits on 100k studies) is too small to settle open-ended expansion. Please tighten abstract/intro claims to distinguish (a) cheap reconfiguration of extractable concepts from (b) truly novel findings that may sti
  2. [Abstract; §4.3; Table 5; Table 17; §5.8] Image utility on long-tail labels (Table 5; Table 17) is reported as ~0.78 macro-AUROC for a ConvNeXt-T probe trained on MIMIC weak labels and tested on CXR-LT gold. §5.8 and the text correctly note that CXR-LT is a MIMIC subset and is a consistency check, not leakage-free external validation. The abstract and §4.3 headline still read as if this is general long-tail image performance. Please restate the abstract/§4.3 claim as “gold-label consistency on a MIMIC subset” (or retrain excluding CXR-LT patients/studies) so the central “unlocks finer tasks” claim is not overstated relative to the experimental design.
  3. [§4.2; Table 2–3; §5.6; Conclusion] Report-label evaluation is precision-first and appropriately scoped as coverage vs accuracy against CheXbert (Table 2 split into 13 representable vs 12 non-representable labels). Still load-bearing for the “usable labels” claim: there is no systematic false-positive/error analysis or inter-rater protocol for the 58-label dictionary (curator count, freeze date, agreement), which §5.6 defers to a peer-reviewed version. For a methods paper whose product is the dictionary+compiler, a short error taxonomy on CXR-LT/PadChest false positives (ontology mismatch vs SRA error vs alias over-match) and a documented curation protocol are needed to support clinical auditability claims in §3.2 and the conclusion.
minor comments (5)
  1. [Table 1; §4.2] Table 1 “Ours” column is correctly described as marginal recompile cost, but a one-line reminder of one-time SRA compute (100k reports ≈22 min on 3 A100s, §3.1) in the table caption would prevent misreading the pipeline as free end-to-end.
  2. [Appendix A; §4.1] Uncertainty handling: only “definitely present” is positive (U-Zeros). An uncertainty-policy ablation is mentioned as a knob in Appendix A but not reported; a brief sensitivity note would help readers who use U-Ones or multi-class heads.
  3. [Table 6; §4.3] Parent-conditioned child AUROCs (Table 6) are a nice contribution; confidence intervals for hernia_abnormality (n=101, 2 children) are very wide—flag low-support parents more explicitly in the text.
  4. [§1; footer] Typo/consistency: “resent, absent, or uncertain” in §1 should be “present.” arXiv date line “July 9, 2026” / pricing “2026-04” is fine for a preprint but should be normalized for journal production.
  5. [Table 4; §4.2] Raw-text vs SRA gap vanishes on PadChest-GR because inputs are short single-finding sentences (Table 4). State earlier that PadChest-GR is a vocabulary-transfer stress test, not a full-report disambiguation test, so readers do not over-generalize the null ablation.

Circularity Check

0 steps flagged

No circular derivation: reconfiguration cost, coverage, and probe metrics are measured against external gold sets and wall-clock/API pricing, not forced by definition or self-citation.

full rationale

The paper’s load-bearing claims are empirical measurements, not closed-form reductions of their inputs. (1) Recompile time (~196 s on MIMIC-scale cached annotations) and LLM cost (~$6.6K Opus) are wall-clock and token-price estimates, not quantities defined by the alias dictionary. (2) The 43% “outside CheXpert-14” figure is a corpus count under a human-curated 58-label taxonomy; it is not a fit that is then re-presented as a prediction. (3) Report-label precision/recall and image-probe AUROCs are scored against external expert subsets (CheXpert 5x200, CXR-LT Task 2, PadChest-GR) and held-out splits; alias phrases are radiologist-edited inputs, not parameters fitted to maximize those metrics. PadChest synonym edits are described as vocabulary crosswalks from PadChest’s own label definitions, not test-set tuning. (4) Use of RadGraph-XL (Delbrouck et al., overlapping authors) is component reuse of a published SRA, not a uniqueness theorem or ansatz that forces the reconfiguration result; the paper treats other SRAs as drop-in replacements and states that SRA misses cannot be recovered by aliases. No self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation chain is present. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard NLP/ML tooling plus domain choices: that a one-time structured graph plus human phrase aliases is a valid substitute for schema-specific full-corpus labeling. Free parameters are mostly human design choices (taxonomy, alias lists, exclude lists), not numerical fits to maximize AUROC. Invented entities are methodological constructs (Radiological Aliases compiler, frozen 58-label taxonomy), not physical postulates. No machine-checked formal core.

free parameters (4)
  • 58-label frozen taxonomy (48 leaves + 10 parents)
    Hand-chosen label set and hierarchy used for all main experiments; placement of findings under single parents is acknowledged as pragmatic (§5.5).
  • Radiological Aliases phrase lists and exclude lists
    Human-curated synonym/exclude phrases per disease key; coverage bounds recall and exclude gates bound precision (Appendix A, Tables 7–10).
  • LLM cost model assumptions (162 input tokens/report, 800 prompt overhead, 200 output tokens)
    Measured mean input tokens on a sample; overhead/output are assumed constants used to produce Table 1 dollar figures.
  • Uncertainty collapse policy (U-Zeros: only definitely present as positive)
    Training and metrics treat uncertain/absent/missing as negative following CheXpert convention; alternative policies are knobs not ablated in main results.
axioms (4)
  • domain assumption A Structured Report Annotator (RadGraph-XL) produces a semantic graph of Anatomy/Observation nodes with present/uncertain/absent status and relations sufficient for downstream alias matching.
    Invoked in §3.1 as the one-time expensive step; errors propagate by design (§5.2).
  • domain assumption Report-derived labels (present in text) are a valid weak-supervision signal for image probes even when findings may be present on the image but unmentioned.
    Stated in Limitations §5.1; underpins all image experiments in §4.3.
  • ad hoc to paper Closed-vocabulary subset matching of normalized tokens in a relational neighborhood implements the intended clinical label definition.
    Full algorithm in Appendix A; precision gates (exclude, optional require_anatomy) are design choices of this compiler.
  • standard math Standard multi-label classification metrics (macro-P/R/F1, AUROC) with bootstrap CIs adequately summarize labeler and probe quality across imbalanced findings.
    Used throughout §4.2–4.3 and appendices.
invented entities (2)
  • Radiological Aliases compiler independent evidence
    purpose: Map radiologist-edited phrase dictionaries onto cached SRA graphs to produce reconfigurable multi-label matrices with evidence spans.
    Core methodological construct of the paper; independent evidence is the released code/dictionary and measured recompile behavior, not an external physical prediction.
  • 58-label CXR taxonomy with parent groups no independent evidence
    purpose: Provide a frozen schema spanning CheXpert-14, CXR-LT, devices, and long-tail findings for experiments.
    Paper-specific taxonomy; editable by design, not claimed as a universal clinical standard.

pith-pipeline@v1.1.0-grok45 · 26045 in / 3650 out tokens · 41664 ms · 2026-07-11T03:40:02.889085+00:00 · methodology

0 comments
read the original abstract

Public chest-radiograph (CXR) datasets are typically released with small, fixed label schemas such as CheXpert-14. However, the underlying free-text reports describe far more findings -- and which findings matter depends on the task, site, and reader. We release a pipeline that converts free-text reports into multi-label matrices and then reconfigures the label schema through dictionary edits rather than new inference passes, i.e., without relabeling the corpus. After this one-time pass, reconfiguring MIMIC-CXR (223K reports) from cached annotations takes 196 seconds with no API cost, compared to \$6.6K for an equivalent relabeling pass with Claude Opus 4.7. Using a 58-label taxonomy, we show that 43\% of CXR studies contain at least one finding outside CheXpert-14. Image probes trained on these labels match CheXpert-14 probes on shared targets while also reaching 0.78 AUROC on expert-reviewed long-tail labels that CheXpert-14 cannot represent. These results suggest a different unit of work for radiology labeling: once reports are structured, the label schema becomes a configuration to edit, not a corpus to relabel.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · 2 internal anchors

  1. [1]

    H., Boland, G

    Abujudeh, H. H., Boland, G. W., Kaewlai, R., et al. Abdominal and pelvic computed tomography (ct) interpretation: discrepancy rates among experienced radiologists. European Radiology, 20 0 (8), 2010

    work page 2010

  2. [2]

    A., Walker, E

    Bruno, M. A., Walker, E. A., and Abujudeh, H. H. Understanding and confronting our mistakes: the epidemiology of error in radiology and strategies for error reduction. RadioGraphics, 35 0 (6), 2015

    work page 2015

  3. [3]

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

    Bustos, A., Pertusa, A., Salinas, J.-M., and de la Iglesia-Vay \'a , M. PadChest : A large chest x-ray image dataset with multi-label annotated reports. Medical Image Analysis, 66, 2020

    work page 2020

  4. [4]

    Castro, D. C. d., Bustos, A., Bannur, S., Hyland, S. L., et al. PadChest-GR : A bilingual chest x-ray dataset for grounded radiology report generation. arXiv preprint arXiv:2411.05085, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    CheXpert Plus: Augmenting a Large Chest X-ray Dataset with Text Radiology Reports, Patient Demographics and Additional Image Formats

    Chambon, P., Delbrouck, J.-B., Sounack, T., Huang, S.-C., Chen, Z., Varma, M., Truong, S., Chuong, C. T., and Langlotz, C. P. CheXpert Plus : Augmenting a large chest x-ray dataset with text radiology reports, patient demographics and additional image formats. arXiv preprint arXiv:2405.19538, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  6. [6]

    RadGraph-XL : A large-scale expert-annotated dataset for entity and relation extraction from radiology reports

    Delbrouck, J.-B., Chambon, P., Chen, Z., Varma, M., Johnston, A., Blankemeier, L., Van Veen, D., Bui, T., Truong, S., and Langlotz, C. RadGraph-XL : A large-scale expert-annotated dataset for entity and relation extraction from radiology reports. In Findings of the Association for Computational Linguistics (ACL 2024), 2024

    work page 2024

  7. [7]

    J., Jürgensen, L., Donle, L., Mohamad, F

    Dorfner, F. J., Jürgensen, L., Donle, L., Mohamad, F. A., Bodenmann, T. R., Cleveland, M. C., Busch, F., Adams, L. C., Sato, J., Schultz, T., et al. Comparing commercial and open-source large language models for labeling chest radiograph reports. Radiology, 2025

    work page 2025

  8. [8]

    langextract : Structured information extraction from unstructured text with language models

    Google . langextract : Structured information extraction from unstructured text with language models. Software library, https://github.com/google/langextract , 2024

    work page 2024

  9. [9]

    E., Er, S., Almas, F., Simsek, A

    Hamamci, I. E., Er, S., Almas, F., Simsek, A. G., Esirgun, S. N., Dogan, I., Dasdelen, M. F., Wittmann, B., Simsar, E., Simsar, M., et al. A foundation model utilizing chest CT volumes and radiology reports for supervised-level zero-shot detection of abnormalities. arXiv preprint arXiv:2403.17834, 2024

    work page arXiv 2024

  10. [10]

    C., Shih, G., Summers, R

    Holste, G., Wang, S., Jiang, Z., Shen, T. C., Shih, G., Summers, R. M., Peng, Y., and Wang, Z. Long-tailed classification of thorax diseases on chest x-ray: A new benchmark study. MICCAI DALI Workshop, 2022

    work page 2022

  11. [11]

    Holste, G. et al. CXR-LT 2024 : Long-tailed, multi-label thoracic disease classification. MICCAI CVAMD Workshop, 2024

    work page 2024

  12. [12]

    CheXpert : A large chest radiograph dataset with uncertainty labels and expert comparison

    Irvin, J., Rajpurkar, P., et al. CheXpert : A large chest radiograph dataset with uncertainty labels and expert comparison. In AAAI, 2019

    work page 2019

  13. [13]

    Jain, S., Agrawal, A., Saporta, A., Truong, S. Q. H., Duong, D. N., Bui, T., Chambon, P., Zhang, Y., Lungren, M. P., Ng, A. Y., Langlotz, C. P., and Rajpurkar, P. RadGraph : Extracting clinical entities and relations from radiology reports. In NeurIPS Datasets and Benchmarks, 2021

    work page 2021

  14. [14]

    Johnson, A. E. W., Pollard, T. J., Berkowitz, S. J., et al. MIMIC-CXR , a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6 0 (1), 2019

    work page 2019

  15. [15]

    A ConvNet for the 2020s

    Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., and Xie, S. A ConvNet for the 2020s. In CVPR, 2022

    work page 2022

  16. [16]

    J., Schwartz, K

    McDonald, R. J., Schwartz, K. M., Eckel, L. J., et al. The effects of changes in utilization and technological advancements of cross-sectional imaging on radiologist workload. Academic Radiology, 22 0 (9), 2015

    work page 2015

  17. [17]

    Hidden stratification causes clinically meaningful failures in machine learning for medical imaging

    Oakden-Rayner, L., Dunnmon, J., Carneiro, G., and R \'e , C. Hidden stratification causes clinically meaningful failures in machine learning for medical imaging. ACM CHIL, 2020

    work page 2020

  18. [18]

    Caveats in generating medical imaging labels from radiology reports

    Olatunji, T., Yao, L., Covington, B., Rhodes, A., and Upton, A. Caveats in generating medical imaging labels from radiology reports. In MIDL, 2019

    work page 2019

  19. [19]

    DINOv2 : Learning robust visual features without supervision

    Oquab, M., Darcet, T., et al. DINOv2 : Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024

    work page 2024

  20. [20]

    NegBio : a high-performance tool for negation and uncertainty detection in radiology reports

    Peng, Y., Wang, X., Lu, L., Bagheri, M., Summers, R., and Lu, Z. NegBio : a high-performance tool for negation and uncertainty detection in radiology reports. In AMIA Joint Summits on Translational Science, 2018

    work page 2018

  21. [21]

    Smit, A., Jain, S., Rajpurkar, P., Pareek, A., Ng, A., and Lungren, M. P. CheXbert : Combining automatic labelers and expert annotations for accurate radiology report labeling using BERT . In EMNLP, 2020

    work page 2020

  22. [22]

    Delbrouck, Jean-Benoit and Chambon, Pierre and Chen, Zhihong and Varma, Maya and Johnston, Andrew and Blankemeier, Louis and Van Veen, Dave and Bui, Tan and Truong, Steven and Langlotz, Curtis , booktitle=

    work page

  23. [23]

    Smit, Akshay and Jain, Saahil and Rajpurkar, Pranav and Pareek, Anuj and Ng, Andrew and Lungren, Matthew P , booktitle=

    work page

  24. [24]

    Irvin, Jeremy and Rajpurkar, Pranav and others , booktitle=

    work page

  25. [25]

    Holste, Gregory and others , journal=

    work page

  26. [26]

    Castro, Daniel Coelho de and Bustos, Aurelia and Bannur, Shruthi and Hyland, Stephanie L and others , journal=

    work page

  27. [27]

    Johnson, Alistair E W and Pollard, Tom J and Berkowitz, Seth J and others , journal=

    work page

  28. [28]

    Chambon, Pierre and Delbrouck, Jean-Benoit and Sounack, Thomas and Huang, Shih-Cheng and Chen, Zhihong and Varma, Maya and Truong, Steven and Chuong, Chuong Tran and Langlotz, Curtis P , journal=

    work page

  29. [29]

    Transactions on Machine Learning Research , year=

    Oquab, Maxime and Darcet, Timoth. Transactions on Machine Learning Research , year=

    work page

  30. [30]

    Liu, Zhuang and Mao, Hanzi and Wu, Chao-Yuan and Feichtenhofer, Christoph and Darrell, Trevor and Xie, Saining , booktitle=. A

    work page

  31. [31]

    Chen, Zhihong and Varma, Maya and Delbrouck, Jean-Benoit and Paschali, Magdalini and Blankemeier, Louis and others , journal=

    work page

  32. [32]

    Peng, Yifan and Wang, Xiaosong and Lu, Le and Bagheri, Mohammadhadi and Summers, Ronald and Lu, Zhiyong , booktitle=

    work page

  33. [33]

    Wang, Xiaosong and Peng, Yifan and Lu, Le and Lu, Zhiyong and Bagheri, Mohammadhadi and Summers, Ronald M , journal=

    work page

  34. [34]

    Medical Image Analysis , volume=

    Bustos, Aurelia and Pertusa, Antonio and Salinas, Jose-Maria and de la Iglesia-Vay. Medical Image Analysis , volume=

    work page

  35. [35]

    Academic Radiology , volume=

    The effects of changes in utilization and technological advancements of cross-sectional imaging on radiologist workload , author=. Academic Radiology , volume=

    work page

  36. [36]

    RadioGraphics , volume=

    Understanding and confronting our mistakes: the epidemiology of error in radiology and strategies for error reduction , author=. RadioGraphics , volume=

    work page

  37. [37]

    European Radiology , volume=

    Abdominal and pelvic computed tomography (CT) interpretation: discrepancy rates among experienced radiologists , author=. European Radiology , volume=

    work page

  38. [38]

    MIDL , year=

    Caveats in generating medical imaging labels from radiology reports , author=. MIDL , year=

    work page

  39. [39]

    ACM CHIL , year=

    Hidden stratification causes clinically meaningful failures in machine learning for medical imaging , author=. ACM CHIL , year=

    work page

  40. [40]

    MICCAI DALI Workshop , year=

    Long-tailed classification of thorax diseases on chest x-ray: A new benchmark study , author=. MICCAI DALI Workshop , year=

    work page

  41. [41]

    Jain, Saahil and Agrawal, Ashwin and Saporta, Adriel and Truong, Steven Q H and Duong, Du Nguyen and Bui, Tan and Chambon, Pierre and Zhang, Yuhao and Lungren, Matthew P and Ng, Andrew Y and Langlotz, Curtis P and Rajpurkar, Pranav , booktitle=

    work page

  42. [42]

    Radiology , year=

    Comparing commercial and open-source large language models for labeling chest radiograph reports , author=. Radiology , year=

    work page

  43. [43]

    A Foundation Model Utilizing Chest

    Hamamci, Ibrahim Ethem and Er, Sezgin and Almas, Furkan and Simsek, Ayse Gulnihan and Esirgun, Sevval Nil and Dogan, Irem and Dasdelen, Muhammed Furkan and Wittmann, Bastian and Simsar, Enis and Simsar, Mehmet and others , journal=. A Foundation Model Utilizing Chest

    work page