Pith. sign in

REVIEW 3 major objections 5 minor 46 references

RadPRISM: Schema-stratified radiology-report supervision for concept-disentangled image representations and visual grounding

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that structuring radiology reports into per-concept text spans and aligning each clinical concept in its own visual subspace yields chest X-ray representations that outperform global report-image alignment in zero-shot cla

desk verdict Genuinely useful method and strong external grounding results, but the headline internal comparison is confounded: the global baseline sees raw reports while RadPRISM gets LLM-structured, imputed spans, so the 0.151 AUROC gain is not cleanly attributable to subspace geometry. read the letter →

arxiv 2608.00147 v1 pith:OFTDR5A2 submitted 2026-07-31 cs.CV cs.LG

Fabian Drexel , Marlene Fritzsche , Era Stambollxhiu , Miriam Kumpf , Lena Schmitzer , Lea Schumann , Jannik Kahmann , Friedrich Puttkammer
show 10 more authors
Johannes Moll Jannik Lübberstedt Zeineb Ben Chaaben Anirudh Narayanan Cosmin I. Bercea Sebastian Ziegelmayer Marcus R. Makowski Daniel Rueckert Lisa C. Adams Keno K. Bressem
This is my paper
classification cs.CVcs.LG
keywords vision-languagepretrainingchestradiographconcept-disentangledrepresentationsschema-stratifiedsupervisionzero-shotclassificationvisualgroundingcontrastivelearningreportstructuring
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 tries to show that the way supervision is organized matters as much as the data: instead of aligning a whole chest X-ray to a whole free-text radiology report, RadPRISM first has a large language model split each report into short text spans for 19 clinically defined concepts (heart, pneumonia, pleural effusion, support devices, etc.), then aligns each concept's span to its own dedicated visual subspace in the image. On an internal archive of over 200,000 examinations, this schema-stratified supervision raised zero-shot classification from 0.717 to 0.868 macro AUROC over a matched baseline that aligned the full report, and on an external benchmark the model's visual grounding of cardiomegaly matched the human reference (0.983 vs 0.972 hit rate) with no localization supervision. A radiologist reader study also confirmed that per-concept retrieval can surface descriptive text (location, extent, certainty) that fixed label sets cannot express. The authors argue this is because routing each clinical concept into its own subspace prevents rare descriptive details from being averaged away by the bulk of the report.

What carries the argument

The central machinery is a clinician-defined schema acting as a stratification axis, implemented as 19 concept tokens in a cross-attention block on top of a frozen chest-X-ray vision transformer, with a frozen multilingual text encoder producing concept-specific text embeddings. Each concept's text span extracted by an on-premise LLM is aligned to its visual subspace by a symmetric per-concept contrastive loss with learnable temperatures, masking batches where a concept's text is missing, and oversampling under-represented concepts. Synthetic absence descriptions are imputed for rare concepts before training, and binary labels for evaluation are derived by collapsing four-class annotations.

What would settle it

Two checks would settle it: (1) train the global-alignment baseline on the identical structured, imputed per-concept spans (concatenated) while holding all else fixed—if zero-shot AUROC approaches 0.868, the advantage is textual, not geometric; (2) re-score the internal test split against a much larger set of human-corrected labels (well beyond 200 reports)—if the AUROC gap narrows substantially, LLM label bias drove the reported comparison.

Watch

Extended reading notes

Core claim

The central claim is that the geometry of supervision, not just the scale of data, determines what a medical vision-language model learns. The authors propose that decomposing each free-text radiology report into per-concept text spans along a clinician-defined 19-concept schema, and aligning each concept to its own dedicated visual subspace via per-concept contrastive losses, produces representations that are discriminative, spatially faithful, and inspectable per finding. They report that on an internal archive of 203,602 examinations, RadPRISM reaches 0.868 macro AUROC in zero-shot classification, surpassing a matched global-alignment baseline trained on the same encoders and data (0.717)

Load-bearing premise

The headline gain is attributed to the per-concept subspace geometry, which requires that the global baseline is not simply disadvantaged by missing the structured, imputed text inputs, and that the LLM-generated labels used as internal ground truth are unbiased enough not to favour the proposed model.

Editorial extensions

If this is right

  • Concept-stratified supervision alone—with identical encoders, adapters, and training data—raises internal zero-shot macro AUROC from 0.717 to 0.868 (0.151 absolute gain) relative to global report alignment.
  • The learned per-concept subspaces support zero-shot classification of all 19 findings and, under linear probing, exceed a label-supervised baseline trained directly on the frozen image encoder (0.890 vs 0.859 AUROC).
  • The representation transfers across simultaneous shifts in institution and label vocabulary: external zero-shot classification remains competitive with a purpose-built reference, and visual grounding on several findings beats that reference by up to 4.3x at matched input resolution.
  • Cardiomegaly grounding reaches a pointing-game hit rate of 0.983, statistically indistinguishable from the human reference of 0.972, with no localization supervision during training.
  • Concept-stratified text retrieval, validated by radiologists at 0.78 macro top-3 correctness, can surface descriptive findings (device subtypes, extent, laterality, diagnostic certainty) that are structurally absent from fixed-label vocabularies, and agreement between classification and retrieval provides a per-concept confidence signal.

Reading between the lines

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

  • A direct test of the paper's central attribution: retrain the global-alignment baseline on the same structured, imputed per-concept spans, concatenated into one report embedding; if its AUROC climbs toward 0.868, the gain is driven by the improved text, not the subspace geometry.
  • If the subspace-geometry claim is right, the schema-stratified recipe should transfer to volumetric modalities such as CT, where reports describe findings across three spatial axes; the paper identifies this as future work and does not test it.
  • The documented ceiling effect of LLM structuring quality suggests that investing in human-corrected spans for the highest-disagreement concepts (pulmonary edema, fibrosis, mass lesion) could raise the lower per-concept AUROCs (0.76–0.79) more than any architectural change.
  • The classification-retrieval consistency signal could be operationalized as an inference-time, finding-level uncertainty estimate without retraining, since radiologist-validated correctness rises sharply when the two outputs agree; the paper presents this as an observation, not a validated uncertainty method.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. RadPRISM introduces a concept-stratified vision-language pretraining framework for chest radiographs. Free-text German radiology reports are decomposed by a locally hosted LLM (GPT-OSS-120B) into per-concept text spans and categorical labels according to a 19-concept clinician-defined schema. Each concept is aligned in its own visual subspace via per-concept contrastive losses with learnable temperatures, using a frozen DINOv2-based CXR encoder and a frozen Qwen3-Embedding text encoder plus trainable projections and a cross-attention concept-token block. On an internal multi-year TUM dataset (203,602 examinations), the authors report a macro AUROC of 0.868 for zero-shot classification versus 0.717 for a report-level global-alignment baseline, competitive external zero-shot classification on CheXlocalize against CARZero, and strong pointing-game grounding (e.g., Cardiomegaly hit rate 0.983 at 518 px, matching a human reference). A radiologist reader study on 200 cases adds per-concept classification, attention-map, and top-3 text-retrieval correctness assessments, with retrieval@rank-3 macro correctness 0.78. The central claim is that schema-stratified supervision outperforms global alignment because of the supervisory geometry, and that the resulting representation is discriminative, spatially faithful, and concept-disentangled.

Significance. If the central causal claim is upheld, RadPRISM would be a valuable step toward interpretable, concept-level medical vision-language representations that can be built from institutional free-text archives without manual annotation. The paper is methodologically rich: it includes a large single-institution cohort, a structured reader study of LLM extraction quality, quantitative external grounding benchmarks, a radiologist reader study for downstream tasks, seed-variability analysis, and a code repository. The external grounding comparison is particularly useful because it provides independent evidence that concept-stratified attention can localize findings without localization supervision. However, the headline internal advantage over the global-alignment baseline is confounded by differences in the input text (raw reports versus LLM-extracted spans plus synthetic imputation) and the internal test labels are generated by the same LLM pipeline that produces the training spans. These issues must be addressed before the 0.151 macro-AUROC gap can be attributed to supervisory geometry rather than to improved textual supervision or label self-consistency.

major comments (3)
  1. [Methods, Baseline and ablation setups / Supplementary Section 1.1.11; Discussion, second paragraph] The claim that the 0.151 macro-AUROC advantage over the global-alignment baseline 'isolates supervisory geometry as the active ingredient' is not supported by the described experiments. The report-level baseline 'embeds the full report directly' and uses a single query token, while RadPRISM trains on per-concept spans extracted by GPT-OSS-120B and on rule-based synthetic absence imputation (Methods, Data preprocessing; Supplementary Table 5). Thus the comparison varies the input text and its coverage in addition to the alignment geometry. A matched-text control is needed: training the global baseline on the same concatenated concept spans with the same imputation (and/or training RadPRISM without imputation) would isolate the geometry effect. As reported, the AUROC gap could be driven largely by cleaner, concept-specific text and synthetic negative supervision rather than by subspace geo
  2. [Methods, LLM-based CXR report structuring and labeling; Methods, Image classification; Supplementary Tables 4–5] The internal zero-shot and linear-probing evaluation labels are generated by the same GPT-OSS-120B structuring pipeline that produces the per-concept training spans and labels, with human validation on only 200 reports (median macro F1 0.89, but no per-test-label audit). For pathology concepts with unknown (label 0) states, negative labels are additionally imputed under a surrounding-coverage rule during evaluation-label generation. This introduces partial circularity: part of the reported internal performance may measure self-consistency between the LLM's text structuring and the model's learned representation rather than clinically independent truth. The external CheXlocalize results provide some independent support, but the central internal head-to-head claim requires either an independently labeled test set or reader-validated labels on a substantially larger and more representative
  3. [Methods, Data preprocessing; Supplementary Table 5] The synthetic imputation materially changes the training and evaluation distributions for rare concepts. For example, text coverage for Pulmonary edema rises from 0.56% to 1.33% and its label masking rate decreases from 99.42% to 98.67%; similar changes occur for Emphysema, Fibrosis, and several support-device concepts. These changes mean that 'the same training corpus' stated in the Discussion is not literally true for the text inputs: the concept-stratified model receives additional synthetic spans that the global baseline never sees. This amplifies the confound described in the first comment and should be explicitly acknowledged and controlled for in the experimental design, not only in the limitations narrative.
minor comments (5)
  1. [Abstract] The phrase 'matched global-alignment baseline' is misleading given the differing text inputs; consider 'global-alignment baseline trained on the same images and reports' or add a parenthetical noting the difference.
  2. [Methods, Equation (1)] The notation for the concept-wise loss components uses both superscripts (L(k)) and subscripts (L_i→t, L_t→i) inconsistently. Please standardize.
  3. [Supplementary Table 10] The 'Report date' row has no values; a footnote explaining that no report-date entries occurred in the 200-report reader sample is already present in the main text but could be repeated in the table caption for self-containedness.
  4. [Results, Reader study and concept-disentangled retrieval] The reader study used a single reader per case (as acknowledged in Methods); this is acceptable but should be restated in the Results or Figure 4 legend when reporting aggregate correctness rates.
  5. [Discussion, limitations paragraph] The limitations paragraph is candid, but it does not mention the evaluation-label circularity or the text-input confound in the global-baseline comparison. Adding those points would strengthen the manuscript's framing.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: internal evaluation labels and training spans come from the same GPT-OSS-120B pipeline, and the headline geometry comparison is confounded by different text inputs.

  1. self definitional [Methods – LLM-based CXR report structuring and labeling; Methods – Data preprocessing (label generation); internal results in Results – Concept-stratified representation learning]
    "structuring was performed by a locally hosted endpoint of the GPT-OSS-120B large language model ... The prespecified schema-extracted spans and labels provided the per-concept supervision for the proposed RadPRISM model. ... For the concept-wise supervised linear-head finetuning and classification evaluation, binary labels were generated by mapping support-device annotations directly to present/absent and by collapsing thoracic-organ and pathology annotations into negative (LLM-assigned label 1, normal) versus positive (LLM-assigned labels 2 and 3, ambiguous/borderline or clearly pathological)"

    The internal test labels and the per-concept training supervision are both outputs of the same GPT-OSS-120B structuring pipeline. RadPRISM's concept subspaces are aligned to the LLM-extracted spans, and the reported internal zero-shot and linear-probe AUROC values are then computed against the same LLM's binary labels. The headline 0.868 vs 0.717 gain therefore partly measures how well the model reproduces the structuring LLM's own annotation scheme rather than an independent clinical ground truth. The 200-report human validation (median macro F1 0.89) provides only a limited external anchor, so the main internal comparison remains substantially self-referential.

full rationale

The clearest circular element is the internal evaluation: both the per-concept training text/labels and the test-split ground-truth labels are generated by the same GPT-OSS-120B structuring workflow, so the internal AUROC advantage partially reflects self-consistency with the LLM rather than independent clinical truth. This is compounded by a non-circular but serious confound: the global-alignment baseline 'embeds the full report directly' (Supplementary Section 1.1.11), while RadPRISM trains on LLM-structured concept spans plus rule-based synthetic absence imputation (Methods, Data preprocessing), so the claimed isolation of 'supervisory geometry as the active ingredient' is not established by the matched-baseline comparison. However, the external CheXlocalize classification, the pointing-game grounding against CARZero and human references, and the radiologist reader study are independent of the internal LLM label loop and provide genuine supporting evidence for the model's utility. The circularity is therefore partial: it undermines the internal attribution claim but does not reduce the whole paper to a tautology. Score 4 reflects one load-bearing self-referential evaluation loop with independent external content elsewhere.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced; the 19 concept tokens are learned parameters. The central claim rests on a hand-specified clinician schema, synthetic imputation rules, hand-authored zero-shot prompts, and an LLM structuring pipeline whose labels also serve as the internal evaluation ground truth.

free parameters (3)
  • Per-concept learnable contrastive temperatures tau_k = initialized 0.07, constrained to 0.02-0.12; final values not reported
    Per-concept alignment sharpness trained on data; part of the objective, not a scientific constant.
  • Synthetic absence imputation rules = 65% local coverage threshold; cap at observed real-data coverage; random subset imputed
    Hand-chosen by a radiologist to fill empty concept text/label fields; affects training distribution and may contribute to gains over the raw-report baseline.
  • Zero-shot prompt prototypes = 3 positive and 3 negative German sentences per concept
    Hand-authored prompt anchors used to define positive/negative prototypes; zero-shot performance depends on this wording.
assumptions (4)
  • domain assumption Frozen RAD-DINO-MAIRA-2 ViT patch features are a sufficient basis from which per-concept cross-attention can extract localized visual evidence.
    Methods/Architecture: image encoder remains frozen throughout training; all per-concept visual content must come from its patch tokens.
  • domain assumption GPT-OSS-120B structured outputs are accurate enough to serve as both training supervision and internal evaluation labels.
    Validated on 200 reports by residents (median label macro F1 0.89), but not on the full 32.5k-image test set.
  • ad hoc to paper Empty concept text fields can be imputed with synthetic absence descriptions and unknown labels can be set negative when surrounding pathology coverage exceeds 65%.
    Methods/Data preprocessing and Supplementary 1.1.4; these rules are defined by clinician judgment, not an independent benchmark.
  • domain assumption CheXpert/CheXlocalize classes map to the TUM 19-concept schema, and max-score aggregation for non-one-to-one classes is a valid evaluation proxy.
    Supplementary Table 7; external evaluation is only possible through this mapping, which can inflate or blur per-concept scores.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RadPRISM: Schema-stratified radiology-report supervision for concept-disentangled image representations and visual grounding." pith.science (2026). https://pith.science/paper/OFTDR5A2

@misc{pith2026260800147,
  author       = {Pith},
  title        = {Pith review of: RadPRISM: Schema-stratified radiology-report supervision for concept-disentangled image representations and visual grounding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OFTDR5A2}},
  note         = {Machine review of arXiv:2608.00147}
}
abstract

Vision-language pretraining learns rich medical image representations from radiology reports, but previous model variants commonly operate within a single shared embedding space, so concept-level structure and interpretability must be recovered post hoc, limiting model transparency and, hence, clinical utility. We introduce RadPRISM, which makes a clinician-defined radiology schema a designated stratification axis: an on-premise large language model extracts per-concept text spans from free-text reports, and each clinical concept is aligned in its own dedicated visual subspace, turning concept stratification into direct, top-level alignment supervision. Instantiated on chest radiographs with a 19-concept schema over $203{,}602$ examinations from an internal multi-year archive, RadPRISM improved internal dataset zero-shot classification from $0.717$ (95% CI, $0.710-0.723$) to $0.868$ (95% CI, $0.863-0.872$) macro AUROC over a matched global-alignment baseline, performed on par with the purpose-built CARZero reference in external zero-shot classification while substantially outperforming it (up to 4.3-fold) in pointing-game visual grounding. In addition, a radiologist reader study demonstrated concept-stratified retrieval ability ($0.78$ macro retrieval correctness rate within rank 3), surfacing disentangled descriptive findings that report-level retrieval and fixed-label vocabularies cannot express. RadPRISM yields discriminative, spatially faithful, natively concept-stratified representations shaped by and transparently inspectable by clinicians.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 5 linked inside Pith

  1. [1]

    Tiu, E.et al.Expert-level detection of pathologies from unannotated chest x-ray images via self-supervised learning.Nature Biomedical Engineering6, 1399–1406 (2022)

    2022

  2. [2]

    Zhang, Y ., Jiang, H., Miura, Y ., Manning, C. D. & Lan- glotz, C. P. Contrastive learning of medical visual rep- resentations from paired images and text. InMachine learning for healthcare conference, 2–25 (PMLR, 2022)

    2022

  3. [3]

    E.et al.Generalist foundation models from a multimodal dataset for 3d computed tomography

    Hamamci, I. E.et al.Generalist foundation models from a multimodal dataset for 3d computed tomography. Nature Biomedical Engineering1–19 (2026)

    2026

  4. [4]

    Huang, S.-C., Shen, L., Lungren, M. P. & Yeung, S. GLo- RIA: A multimodal global-local representation learning framework for label-efficient medical image recognition. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision (ICCV), 3942–3951 (2021)

    2021

  5. [5]

    In European Conference on Computer Vision (ECCV), 1–21 (2022)

    Boecking, B.et al.Making the most of text semantics to improve biomedical vision–language processing. In European Conference on Computer Vision (ECCV), 1–21 (2022)

    2022

  6. [6]

    & Xie, W

    Wu, C., Zhang, X., Zhang, Y ., Wang, Y . & Xie, W. Med- KLIP: Medical knowledge enhanced language-image pre-training for X-ray diagnosis. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)(2023)

    2023

  7. [7]

    & Wang, Y

    Zhang, X., Wu, C., Zhang, Y ., Xie, W. & Wang, Y . Knowledge-enhanced visual–language pre-training on chest radiology images.Nature Communications14, 4542 (2023)

    2023

  8. [8]

    & Choi, K

    Park, J., Yoon, B., Kim, S. & Choi, K. Radzero: Similarity-based cross-attention for explainable vision- language alignment in chest x-ray with zero-shot multi- task capability.Advances in Neural Information Process- ing Systems38, 56008–56034 (2026)

    2026

Show all 46 references
  1. [9]

    & Lee, Y

    Imran, M. & Lee, Y . Multimodal vision–language models in medical imaging: A survey of retrieval, interpretability, and trust.IEEE Access(2026)

    2026

  2. [10]

    & Metaxas, D

    Gu, D., Gao, Y ., Zhou, Y ., Zhou, M. & Metaxas, D. Radalign: Advancing radiology report generation with vision-language concept alignment. InInternational Con- ference on Medical Image Computing and Computer- Assisted Intervention, 484–494 (Springer, 2025)

    2025

  3. [11]

    & Metaxas, D

    Gu, D., Gao, Y ., Zhou, M. & Metaxas, D. Anatomy- vlm: A fine-grained vision-language model for medical interpretation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2838– 2847 (2026)

    2026

  4. [12]

    Benchmarking saliency methods for chest x-ray inter- pretation.Nature Machine Intelligence4, 867–878 (2022)

    Saporta, A., Gui, X., Agrawal, A., Pareek, A.et al. Benchmarking saliency methods for chest x-ray inter- pretation.Nature Machine Intelligence4, 867–878 (2022). Available at https://www.nature.com/ articles/s42256-022-00536-x

    2022

  5. [13]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 11137–11146 (2024)

    Lai, H.et al.Carzero: Cross-attention alignment for radiology zero-shot classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 11137–11146 (2024)

    2024

  6. [14]

    gpt-oss-120b & gpt-oss-20b model card (2025)

    OpenAI. gpt-oss-120b & gpt-oss-20b model card (2025). Available at https://arxiv.org/abs/ 2508.10925.2508.10925

    2025 arXiv

  7. [15]

    Computing krippendorff’s alpha- reliability.Departmental Papers (ASC), University of Pennsylvania(2011)

    Krippendorff, K. Computing krippendorff’s alpha- reliability.Departmental Papers (ASC), University of Pennsylvania(2011)

    2011

  8. [16]

    InProceedings of the AAAI Conference on Ar- tificial Intelligence, vol

    Irvin, J.et al.Chexpert: A large chest radiograph dataset with uncertainty labels and expert compari- son. InProceedings of the AAAI Conference on Ar- tificial Intelligence, vol. 33, 590–597 (2019). Avail- able at https://ojs.aaai.org/index.php/ AAAI/article/view/3834

    2019

  9. [17]

    & Zou, J

    Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D. & Zou, J. When and why vision-language models behave like bags-of-words, and what to do about it? InInterna- tional Conference on Learning Representations (ICLR) (2023). Available at https://openreview.net/ forum?id=KRLUvxh8uaX

    2023

  10. [18]

    InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL), 8253–8280 (2022)

    Parcalabescu, L.et al.V ALSE: A task-independent benchmark for vision and language models centered on linguistic phenomena. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL), 8253–8280 (2022)

    2022

  11. [19]

    Moor, M.et al.Foundation models for generalist medical artificial intelligence.Nature616, 259–265 (2023)

    2023

  12. [20]

    & Ouyang, D

    Christensen, M., Vukadinovic, M., Yuan, N. & Ouyang, D. Vision-language foundation model for echocardio- gram interpretation.Nature Medicine30, 1481–1488 (2024)

    2024

  13. [21]

    & Subramanian, D

    Cheng, S. & Subramanian, D. Rethinking radiology report generation: From narrative flow to topic-guided findings. InThe Fourteenth International Conference on Learning Representations(2026)

    2026

  14. [22]

    B., Langlotz, C

    Miura, Y ., Zhang, Y ., Tsai, E. B., Langlotz, C. P. & 12 Jurafsky, D. Improving factual completeness and consis- tency of image-to-text radiology report generation. In Proceedings of the 2021 Conference of the North Ameri- can Chapter of the Association for Computational Lin-...

    2021

  15. [23]

    Yu, F.et al.Evaluating progress in automatic chest x-ray radiology report generation.Patterns4, 100802 (2023)

    2023

  16. [24]

    Information Fusion76, 243–297 (2021)

    Abdar, M.et al.A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information Fusion76, 243–297 (2021)

    2021

  17. [25]

    & Beam, A

    Kompa, B., Snoek, J. & Beam, A. L. Second opinion needed: communicating uncertainty in medical machine learning.npj Digital Medicine4, 4 (2021)

    2021

  18. [26]

    Z., Zhou, Z., Otalora, S

    Luo, H., Shu, S. Z., Zhou, Z., Otalora, S. & Reyes, M. Xbench: A comprehensive benchmark for visual- language explanations in chest radiography.arXiv preprint arXiv:2510.19599(2025)

    2025

  19. [27]

    Wo´ znicki, P.et al.Automatic structuring of radiology reports with on-premise open-source large language mod- els.European Radiology35, 2018–2029 (2025)

    2018

  20. [28]

    & Beam, A

    Ghassemi, M., Oakden-Rayner, L. & Beam, A. L. The false hope of current approaches to explainable artificial intelligence in health care.The Lancet Digital Health3, e745–e750 (2021)

    2021

  21. [29]

    Radiology: Artificial Intelligence2, e190043 (2020)

    Reyes, M.et al.On the interpretability of artificial in- telligence in radiology: Challenges and opportunities. Radiology: Artificial Intelligence2, e190043 (2020)

    2020

  22. [30]

    J.et al.Preparing medical imaging data for machine learning.Radiology295, 4–15 (2020)

    Willemink, M. J.et al.Preparing medical imaging data for machine learning.Radiology295, 4–15 (2020)

    2020

  23. [31]

    & Cheplygina, V

    Varoquaux, G. & Cheplygina, V . Machine learning for medical imaging: methodological failures and recom- mendations for the future.npj Digital Medicine5, 48 (2022)

    2022

  24. [32]

    J., Daniore, P

    Muehlematter, U. J., Daniore, P. & V okinger, K. N. Ap- proval of artificial intelligence and machine learning- based medical devices in the USA and Europe (2015– 20): a comparative analysis.The Lancet Digital Health 3, e195–e203 (2021)

    2015

  25. [33]

    Lin, J.et al.Ct-glip: 3d grounded language-image pre- training with ct scans and radiology reports for full-body scenarios.arXiv preprint arXiv:2404.15272(2024)

    2024

  26. [34]

    Le Guellec, B.et al.Performance of an open-source large language model in extracting information from free-text radiology reports.Radiology: Artificial Intelligence6, e230364 (2024)

    2024

  27. [35]

    P.et al.Data extraction from oncology imaging reports by large language models: A compara- tive accuracy study.medRxiv2025–12 (2025)

    Passweg, L. P.et al.Data extraction from oncology imaging reports by large language models: A compara- tive accuracy study.medRxiv2025–12 (2025)

    2025

  28. [36]

    A coefficient of agreement for nominal scales

    Cohen, J. A coefficient of agreement for nominal scales. Educational and Psychological Measurement20, 37–46 (1960). Available at https://doi.org/10.1177/ 001316446002000104

    1960

  29. [37]

    Zhang, T., Kishore, V ., Wu, F., Weinberger, K. Q. & Artzi, Y . Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations (2020). Available at https://openreview.net/ forum?id=SkeHuCVFDr

    2020

  30. [38]

    K.et al.medbert.de: A german language model for the medical domain.Artificial Intelligence in Medicine157, 102928 (2024)

    Bressem, K. K.et al.medbert.de: A german language model for the medical domain.Artificial Intelligence in Medicine157, 102928 (2024). Available at https: //www.sciencedirect.com/science/ article/pii/S0933365724002130

    2024

  31. [39]

    Krippendorff, K.Content Analysis: An Introduction to Its Methodology(SAGE Publications, 2018), 4 edn

    2018

  32. [40]

    Gwet, K. L. Computing inter-rater reliability and its vari- ance in the presence of high agreement.British Journal of Mathematical and Statistical Psychology61, 29–48 (2008). Available at https://doi.org/10.1348/ 000711006X126600

    2008

  33. [41]

    Oquab, M.et al.Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  34. [42]

    2401.10815

    Pérez-García, F.et al.RAD-DINO: Exploring scalable medical image encoders beyond text supervision (2024). 2401.10815

    2024 arXiv

  35. [43]

    Bannur, S.et al.Maira-2: Grounded radiology report generation.arXiv preprint arXiv:2406.04449(2024)

    2024 arXiv

  36. [44]

    arXiv preprint arXiv:2506.05176(2025)

    Zhang, Y .et al.Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176(2025)

    2025 arXiv

  37. [45]

    J., Janizek, J

    DeGrave, A. J., Janizek, J. D. & Lee, S.-I. Ai for radio- graphic covid-19 detection selects shortcuts over signal. Nature Machine Intelligence3, 610–619 (2021)

    2021

  38. [46]

    correct” or “incorrect

    Holm, S. A simple sequentially rejective multiple test procedure.Scandinavian journal of statistics65–70 (1979). 13 1 Supplementary Information 1.1 Supplementary methods 1.1.1 Structuring and labeling schema definitions Supplementary Table 1 |Label schema and field definitions...

    1979

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.