Pith. sign in

REVIEW 3 major objections 4 minor 182 references

A chest X-ray VLM co-trains classification and grounding heads with report generation, aligns with clinical rewards, and delegates measurements to tools—claiming top results and 43.6-point F1 gains.

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 ↗

A chest X-ray VLM co-trained with classification and grounding heads, tuned with DAPO reinforcement learning, and augmented with deterministic measurement tools outperforms prior radiology VLMs on report generation, VQA, and measurement-based diagnosis.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid, ambitious engineering; the CRIMSON result does not actually support the DAPO-specific claim, but the core system is real and worth refereeing. the 3 major comments →

arxiv 2608.03890 v1 pith:RKNGIOBP submitted 2026-08-04 cs.CV cs.AI

CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement

Mercy Prasanna Ranjit , Anirban Porya , Sathvik Joel , Niharika Vadlamudi , Nikhilesh Chowdary Eathamukkala , Prasanth V V , Abhyuday Kumara Swamy , Pranay Narhari Umredkar
show 3 more authors
Pradeep Narayan Vivek Rajagopal Tanuja Ganu
This is my paper
classification cs.CV cs.AI
keywords chest X-rayvision-language modelauxiliary supervisionreinforcement learningspatial groundingreport generationtool-augmented measurementclinical evaluation
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.

The reading

This paper tries to establish that a chest X-ray vision-language model—a model that reads an image and produces text—can be made clinically useful by combining three things usually kept separate: thresholdable diagnostic classification, spatial grounding, and free-text report generation. CARE-X attaches focal-loss classification heads and a composite-loss bounding-box head to an autoregressive generative backbone and co-trains them with the language-modeling objective; the authors report that this co-training improves the generative outputs themselves, not just the auxiliary predictions. They then apply DAPO, a reinforcement-learning method with task-specific clinical rewards, and report gains across report generation, closed visual question answering, and spatial grounding, including on a held-out clinical metric that was never used as a reward. For diagnoses that depend on measurements, they show that letting a tool-calling VLM perceive the radiograph while deterministic tools compute ratios and thresholds yields large F1 improvements over perception-only inference. If these claims hold, a single radiology VLM could support both flexible reporting and structured, thresholdable, measurement-grounded decision support.

Core claim

CARE-X's central claim is that discriminative and generative objectives reinforce one another rather than compete. The model attaches two focal-loss binary classifiers (abnormality presence; abnormal tube/line placement) and a bounding-box detection head to the shared hidden layer of a 3.8B autoregressive language backbone, and co-trains these heads with the causal language-modeling loss. In an ablation where both model variants are evaluated only through generated text, co-training improves classification F1 and every one of eight grounding metrics, which the paper reads as evidence that structured supervision enriches the shared representation. On top of this, DAPO—a reinforcement-learning

What carries the argument

Two mechanisms carry the argument. The first is auxiliary-head co-training: focal-loss classifiers and a bounding-box head attached to the same hidden representation that feeds autoregressive text generation, trained jointly with the language-modeling loss, so discriminative gradients flow into the shared representation. The isolation experiment compares an identical generative-only model against the co-trained model, both read out in generative mode, to attribute gains to representation enrichment. The second is DAPO (Decoupled Clip and Dynamic sAmpling Policy Optimization), an RL algorithm with asymmetric clipping and token-level loss normalization chosen to prevent entropy collapse and to

Load-bearing premise

The load-bearing premise is that co-training the auxiliary heads improves the shared representation itself; if the observed gains come instead from the particular loss-weight balance or from the extra supervision acting as a generic regularizer, the paper's generalizable claim that discriminative supervision and generation reinforce one another weakens, even though the concrete model may still perform well.

What would settle it

Train the same CARE-X configuration with the auxiliary heads' gradients blocked from reaching the shared image–text backbone (a stop-gradient ablation), keeping data and loss weights identical; if generative report and grounding metrics still improve by the same margin, the claimed representation enrichment is not the cause of the gains.

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

If this is right

  • A single model can serve both deployment modes: free-text reports from generative decoding and structured probabilities or boxes from auxiliary heads, with classification thresholds tunable between high-sensitivity screening and high-specificity confirmation.
  • At deployment, clinicians may not need a separate detection head: DAPO-trained autoregressive grounding is close to or better than the SFT auxiliary head on some benchmarks.
  • Report-quality gains are not purely reward overfitting: CRIMSON, a severity-weighted clinical metric never used as a reward, improves alongside the rewarded metrics on all four report benchmarks.
  • Measurement-dependent diagnoses (cardiomegaly, mediastinal widening, aortic enlargement) are better computed than perceived, so hybrid VLM-plus-tools inference is a practical route to threshold-based screening.
  • The co-training-plus-reward-alignment recipe is put forward as a transferable design principle for other medical domains needing both thresholdable predictions and flexible text generation.

Where Pith is reading between the lines

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

  • A natural next test, beyond this paper, is to fuse the two pipelines: use CARE-X's own grounding head to pick the landmarks that the deterministic measurement tools then measure, rather than a separate general-purpose tool-calling VLM; this could address the residual errors near decision thresholds.
  • The division-of-labour principle should transfer to other imaging tasks where diagnosis is threshold-based and deterministic measurement tools exist—for example echocardiographic dimensions, fetal ultrasound biometry, or retinal vessel calibre.
  • If the mutual-reinforcement claim is a general property, its benefit should be largest when radiology text is scarce but labels are available; the paper's rare-ICU cohort results are consistent with this but were not designed to isolate the mechanism.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper introduces CARE-X, a chest X-ray VLM that co-trains focal-loss classification heads and composite-loss grounding heads with the generative language-modeling objective, then applies DAPO reinforcement learning with task-specific rewards for report generation, closed VQA, and spatial grounding. A second, separate pipeline augments Qwen3-VL-4B-Instruct with deterministic measurement tools for ratio-based diagnoses such as cardiomegaly and mediastinal widening. The paper reports state-of-the-art results on the majority of metrics across four report-generation benchmarks, 94.0% accuracy on ReXVQA, generative spatial decoding approaching or surpassing the auxiliary detection head, and +43.6 pp average F1 for tool-augmented measurement over perception-only inference. It also presents clinical validation on a Narayana Health ICU cohort for rare conditions and on a CT-confirmed outpatient cohort for enlargement conditions.

Significance. If the results hold, the paper makes a useful contribution by integrating three otherwise separate lines of work: auxiliary discriminative heads, reward-aligned RL, and tool-augmented quantitative reasoning. The controlled comparisons are a genuine strength: Appendix B isolates co-training from the presence of auxiliary readouts, the main text compares SFT vs SFT+DAPO, and Appendix F.1 quantifies threshold-selection optimism with cross-validation. The NH clinical evaluations, especially the CT-confirmed enlargement cohort, are valuable and rare. However, two load-bearing pieces of evidence are weaker than the text claims: the CRIMSON result lacks a pre-RL comparator, and the public quantitative-reasoning evaluation shares its label-generation and tool-measurement machinery. Both are fixable, but the stated claims in Sections 5.1.1 and 5.3 currently go beyond what the experiments establish.

major comments (3)
  1. [§5.1.1, Table 2] The claim that CRIMSON "independently confirms that DAPO improvements reflect genuine clinical quality" is not supported by the reported evidence. CRIMSON is listed only for CARE-X-RL; CARE-X-SFT is a dash because the metric became available after the SFT checkpoint was submitted. Since the DAPO report-generation rewards are exactly BERTScore, RadGraph, and GREEN (Table 1, Appendix I.2), every reported SFT-to-RL gain on those metrics is at least partly a reward-optimization artifact. CRIMSON was the only independent check, but without a CRIMSON score for CARE-X-SFT it cannot attribute the CRIMSON result to DAPO rather than to the SFT/auxiliary-co-training foundation. Please compute CRIMSON for the SFT checkpoint, or revise the claim to state that CARE-X-RL achieves strong CRIMSON relative to external baselines, without attributing the gain to DAPO.
  2. [§3.2.1, §K.2, Table 6] The public quantitative-reasoning evaluation is partially circular. Appendix K.2 states that ground-truth labels for the five conditions are derived by applying the CheXStruct thresholds to CheXStruct-computed measurements, and the tool-augmented pipeline loads CXAS segmentation masks and applies the same thresholds. Thus the large F1 gains in Table 6 may reflect alignment between the label-construction machinery and the tool-pipeline machinery rather than clinically validated measurement. The CT-confirmed NH Study 2 is a better external check, but it reports only recall, not F1 or specificity. Please either report full sensitivity/specificity/F1 on an externally labeled cohort, validate the segmentation on an independent set, or explicitly frame Table 6 as demonstrating tool-orchestration capability rather than clinical measurement validity.
  3. [§3.1.2, Appendix B] The paper claims that auxiliary supervision "enriches the shared representation" and that structured prediction and generation "reinforce one another." The Appendix B ablation does show large generative-mode improvements from co-training, which is meaningful. However, the mechanism is not established: the gains could come from the auxiliary losses acting as a regularizer, from improved early-training gradient dynamics, or from loss-weight balancing, rather than from a representational enrichment that transfers to generation. Since this is stated as one of the paper's three generalizable insights, please either soften the mechanistic language or provide representation-level evidence (e.g., probing, feature-similarity analysis, or a loss-weight sensitivity study).
minor comments (4)
  1. [§3.3.1, Table 15] The text says the evaluation cohort is "annotated for five binary conditions," but Table 15 lists six conditions including Tubes & Lines Presence and Tubes & Lines Abnormal Placement. Clarify the counting (e.g., five target pathologies plus the prerequisite presence task).
  2. [§5.4, Table 8] The NH Study 2 comparison reports recall only, because all cases are CT-confirmed positives. The sentence "corroborating our central findings (Table 27)" should be qualified: recall improvement is not the same as the F1 improvement reported on the public benchmarks, and no specificity can be estimated in this cohort.
  3. [§5.1.1] The sentence beginning "Because CRIMSON is entirely independent of our training rewards" should be moved closer to the table footnote and rephrased, since the independence holds for the metric itself but not for the DAPO attribution.
  4. [Table 6, DA row] The descending-aorta row has only 7 test samples and the footnote already cautions interpretation; please also carry this caveat into the main-text summary sentence that reports the +43.6 pp average.

Circularity Check

2 steps flagged

Tool-augmented measurement evaluation is partially self-definitional (ground truth generated by the same tools being evaluated), and the report-generation RL gains are measured on the exact rewards DAPO optimizes, with the only held-out check lacking a pre-RL baseline.

specific steps
  1. self definitional [Section 3.2.1 and Appendix K.1, Table 24; results in Table 6]
    "Ground-truth labels for each sample are derived by applying these thresholds to the CheXStruct-computed measurements. (Appendix K.1) ... classifying as abnormal when CTR≥0.50 (PA) or≥0.55 (AP). (Section 3.2.1)"

    The quantitative-reasoning evaluation defines the ground-truth label by running CheXStruct's deterministic measurement tools and thresholds. The tool-augmented pipeline under test invokes those same CheXStruct tools (measure cardiac width, measure thoracic width, compute ctr, etc.) and applies the same thresholds. A correct execution of the tool chain therefore trivially reproduces the label. The +43.6 pp F1 gain over perception-only is thus partly a tautology: it measures how faithfully the VLM calls the oracle tool chain, not whether the measurements are clinically valid against an independent standard. The NH Study-2 CT-confirmed cohort is separate, but the headline Table 6 result rests on tool-derived labels.

  2. fitted input called prediction [Section 3.1.2 (DAPO Rewards) and Section 5.1.1; Table 2]
    "Report Generation: We use BERTScore (Zhang et al., 2020), RadGraph (Delbrouck et al., 2022), and GREEN (Ostmeier et al., 2024) as the rewards."

    The report-generation metrics on which CARE-X-RL is claimed to improve and reach SOTA — BERTScore, RadGraph, GREEN — are exactly the three reward signals used in the DAPO curriculum (Table 1, Appendix I.2). The SFT→RL gains on those columns are therefore the direct result of optimizing those same metrics, not independent evidence of clinical-quality improvement. The paper attempts to break this with CRIMSON, a metric never used as a reward, but CRIMSON is reported only for CARE-X-RL and not for CARE-X-SFT, so it cannot isolate the DAPO increment; the claim that CRIMSON 'independently confirms that DAPO improvements reflect genuine clinical quality' is missing the pre-RL comparator needed to support it.

full rationale

The core CARE-X-RL contributions are not fundamentally circular: the auxiliary-head co-training ablation in Appendix B compares a generative-only backbone with a co-trained backbone under identical generative decoding, which is a genuine controlled experiment; DAPO is an externally published method; and the SOTA comparisons against MedGemma, MedVersa, and CheXOne on held-out test sets use external baselines. There is no load-bearing self-citation or imported uniqueness theorem. However, two significant reductions weaken the paper's strongest claims. First, the tool-augmented measurement experiments define ground truth with the same CheXStruct measurement tools and thresholds that the tool-calling pipeline invokes, making the headline +43.6 pp F1 improvement partly self-confirming by construction. Second, the report-generation RL results are reported on the exact BERTScore/RadGraph/GREEN rewards that DAPO optimizes, so those gains are not independent evidence of clinical fidelity; the only held-out metric, CRIMSON, was not scored for CARE-X-SFT, leaving no baseline against which the DAPO-specific improvement can be verified. These are partial circularities, not a complete collapse of the derivation, so the score is 6 rather than higher.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

CARE-X introduces no new physical or computational entities, but it does introduce several hand-set quantities: per-condition decision thresholds, spatial loss weights, and a reward curriculum. The threshold tuning on the evaluation cohort is the most notable free-parameter issue; the paper partially mitigates it with 3-fold CV. The heaviest unvalidated assumption is the CXAS segmentation quality on the NH cohort, which the authors themselves flag.

free parameters (4)
  • Per-condition decision thresholds for NH ICU conditions = Fracture 0.50, Med. Shift 0.55, Pneumoperitoneum 0.65, Pneumothorax 0.65, Tubes & Lines 0.40
    Tuned on the same 1,047 radiographs used for evaluation (Section 3.3.1, Table 7). The paper quantifies selection optimism via 3-fold CV in Appendix F.1, showing small differences, so this is a free parameter with a partially addressed caveat.
  • Composite spatial loss weights (GIoU 5.0, mIoU 0.8, L1 7.0, focal 2.0) = Weights reported in Appendix C
    The grounding head loss weighting is chosen by hand, not derived from first principles, and the grounding results depend on it.
  • Reward curriculum schedule and weights = Stage 1: BERTScore 0.5, RadGraph 0.5; Stage 2: 0.25/0.25/0.5 GREEN; Stage 3: 0.5/0.5
    The three-stage reward curriculum and its stage boundaries/weights are design choices selected to avoid reward hacking; the reported results depend on them.
  • AP/PA view-dependent thresholds = CM: CTR 0.50 (PA), 0.55 (AP); MW: MCR 0.25 (PA), 0.33 (AP)
    These are clinical thresholds from the literature used by the CheXStruct/measurement tools, not fit to data. They are included as parameters the central measurement claim depends on, but they are externally sourced.
axioms (5)
  • domain assumption Focal loss and composite spatial loss provide useful gradient signal for the generative decoder when trained jointly.
    The co-training benefit (Appendix B) is an empirical result, not a derived property. It is also the paper's generalizable insight #1.
  • domain assumption The reward models (BERTScore, RadGraph, GREEN, CRIMSON) correlate with clinical quality sufficiently to serve as optimization targets and validation.
    The RL claims rest on reward models being faithful proxies; the paper itself notes in Section 7 that 'DAPO rewards are imperfect proxies for clinical quality.'
  • domain assumption CXAS segmentation masks are accurate enough to support the measurement tools.
    The tool thresholds operate on widths/areas derived from CXAS masks. The paper admits in Section 7 that it did not independently validate CXAS quality on the NH cohort.
  • domain assumption The GPT-5.1-extracted Tubes & Lines placement labels used in training are accurate.
    Appendix E.1 uses GPT-5.1 to extract device status from radiology reports without reporting a validation of the extraction quality on the training data.
  • standard math DAPO is an appropriate RL algorithm for multi-task radiology optimization.
    The paper applies DAPO as an off-the-shelf method from Yu et al. 2025, citing its Clip-Higher and token-level normalization properties.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement." pith.science (2026). https://pith.science/paper/RKNGIOBP

@misc{pith2026260803890,
  author       = {Pith},
  title        = {Pith review of: CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RKNGIOBP}},
  note         = {Machine review of arXiv:2608.03890}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

A clinically useful chest X-ray system must go beyond fluent report generation: it should classify findings with tunable decision thresholds, localize them spatially, and derive the anatomical measurements upon which many diagnoses depend. Today's Vision-Language Models (VLMs) treat these as separate problems, if they address them at all, leaving a gap between what radiologists need and what generative models provide. We introduce CARE-X, a chest X-ray VLM that narrows this gap by unifying auxiliary discriminative supervision with reward-aligned generation. CARE-X augments its generative backbone with focal-loss classification and composite-loss grounding heads, co-trained alongside the language-modeling objective. This auxiliary supervision produces discriminative diagnostic predictions with tunable decision thresholds and precise spatial localization while also improving report quality, providing evidence that structured prediction and generation reinforce one another. Building on this foundation, Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) leverages task-specific reward signals for report generation, visual question answering (VQA), and spatial grounding, directly optimizing the clinical quality metrics that matter in practice. The result is state-of-the-art performance on the majority of metrics across four report-generation benchmarks, 94.0% VQA accuracy on ReXVQA (+6.0 pp over the next-best baseline), and generative spatial decoding that reaches near parity with dedicated detection heads. Separately, to address measurement-dependent diagnoses, we couple Qwen3-VL-4B-Instruct with native tool-calling capabilities for invoking deterministic measurement tools, while retaining full visual access to the image. This hybrid inference yields +43.6 pp average F1 over perception-only baselines across five measurement-dependent conditions.

Figures

Figures reproduced from arXiv: 2608.03890 by Abhyuday Kumara Swamy, Anirban Porya, Mercy Prasanna Ranjit, Niharika Vadlamudi, Nikhilesh Chowdary Eathamukkala, Pradeep Narayan, Pranay Narhari Umredkar, Prasanth V V, Sathvik Joel, Tanuja Ganu, Vivek Rajagopal.

Figure 1
Figure 1. Figure 1: The CARE-X model. (Left) Supervised fine-tuning with three task-specific heads for classification, grounding, and report generation/VQA. (Right) DAPO with task-specific rewards. See Section 3.1.1 for details. Decoupled Clip and Dynamic sAmpling Policy Optimization (DAPO). In the second phase, we apply DAPO (Yu et al., 2025) to directly optimize task-specific reward signals that capture clinical quality dim… view at source ↗
Figure 2
Figure 2. Figure 2: Quantitative Reasoning Inference Pipeline [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: NH Dataset Prevalence Distribution for Inpatient & ICU Conditions [PITH_FULL_IMAGE:figures/full_fig_p030_3.png] view at source ↗

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

182 extracted references · 24 canonical work pages · 6 internal anchors

  1. [1]

    2026 , eprint=

    A Reasoning-Enabled Vision-Language Foundation Model for Chest X-ray Interpretation , author=. 2026 , eprint=

    2026

  2. [2]

    2025 , url=

    Ranjit, Mercy Prasanna and Porya, Anirban and Srivastav, Shaury and Vadlamudi, Niharika and Eathamukkala, Nikhilesh Chowdary and Udyavar, Shashank and Kumar, Rahul and Ganu, Tanuja , booktitle=. 2025 , url=

    2025

  3. [3]

    2025 , eprint=

    Deperrois, Nicolas and Matsuo, Hidetoshi and Ruip. 2025 , eprint=

    2025

  4. [4]

    2026 , eprint=

    Scaling medical imaging report generation with multimodal reinforcement learning , author=. 2026 , eprint=

    2026

  5. [5]

    2025 , eprint=

    Enhancing Radiology Report Generation and Visual Grounding using Reinforcement Learning , author=. 2025 , eprint=

    2025

  6. [6]

    GREEN : Generative Radiology Report Evaluation and Error Notation

    Ostmeier, Sophie and Xu, Justin and Chen, Zhihong and Varma, Maya and Blankemeier, Louis and Bluethgen, Christian and Michalson, Arne Edward and Moseley, Michael and Langlotz, Curtis and Chaudhari, Akshay S and Delbrouck, Jean-Benoit. GREEN : Generative Radiology Report Evaluation and Error Notation. Findings of the Association for Computational Linguisti...

    doi:10.18653/v1/2024.findings-emnlp.21 2024

  7. [7]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Sigmoid loss for language image pre-training , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  8. [8]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=. 2022 , url=

    2022

  9. [9]

    Weinberger and Yoav Artzi , booktitle=

    Tianyi Zhang and Varsha Kishore and Felix Wu and Kilian Q. Weinberger and Yoav Artzi , booktitle=. 2020 , url=

    2020

  10. [10]

    Improving the Factual Correctness of Radiology Report Generation with Semantic Rewards

    Delbrouck, Jean-Benoit and Chambon, Pierre and Bluethgen, Christian and Tsai, Emily and Almusa, Omar and Langlotz, Curtis. Improving the Factual Correctness of Radiology Report Generation with Semantic Rewards. Findings of the Association for Computational Linguistics: EMNLP 2022. 2022. doi:10.18653/v1/2022.findings-emnlp.319

    doi:10.18653/v1/2022.findings-emnlp.319 2022

  11. [11]

    Topol and Pranav Rajpurkar , title =

    Hong-Yu Zhou and Julián Nicolás Acosta and Subathra Adithan and Suvrankar Datta and Eric J. Topol and Pranav Rajpurkar , title =. NEJM AI , volume =. 2026 , doi =. https://ai.nejm.org/doi/pdf/10.1056/AIoa2500595 , abstract =

    doi:10.1056/aioa2500595 2026

  12. [12]

    2503.14476 , archivePrefix=

    Qiying Yu and Zheng Zhang and Ruofei Zhu and Yufeng Yuan and Xiaochen Zuo and Yu Yue and Weinan Dai and Tiantian Fan and Gaohong Liu and Lingjun Liu and Xin Liu and Haibin Lin and Zhiqi Lin and Bole Ma and Guangming Sheng and Yuxuan Tong and Chi Zhang and Mofan Zhang and Wang Zhang and Hang Zhu and Jinhua Zhu and Jiaze Chen and Jiangjie Chen and Chengyi W...

    Pith/arXiv arXiv

  13. [13]

    and Miller, Josh and Huang, Ouwen and Rajpurkar, Pranav , booktitle =

    Zhang, Xiaoman and Zhou, Hong-Yu and Yang, Xiaoli and Banerjee, Oishi and Acosta, Juli\'an N. and Miller, Josh and Huang, Ouwen and Rajpurkar, Pranav , booktitle =. 2025 , editor =

    2025

  14. [14]

    GitHub repository , howpublished =

    Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallouédec , title =. GitHub repository , howpublished =. 2022 , publisher =

    2022

  15. [15]

    2026 , eprint=

    CRIMSON: A Clinically-Grounded LLM-Based Metric for Generative Radiology Report Evaluation , author=. 2026 , eprint=

    2026

  16. [16]

    2025 , eprint =

    Lee, Hyungyung and Choi, Geon and Lee, Jung-Oh and Yoon, Hangyul and Hong, Hyuk Gi and Choi, Edward , booktitle =. 2025 , eprint =

    2025

  17. [17]

    2602.23276 , archivePrefix =

    Lee, Hyungyung and Yoon, Hangyul and Choi, Edward , year =. 2602.23276 , archivePrefix =

    work page arXiv

  18. [18]

    arXiv preprint arXiv:2306.03934 , year =

    Accurate Fine-Grained Segmentation of Human Anatomy in Radiographs via Volumetric Pseudo-Labeling , author =. arXiv preprint arXiv:2306.03934 , year =

    Pith/arXiv arXiv

  19. [19]

    Bai, Jinze and Bai, Shuai and Yang, Shusheng and Wang, Shijie and Tan, Sinan and Wang, Peng and Lin, Junyang and Zhou, Chang and Zhou, Jingren , journal =

  20. [20]

    and Gonzalez, Joseph and Zhang, Hao , booktitle =

    Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Ying, Ying and Liang, Eric and Stoica, Ion and Xing, Eric P. and Gonzalez, Joseph and Zhang, Hao , booktitle =. Efficient Memory Management for Large Language Model Serving with

  21. [21]

    Guyon and A

    I. Guyon and A. Elisseeff. An Introduction to Variable and Feature Selection. JMLR

  22. [22]

    Guyon and C

    I. Guyon and C. Aliferis and A. Elisseeff , title =

  23. [23]

    and Mosquera, C

    Gaggion, N. and Mosquera, C. and Aineseder, M. and Mansilla, L. and Milone, D. and Ferrante, E. , title =. 2023 , howpublished =

    2023

  24. [24]

    2401.10815 , archivePrefix=

    P. 2401.10815 , archivePrefix=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2407.10671 , year=

    Qwen2 technical report , author=. arXiv preprint arXiv:2407.10671 , year=

    Pith/arXiv arXiv

  26. [26]

    Improving the Factual Correctness of Radiology Report Generation with Semantic Rewards

    Improving the factual correctness of radiology report generation with semantic rewards , author=. arXiv preprint arXiv:2210.12186 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    Findings of the Association for Computational Linguistics ACL 2024 , pages=

    Radgraph-xl: A large-scale expert-annotated dataset for entity and relation extraction from radiology reports , author=. Findings of the Association for Computational Linguistics ACL 2024 , pages=

    2024

  28. [29]

    and others , title =

    Jackson, Valerie P. and others , title =. Journal of the American College of Radiology , volume =. 2009 , doi =

    2009

  29. [30]

    2025 , url =

    OpenAI , title =. 2025 , url =

    2025

  30. [31]

    arXiv preprint arXiv:2411.15122 , year=

    Rexrank: A public leaderboard for ai-powered radiology report generation , author=. arXiv preprint arXiv:2411.15122 , year=

    Pith/arXiv arXiv

  31. [32]

    1895 , publisher=

    Ueber eine neue Art von Strahlen , author=. 1895 , publisher=

  32. [33]

    Boosting Radiology Report Generation by Infusing Comparison Prior

    Kim, Sanghwan and Nooralahzadeh, Farhad and Rohanian, Morteza and Fujimoto, Koji and Nishio, Mizuho and Sakamoto, Ryo and Rinaldi, Fabio and Krauthammer, Michael. Boosting Radiology Report Generation by Infusing Comparison Prior. The 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks. 2023. doi:10.18653/v1/2023.bionlp-1.4

    work page doi:10.18653/v1/2023.bionlp-1.4 2023

  33. [34]

    Nature Biomedical Engineering , pages=

    A vision--language foundation model for the generation of realistic chest x-ray images , author=. Nature Biomedical Engineering , pages=. 2024 , publisher=

    2024

  34. [35]

    arXiv preprint arXiv:2405.07960 , year=

    AgentClinic: a multimodal agent benchmark to evaluate AI in simulated clinical environments , author=. arXiv preprint arXiv:2405.07960 , year=

    Pith/arXiv arXiv

  35. [36]

    Proceedings of the 18th NTCIR Conference on Evaluation of Information Access Technologies , YEAR =

    Overview of the NTCIR-18 HIDDEN-RAD Task: Hidden Causality Inclusion in Radiology Report Generation , author=. Proceedings of the 18th NTCIR Conference on Evaluation of Information Access Technologies , YEAR =

  36. [37]

    2024 , organization =

    Llama 3.3 70B: Model Card , author =. 2024 , organization =

    2024

  37. [38]

    2506.04353 , archivePrefix=

    Ankit Pal and Jung-Oh Lee and Xiaoman Zhang and Malaikannan Sankarasubbu and Seunghyeon Roh and Won Jung Kim and Meesun Lee and Pranav Rajpurkar , year=. 2506.04353 , archivePrefix=

    Pith/arXiv arXiv

  38. [39]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  39. [40]

    2019 , url =

    Johnson, Alistair EW and Pollard, Tom J and Greenbaum, Nathaniel R and Lungren, Matthew P and Deng, Chih-ying and Peng, Yifan and Lu, Zhiyong and Mark, Roger G and Berkowitz, Seth J and Horng, Steven , journal=. 2019 , url =

    2019

  40. [41]

    MAIRA-Seg: Enhancing Radiology Report Generation with Segmentation-Aware Multimodal Large Language Models

    MAIRA-Seg: Enhancing Radiology Report Generation with Segmentation-Aware Multimodal Large Language Models , author=. arXiv preprint arXiv:2411.11362 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [42]

    arXiv preprint arXiv:2409.12191 , year=

    Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    Pith/arXiv arXiv

  42. [43]

    Hidden-Rad Evaluation Scheme of NTCIR-18 , year =

    Cho, Yousang and Choi, Key-Sun and. Hidden-Rad Evaluation Scheme of NTCIR-18 , year =

  43. [44]

    Diagnostics , volume =

    Enhancing Radiologist Productivity with Artificial Intelligence in Magnetic Resonance Imaging (MRI): A Narrative Review , author =. Diagnostics , volume =. 2025 , publisher =. doi:10.3390/diagnostics15091146 , url =

    work page doi:10.3390/diagnostics15091146 2025

  44. [45]

    Medical image analysis , volume=

    Bustos, Aurelia and Pertusa, Antonio and Salinas, Jose-Maria and de la Iglesia-Vay. Medical image analysis , volume=. 2020 , publisher=

    2020

  45. [46]

    arXiv preprint arXiv:2411.05085 , year=

    PadChest-GR: A Bilingual Chest X-ray Dataset for Grounded Radiology Report Generation , author=. arXiv preprint arXiv:2411.05085 , year=

    Pith/arXiv arXiv

  46. [47]

    arXiv preprint arXiv:2405.03595 , year=

    GREEN: Generative Radiology Report Evaluation and Error Notation , author=. arXiv preprint arXiv:2405.03595 , year=

    Pith/arXiv arXiv

  47. [48]

    European conference on computer vision , pages=

    Making the most of text semantics to improve biomedical vision--language processing , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  48. [49]

    European Conference on Computer Vision , pages=

    Any-resolution training for high-resolution image synthesis , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  49. [50]

    European journal of radiology , volume=

    Radiologist burnout: trends in medical imaging utilization under the national health insurance system with the universal code bundling strategy in an academic tertiary medical centre , author=. European journal of radiology , volume=. 2022 , publisher=

    2022

  50. [51]

    Proceedings of the IEEE conference on computer vision and pattern recognition , year=

    You only look once: Unified, real-time object detection , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , year=

  51. [52]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Faster R-CNN: Towards real-time object detection with region proposal networks , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2016 , publisher=

    2016

  52. [53]

    Nature , volume=

    Large language models encode clinical knowledge , author=. Nature , volume=. 2023 , publisher=

    2023

  53. [54]

    NEJM AI , volume=

    Towards generalist biomedical AI , author=. NEJM AI , volume=. 2024 , publisher=

    2024

  54. [55]

    Advances in Neural Information Processing Systems , year=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , year=

  55. [56]

    arXiv preprint arXiv:2005.14165 , year=

    Language models are few-shot learners , author=. arXiv preprint arXiv:2005.14165 , year=

    Pith/arXiv arXiv 2005

  56. [58]

    OpenAI blog , volume=

    Language models are unsupervised multitask learners , author=. OpenAI blog , volume=

  57. [59]

    arXiv preprint arXiv:2109.01652 , year=

    Finetuned language models are zero-shot learners , author=. arXiv preprint arXiv:2109.01652 , year=

    Pith/arXiv arXiv

  58. [60]

    arXiv preprint arXiv:2303.08774v6 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774v6 , year=

    Pith/arXiv arXiv

  59. [61]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: a family of highly capable multimodal models , author=. arXiv preprint arXiv:2312.11805 , year=

    Pith/arXiv arXiv

  60. [62]

    International conference on machine learning , pages=

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  61. [63]

    , author=

    Building and better understanding vision-language models: insights and future directions. , author=. 2024 , eprint=

    2024

  62. [64]

    Nature Biomedical Engineering , volume=

    Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning , author=. Nature Biomedical Engineering , volume=. 2022 , publisher=

    2022

  63. [65]

    BMC Medical Informatics and Decision Making , volume=

    BarlowTwins-CXR: enhancing chest X-ray abnormality localization in heterogeneous data with cross-domain self-supervised learning , author=. BMC Medical Informatics and Decision Making , volume=. 2024 , publisher=

    2024

  64. [66]

    2022 IEEE international conference on bioinformatics and biomedicine (BIBM) , pages=

    Research on chest abnormality detection based on improved YOLOv7 algorithm , author=. 2022 IEEE international conference on bioinformatics and biomedicine (BIBM) , pages=. 2022 , organization=

    2022

  65. [67]

    arXiv preprint arXiv:2310.07704 , year=

    Ferret: Refer and ground anything anywhere at any granularity , author=. arXiv preprint arXiv:2310.07704 , year=

    Pith/arXiv arXiv

  66. [68]

    arXiv preprint arXiv:2404.07973 , year=

    Ferret-v2: An Improved Baseline for Referring and Grounding with Large Language Models , author=. arXiv preprint arXiv:2404.07973 , year=

    Pith/arXiv arXiv

  67. [69]

    arXiv preprint arXiv:2212.10560 , year=

    Self-instruct: Aligning language models with self-generated instructions , author=. arXiv preprint arXiv:2212.10560 , year=

    Pith/arXiv arXiv

  68. [70]

    Foundations and Trends

    Multimodal foundation models: From specialists to general-purpose assistants , author=. Foundations and Trends. 2024 , publisher=

    2024

  69. [71]

    arXiv preprint arXiv:2309.11998 , year=

    Lmsys-chat-1m: A large-scale real-world llm conversation dataset , author=. arXiv preprint arXiv:2309.11998 , year=

    Pith/arXiv arXiv

  70. [72]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Vila: On pre-training for visual language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  71. [73]

    Advances in Neural Information Processing Systems , volume=

    Judging llm-as-a-judge with mt-bench and chatbot arena , author=. Advances in Neural Information Processing Systems , volume=

  72. [74]

    arXiv preprint arXiv:2211.05719 , year=

    Mmdialog: A large-scale multi-turn dialogue dataset towards multi-modal open-domain conversation , author=. arXiv preprint arXiv:2211.05719 , year=

    Pith/arXiv arXiv

  73. [75]

    2015 IEEE international conference on bioinformatics and biomedicine (BIBM) , pages=

    Foreign object detection in chest X-rays , author=. 2015 IEEE international conference on bioinformatics and biomedicine (BIBM) , pages=. 2015 , organization=

    2015

  74. [76]

    ChEX: Interactive Localization and Region Description in Chest X-rays

    ChEX: Interactive Localization and Region Description in Chest X-rays , author=. arXiv preprint arXiv:2404.15770 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  75. [77]

    Image and Vision Computing , volume=

    Weighted boxes fusion: Ensembling boxes from different object detection models , author=. Image and Vision Computing , volume=. 2021 , publisher=

    2021

  76. [78]

    arXiv preprint arXiv:2004.09167 , year=

    CheXbert: combining automatic labelers and expert annotations for accurate radiology report labeling using BERT , author=. arXiv preprint arXiv:2004.09167 , year=

    Pith/arXiv arXiv 2004

  77. [79]

    Langlotz , year=

    Pierre Chambon and Jean-Benoit Delbrouck and Thomas Sounack and Shih-Cheng Huang and Zhihong Chen and Maya Varma and Steven QH Truong and Chu The Chuong and Curtis P. Langlotz , year=. 2405.19538 , archivePrefix=

    Pith/arXiv arXiv

  78. [80]

    arXiv preprint arXiv:2408.16213 , year=

    M4CXR: Exploring Multi-task Potentials of Multi-modal Large Language Models for Chest X-ray Interpretation , author=. arXiv preprint arXiv:2408.16213 , year=

    Pith/arXiv arXiv

  79. [81]

    WoLF: Wide-scope Large Language Model Framework for CXR Understanding

    WoLF: Large Language Model Framework for CXR Understanding , author=. arXiv preprint arXiv:2403.15456 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  80. [82]

    International Conference on Machine Learning , pages=

    Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.