Pith. sign in

REVIEW 4 major objections 5 minor 131 references

Interpreting Chest X-rays Like a Radiologist: A Benchmark with Clinical Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Chest X-ray AI gets an eight-stage clinical reasoning benchmark and a model that follows it.

desk verdict Large 8-stage CXR VQA dataset with real potential, but the generalization claim collapses because the 'external' benchmarks are in the training mixture. read the letter →

arxiv 2505.23143 v1 pith:KBKTMIP7 submitted 2025-05-29 cs.CV

Jinquan Guan , Qi Chen , Lizhou Liang , Yuhang Liu , Vu Minh Hieu Phan , Minh-Son To , Jian Chen , Yutong Xie This is my paper
classification cs.CV
keywords chestX-rayclinicalreasoningvisualquestionansweringvision-languagemodelmulti-stagemedicalbenchmarkradiologyreportgeneration
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

The paper aims to show that chest X-ray interpretation by AI should mirror the sequential reasoning of radiologists, not just map image to answer. To that end, it builds CXRTrek, a dataset of 428,966 X-rays with over 11 million question-answer pairs organized into eight diagnostic stages, from image validation to final report. It then trains CXRTrekNet, a vision-language model that consumes previous stages' answers as context, and reports that this model outperforms existing medical VLLMs on the benchmark and transfers well to five external datasets. A reader would care because the claim, if true, suggests a path toward more traceable and clinically aligned AI radiology assistants.

What carries the argument

The load-bearing mechanism is the eight-stage clinical reasoning flow, a fixed sequence of stages (X-ray validation, findings recognition, attribute recognition, relationships, image comparison, future risk, diagnosis advice, summarizing) paired with a stage-wise training objective. At each stage i, the model generates answer Âᵢ conditioned on the current question Qᵢ, the images X, and accumulated context Ctxᵢ = {Q₁, Â₁, …, Qᵢ₋₁, Âᵢ₋₁}, so earlier conclusions feed later reasoning in an autoregressive, teacher-forced loop.

What would settle it

Have a panel of radiologists independently annotate a random sample of the 200 CXRTrek test studies for all eight stages, then compute per-stage agreement with the dataset's auto-generated answers; if agreement falls substantially below the reported BERTScore of 92.53—especially on relationships, future risk, and advice—the benchmark's validity as a measure of clinical reasoning collapses.

Watch

Extended reading notes

Core claim

The central claim is that explicitly modeling the stage-by-stage diagnostic flow—where each answer conditions on prior answers and the current question—yields substantially better chest X-ray interpretation than joint, single-stage training. The paper reports that CXRTrekNet, trained on CXRTrek with this sequential teacher-forced objective, achieves an average score of 69.66% across the eight stages, versus 33.38% for the strongest baseline (CheXagent), and that even training on 1% of the data with stage-wise reasoning outperforms a model trained on 100% of the data without it (60.01% vs. 52.86%). It also claims superior generalization on VQA, classification, and report generation tasks across five external datasets, attributing the gains to both the dataset's fine-grained supervision and the model's context-aware reasoning design.

Load-bearing premise

The automatically generated labels, produced by a language model and simple rules from text reports, are clinically accurate enough to serve as reliable ground truth for both training and evaluating clinical reasoning.

Editorial extensions

If this is right

  • If the claims hold, medical VLLMs trained on multi-stage reasoning data should generalize better to diverse radiology tasks than models trained on monolithic image-to-text pairs.
  • The explicit context chain could make model errors more traceable, since a misstep at an early stage can be identified as the cause of downstream mistakes.
  • The benchmark provides a standardized way to evaluate not just final answers but intermediate reasoning stages, which may become a template for other imaging modalities.
  • The reported 1%-data result suggests that stage-wise structure, not raw data volume, drives much of the performance gain, a hypothesis worth testing on other medical domains.

Reading between the lines

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

  • The automatic annotation pipeline, built on Llama 3.2 8B extraction with a medical vocabulary filter, likely injects label noise that propagates into the benchmark and the trained model; the paper's BERTScore validation on 100 samples is too small to rule this out.
  • A direct consequence, not stated by the authors, is that the benchmark's value for clinical deployment hinges on whether the auto-annotated stages (especially relationships, future risk, and advice) match expert opinion; a radiologist-reviewed subset of the test set would settle this.
  • The eight-stage flow could be repurposed as a prompting template for general medical imaging models, potentially improving their interpretability without retraining, though the paper does not test this.
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

4 major / 5 minor

Summary. The paper introduces CXRTrek, a large multi-stage visual question answering dataset for chest X-ray interpretation, spanning 428,966 studies and over 11 million question-answer pairs across 8 sequential diagnostic stages. It also proposes CXRTrekNet, a vision-language model initialized from CheXagent and fine-tuned with LoRA to carry contextual information from previous diagnostic stages into subsequent ones. The authors claim that CXRTrekNet consistently outperforms existing medical VLLMs on the CXRTrek test set and generalizes across five external datasets covering VQA, classification, and report generation. The paper includes extensive construction details, a hybrid LLM/rule-based annotation framework, ablations on stage-wise training and data scale, and appendices with evaluation and training protocols.

Significance. If the evaluation were clean, CXRTrek would be a substantial resource: it is one of the largest structured CXR-reasoning benchmarks, with explicit stage dependencies, four answer formats, and an accompanying reasoning-aware model. The 8-stage design is grounded in radiology guidelines, and the annotation framework, including the medical vocabulary and synonym dictionary, is described in unusual detail. The central empirical claims, however, are currently not established because the training mixture overlaps with the datasets used for 'external' evaluation, and because the baseline comparisons do not clearly control for fine-tuning exposure. The resource itself is valuable and the core idea is defensible, but the paper needs additional experiments and reporting before the claimed generalization and superiority can be accepted.

major comments (4)
  1. [Appendix A.2 / Figure A] The training mixture shown in Figure A explicitly includes MIMIC-CXR (21.81%), VINDR (13.96%), VQA-RAD (1.78%), IU-Xray (0.99%), and rad-restruct (1.67%), which are exactly the five datasets later called 'external' in Tables 2 and 3. The paper never states which splits of these datasets are used in the mixture. If any training samples come from the evaluation splits of these datasets, the abstract's claim of 'superior generalization across multiple tasks on five diverse external datasets' is not supported. Even if the splits are non-overlapping by study identifier, the model has been trained on the same label distributions and domains as the evaluation sets, so the term 'external' is misleading. Please report exact split provenance, confirm no evaluation study appears in the mixture, and ideally re-run the external evaluation on datasets or tasks that were not part of the training mixture.
  2. [Section 5.1 / Tables 2 and 3] The text says the external benchmarks are evaluated 'via fine-tuning', but it is not stated whether the baseline models (LLaVA-Med, MiniGPT-Med, CheXagent, XrayGPT) are fine-tuned on those datasets or evaluated off-the-shelf. The prompts in Appendix A.1 and the comparison protocols suggest some baselines may be zero-shot or prompted without training. If CXRTrekNet is fine-tuned on each target dataset while baselines are not, the comparison conflates model quality with training exposure. Please specify the fine-tuning setup for every model on every external dataset, including the number of steps, LoRA settings, and data splits used, or add a controlled comparison where all models receive the same fine-tuning data.
  3. [Section 3.2 / Appendix C / Table D] The supervision for most of the 8 stages is generated by LLaMA 3.2 8B extraction and rule-based modules, and the reported validation is a BERTScore of 92.53 against human annotations on only 100 samples. The 200-sample CXRTrek test set is described as 'human verified', but no expert clinical review, annotator qualifications, or inter-annotator agreement is reported. Because these automatically generated labels are used both as training targets for CXRTrekNet and as ground truth for benchmarking, errors in entity, attribute, relationship, or advice extraction propagate directly into the model's behavior and the evaluation. Please provide a clinician-reviewed error analysis on the test set, report per-field agreement beyond BERTScore, and quantify how annotation errors affect the headline stage-wise numbers.
  4. [Table 1 / Table 4] The main benchmark comparison in Table 1 shows CXRTrekNet against baselines that were not trained on CXRTrek. The large gap is therefore unsurprising and does not isolate whether the improvement comes from the new dataset, the stage-wise context mechanism, or simply from in-distribution training. The ablation in Table 4 partially addresses this by comparing joint and stage-wise training, but a direct comparison with baselines fine-tuned on the same CXRTrek training split (with and without stage-wise context) is needed to support the claim that CXRTrekNet's design is responsible for the gains. Please add such a controlled experiment.
minor comments (5)
  1. [Appendix A.2 / Figure A] The figure caption 'NUM COUNT (Horizontally Aligned Labels)' is unclear, and the axis has no unit label; the numbers appear to be percentages, but this should be stated explicitly.
  2. [Throughout] There are several typos and inconsistent names: 'Software Engneering' in the author block, 'orignal dataset' in Section 3.2, 'VINDR' vs 'VinDr-CXR', and 'CheXpert-plus' vs 'CheXpert plus'. Please proofread and standardize terminology.
  3. [Checklist items 5 and 13] The checklist states that code, data, and model are not yet open, while the abstract and repository link suggest they are available. Please clarify the current availability status and the expected release terms.
  4. [Section 5 / Checklist item 7] No error bars or multiple-seed results are reported, which is acknowledged in the checklist. For the headline claim of consistent superiority, at least a few seeds or bootstrap confidence intervals on the main tables would substantially strengthen the report.
  5. [Appendix A.1 / Table A] Several Multi-Stage baseline scores are reported as negative BERTScore values (e.g., CheXagent Stage-3 Open BS -4.84). BERTScore can be negative in principle, but such values are unusual and should be explained or verified, as they affect the credibility of the evaluation pipeline.

Circularity Check

2 steps flagged · score 6.0 of 10

External-generalization claim is undermined because the five 'external' evaluation benchmarks appear in the paper's own training mixture.

  1. fitted input called prediction [Appendix A.2, Figure A; Section 5.1]
    "To mitigate catastrophic forgetting, we adopt a data mixture strategy for training. This approach primarily involved using our proprietary CXRTrek dataset, supplemented with smaller portions of data from other relevant datasets [24, 32, 35–37, 39, 45, 47, 48]. ... Figure A: Distribution of training samples, illustrating the proportional contribution of each dataset to the training mixture. ... CXRTrek 51.08 MIMIC-CXR 21.81 VINDR 13.96 RSNA 3.87 MS-CXR 3.97 VQA-RAD 1.78 VQA-MED 0.74 SLake 0.10 IU X-ray 0.99 rad-restruct 1.67."

    The five benchmarks used to claim 'superior generalization across multiple tasks on five diverse external datasets' (Rad-ReStruct, VQA-RAD, IU-Xray, MIMIC-CXR, VinDr-CXR) all appear in the Appendix A.2 training mixture shown in Figure A. The paper does not state that these mixture components were restricted to non-evaluation splits, and the assets are not released. Therefore the 'external' results are at least partly a consequence of training on those datasets, rather than independent evidence of generalization. The claimed prediction is an input to training.

  2. self definitional [Section 3.3 vs. Section 5.1]
    "The CXRTrek training set is constructed from the MIMIC-CXR [32] training split and the CheXpert-plus [33] public dataset. ... we evaluate its performance on five external clinical benchmarks via fine-tuning ... MIMIC-CXR [32] for radiology report generation."

    MIMIC-CXR is not external to the model's training by the paper's own construction: CXRTrek's training set is built from the MIMIC-CXR training split, and MIMIC-CXR is also listed in the Appendix A.2 training mixture. Calling MIMIC-CXR one of the 'five external clinical benchmarks' and then reporting strong performance on it is a self-referential evaluation, not a generalization test.

full rationale

The central circularity is in the generalization claim. The five datasets used as 'external' benchmarks (VQA-RAD, Rad-ReStruct, VinDr-CXR, IU-Xray, and MIMIC-CXR) are all present in the Appendix A.2 training mixture (Figure A), and the paper neither restricts these mixture components to non-evaluation splits nor discloses the splits. Consequently, the headline result 'superior generalization across multiple tasks on five diverse external datasets' is at least partly a restatement of the training mixture rather than an independent transfer result. The MIMIC-CXR case is especially direct: CXRTrek is constructed from MIMIC-CXR's training split, so labeling MIMIC-CXR an external benchmark is self-contradictory. The CXRTrek test-set results themselves are not circular, because the 200-sample test set is drawn from MIMIC-CXR's test split while training uses the training split, which is a normal held-out evaluation. The concern about automatically generated labels is an annotation-quality issue, not circularity. No load-bearing self-citation was found; the paper's citations to prior work provide supporting context but do not by themselves force the main results. Overall, one or more of the paper's central 'predictions' reduce, at least in part, to the data on which the model was trained, yielding partial circularity.

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

The central claim rests on three key premises: the clinical validity of the 8-stage flow, the accuracy of the LLM-generated labels, and the adequacy of the evaluation metrics. The free parameters are hand-chosen thresholds in the annotation pipeline. No new physical entities are introduced.

free parameters (3)
  • entity token overlap threshold = 50%
    Hand-chosen threshold to retain LLM-extracted entities only if at least 50% of tokens match the medical vocabulary. Directly affects Stage-2 and Stage-3 labels.
  • synonym similarity threshold = 0.90
    Hand-chosen string similarity cutoff to collapse synonymous terms. Affects answer generation and label consistency across all stages.
  • medical term frequency cutoff = 10 occurrences
    Terms occurring fewer than ten times in the corpus are excluded from the medical vocabulary, influencing entity extraction and filtering.
assumptions (3)
  • domain assumption The 8-stage clinical reasoning flow is a core and necessary decomposition of radiologist interpretation.
    Based on standard radiology guidelines [16-18], but this is a modeling choice that the paper itself makes; there is no empirical proof that exactly these 8 stages are necessary or sufficient.
  • domain assumption LLM (Llama 3.2 8B) based extraction produces clinically accurate gold labels for entities, attributes, relationships, and advice.
    Validation is only BERTScore 92.53 against human annotations on 100 samples (Table D), not expert clinical review. The assumption is load-bearing for the entire dataset quality.
  • domain assumption BERTScore, Macro-F1, and mIoU adequately capture the quality of diagnostic reasoning.
    The paper itself notes in Appendix D that these metrics 'may not fully capture the complete spectrum of diagnostic quality', so the evaluation relies on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpreting Chest X-rays Like a Radiologist: A Benchmark with Clinical Reasoning." pith.science (2026). https://pith.science/paper/KBKTMIP7

@misc{pith2026250523143,
  author       = {Pith},
  title        = {Pith review of: Interpreting Chest X-rays Like a Radiologist: A Benchmark with Clinical Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KBKTMIP7}},
  note         = {Machine review of arXiv:2505.23143}
}
read the original abstract

Artificial intelligence (AI)-based chest X-ray (CXR) interpretation assistants have demonstrated significant progress and are increasingly being applied in clinical settings. However, contemporary medical AI models often adhere to a simplistic input-to-output paradigm, directly processing an image and an instruction to generate a result, where the instructions may be integral to the model's architecture. This approach overlooks the modeling of the inherent diagnostic reasoning in chest X-ray interpretation. Such reasoning is typically sequential, where each interpretive stage considers the images, the current task, and the contextual information from previous stages. This oversight leads to several shortcomings, including misalignment with clinical scenarios, contextless reasoning, and untraceable errors. To fill this gap, we construct CXRTrek, a new multi-stage visual question answering (VQA) dataset for CXR interpretation. The dataset is designed to explicitly simulate the diagnostic reasoning process employed by radiologists in real-world clinical settings for the first time. CXRTrek covers 8 sequential diagnostic stages, comprising 428,966 samples and over 11 million question-answer (Q&A) pairs, with an average of 26.29 Q&A pairs per sample. Building on the CXRTrek dataset, we propose a new vision-language large model (VLLM), CXRTrekNet, specifically designed to incorporate the clinical reasoning flow into the VLLM framework. CXRTrekNet effectively models the dependencies between diagnostic stages and captures reasoning patterns within the radiological context. Trained on our dataset, the model consistently outperforms existing medical VLLMs on the CXRTrek benchmarks and demonstrates superior generalization across multiple tasks on five diverse external datasets. The dataset and model can be found in our repository (https://github.com/guanjinquan/CXRTrek).

Figures

Figures reproduced from arXiv: 2505.23143 by the authors.

Figure 1
Figure 1. The clinical reasoning flow in CXRTrek through a multi-stage Q&A flow. Each Q&A pair [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Frequency analysis of word-pairs and task distributions in the training set of CXRTrek. (a) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Overview of CXRTrekNet. It takes as input one or more chest X-ray images and a sequence [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Two radar charts comparing stage-wise performance. Stage scores average corresponding [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

131 extracted references · 61 canonical work pages

  1. [1]

    Development and validation of open-source deep neural networks for comprehensive chest x-ray reading: a retrospective, multicentre study

    Yashin Dicente Cid, Matthew Macpherson, Louise Gervais-Andre, Yuanyi Zhu, Giuseppe Franco, Ruggiero Santeramo, Chee Lim, Ian Selby, Keerthini Muthuswamy, Ashik Amlani, et al. Development and validation of open-source deep neural networks for comprehensive chest x-ray reading: a retrospective, multicentre study. The Lancet Digital Health, 6(1):e44–e57, 2024

    2024

  2. [2]

    Early detection of lung cancer: clinical perspectives of recent advances in biology and radiology

    Fred R Hirsch, Wilbur A Franklin, Adi F Gazdar, and Paul A Bunn Jr. Early detection of lung cancer: clinical perspectives of recent advances in biology and radiology. Clinical Cancer Research, 7(1):5–22, 2001

    2001

  3. [3]

    Deep learning for chest x-ray analysis: A survey

    Erdi Çallı, Ecem Sogancioglu, Bram van Ginneken, Kicky G van Leeuwen, and Keelin Murphy. Deep learning for chest x-ray analysis: A survey. Medical image analysis, 72:102125, 2021

    2021

  4. [4]

    A review of recent advances in deep learning models for chest disease detection using radiography

    Adnane Ait Nasser and Moulay A Akhloufi. A review of recent advances in deep learning models for chest disease detection using radiography. Diagnostics, 13(1):159, 2023

    2023

  5. [5]

    Using artificial intelligence to read chest radiographs for tuberculosis detection: A multi-site evaluation of the diagnostic accuracy of three deep learning systems

    Zhi Zhen Qin, Melissa S Sander, Bishwa Rai, Collins N Titahong, Santat Sudrungrot, Sylvain N Laah, Lal Mani Adhikari, E Jane Carter, Lekha Puri, Andrew J Codlin, et al. Using artificial intelligence to read chest radiographs for tuberculosis detection: A multi-site evaluation of the diagnostic accuracy of three deep learning systems. Scientific reports, 9...

    2019

  6. [6]

    Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning

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

    arXiv 2017

  7. [7]

    Chexbert: combining automatic labelers and expert annotations for accurate radiology report labeling using bert

    Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Pareek, Andrew Y Ng, and Matthew P Lungren. Chexbert: combining automatic labelers and expert annotations for accurate radiology report labeling using bert. arXiv preprint arXiv:2004.09167, 2020

    arXiv 2004

  8. [8]

    Expert- level detection of pathologies from unannotated chest x-ray images via self-supervised learning

    Ekin Tiu, Ellie Talius, Pujan Patel, Curtis P Langlotz, Andrew Y Ng, and Pranav Rajpurkar. Expert- level detection of pathologies from unannotated chest x-ray images via self-supervised learning. Nature biomedical engineering, 6(12):1399–1406, 2022

    2022

Show all 131 references
  1. [9]

    Medclip: Contrastive learning from unpaired medical images and text

    Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing, ...

    2022

  2. [10]

    Generating radiology reports via memory- driven transformer

    Zhihong Chen, Yan Song, Tsung-Hui Chang, and Xiang Wan. Generating radiology reports via memory- driven transformer. arXiv preprint arXiv:2010.16056, 2020

    2010 arXiv

  3. [11]

    Act like a radiologist: Radiology report generation across anatomical regions

    Qi Chen, Yutong Xie, Biao Wu, Xiaomin Chen, James Ang, Minh-Son To, Xiaojun Chang, and Qi Wu. Act like a radiologist: Radiology report generation across anatomical regions. In Proceedings of the Asian Conference on Computer Vision, pages 1–17, 2024

    2024

  4. [12]

    R2gengpt: Radiology report generation with frozen llms

    Zhanyu Wang, Lingqiao Liu, Lei Wang, and Luping Zhou. R2gengpt: Radiology report generation with frozen llms. Meta-Radiology, 1(3):100033, 2023

    2023

  5. [13]

    Xraygpt: Chest radiographs summa- rization using medical vision-language models

    Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mullappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. Xraygpt: Chest radiographs summa- rization using medical vision-language models. arXiv preprint arXiv:2306.07971, 2023

    2023 arXiv

  6. [14]

    Llava-med: Training a large language-and-vision assistant for biomedicine in one day

    Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Nau- mann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36:28541–...

    2023

  7. [15]

    Radvlm: A multitask conversational vision-language model for radiology

    Nicolas Deperrois, Hidetoshi Matsuo, Samuel Ruipérez-Campillo, Moritz Vandenhirtz, Sonia Laguna, Alain Ryser, Koji Fujimoto, Mizuho Nishio, Thomas M Sutter, Julia E V ogt, et al. Radvlm: A multitask conversational vision-language model for radiology. arXiv preprint arXiv:2502....

    2025

  8. [16]

    A systematic approach to chest radiographic analysis

    Jeffrey S Klein and Melissa L Rosado-de Christenson. A systematic approach to chest radiographic analysis. Diseases of the Chest, Breast, Heart and Vessels 2019-2022: Diagnostic and Interventional Imaging, pages 1–16, 2019

    2019

  9. [17]

    Screening chest x-ray interpretations and radiographic techniques

    SM Gelaw. Screening chest x-ray interpretations and radiographic techniques. IOM guidelines. Global Radiology Coordination and Teleradiology Centre, Makati city, Philippines, 2015

    2015

  10. [18]

    Radiographic image analysis-E-book

    Kathy McQuillen Martensen. Radiographic image analysis-E-book. Elsevier Health Sciences, 2013

    2013

  11. [19]

    Tienet: Text-image embedding network for common thorax disease classification and reporting in chest x-rays

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, and Ronald M Summers. Tienet: Text-image embedding network for common thorax disease classification and reporting in chest x-rays. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9049–9058, 2018

    2018

  12. [20]

    Improving chest X-ray report generation by leveraging warm starting

    Aaron Nicolson, Jason Dowling, and Bevan Koopman. Improving chest X-ray report generation by leveraging warm starting. Artificial Intelligence in Medicine, 144:102633, 2023. 11

    2023

  13. [21]

    Gloria: A multimodal global- local representation learning framework for label-efficient medical image recognition

    Shih-Cheng Huang, Liyue Shen, Matthew P Lungren, and Serena Yeung. Gloria: A multimodal global- local representation learning framework for label-efficient medical image recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3942–3951, 2021

    2021

  14. [22]

    Contrastive learning of medical visual representations from paired images and text

    Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D Manning, and Curtis P Langlotz. Contrastive learning of medical visual representations from paired images and text. In Machine learning for healthcare conference, pages 2–25. PMLR, 2022

    2022

  15. [23]

    Sedigheh Eslami, Christoph Meinel, and Gerard De Melo. Pubmedclip: How much does clip benefit visual question answering in the medical domain? In Findings of the Association for Computational Linguistics: EACL 2023, pages 1181–1193, 2023

    2023

  16. [24]

    Rad-restruct: A novel vqa benchmark and method for structured radiology reporting

    Chantal Pellegrini, Matthias Keicher, Ege Özsoy, and Nassir Navab. Rad-restruct: A novel vqa benchmark and method for structured radiology reporting. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 409–419. Springer, 2023

    2023

  17. [25]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  18. [26]

    Minigpt-med: Large language model as a general interface for radiology diagnosis

    Asma Alkhaldi, Raneem Alnajim, Layan Alabdullatef, Rawan Alyahya, Jun Chen, Deyao Zhu, Ahmed Alsinan, and Mohamed Elhoseiny. Minigpt-med: Large language model as a general interface for radiology diagnosis. arXiv preprint arXiv:2407.04106, 2024

    2024 arXiv

  19. [27]

    Chexagent: Towards a foundation model for chest x-ray interpretation

    Zhihong Chen, Maya Varma, Jean-Benoit Delbrouck, Magdalini Paschali, Louis Blankemeier, Dave Van Veen, Jeya Maria Jose Valanarasu, Alaa Youssef, Joseph Paul Cohen, Eduardo Pontes Reis, et al. Chexagent: Towards a foundation model for chest x-ray interpretation. arXiv preprint ...

    2024 arXiv

  20. [28]

    Biomedgpt: Open multimodal generative pre-trained transformer for biomedicine

    Yizhen Luo, Jiahuan Zhang, Siqi Fan, Kai Yang, Yushuai Wu, Mu Qiao, and Zaiqing Nie. Biomedgpt: Open multimodal generative pre-trained transformer for biomedicine. arXiv preprint arXiv:2308.09442, 2023

    2023 arXiv

  21. [29]

    Towards generalist foundation model for radiology by leveraging web-scale 2d&3d medical data

    Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards generalist foundation model for radiology by leveraging web-scale 2d&3d medical data. arXiv preprint arXiv:2308.02463, 2023

    2023 arXiv

  22. [30]

    Pmc- vqa: Visual instruction tuning for medical visual question answering

    Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc- vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415, 2023

    2023 arXiv

  23. [31]

    Huatuogpt-vision, towards injecting medical visual knowledge into multimodal llms at scale

    Junying Chen, Chi Gui, Ruyi Ouyang, Anningzhe Gao, Shunian Chen, Guiming Hardy Chen, Xidong Wang, Ruifei Zhang, Zhenyang Cai, Ke Ji, et al. Huatuogpt-vision, towards injecting medical visual knowledge into multimodal llms at scale. arXiv preprint arXiv:2406.19280, 2024

    2024 arXiv

  24. [32]

    Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports

    Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019

    2019

  25. [33]

    Chexpert plus: Hundreds of thousands of aligned radiology texts, images and patients

    Pierre Chambon, Jean-Benoit Delbrouck, Thomas Sounack, Shih-Cheng Huang, Zhihong Chen, Maya Varma, Steven QH Truong, Curtis P Langlotz, et al. Chexpert plus: Hundreds of thousands of aligned radiology texts, images and patients. arXiv e-prints, pages arXiv–2405, 2024

    2024

  26. [34]

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

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

    2019

  27. [35]

    Vindr-cxr: An open dataset of chest x-rays with radiologist’s annotations

    Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist’s annotations. Scientific Data, 9(1):429, 2022

    2022

  28. [36]

    Vqa-med: Overview of the medical visual question answering task at imageclef 2019

    Asma Ben Abacha, Sadid A Hasan, Vivek V Datla, Dina Demner-Fushman, and Henning Müller. Vqa-med: Overview of the medical visual question answering task at imageclef 2019. In Proceedings of CLEF (Conference and Labs of the Evaluation Forum) 2019 Working Notes. 9-12 September 2019, 2019

    2019

  29. [37]

    A dataset of clinically generated visual questions and answers about radiology images

    Jason J Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images. Scientific data, 5(1):1–10, 2018

    2018

  30. [38]

    Mimic-ext-mimic-cxr-vqa: A complex, diverse, and large-scale visual question answering dataset for chest x-ray images, 2024

    Seongsu Bae, Daeun Kyung, Jaehee Ryu, Eunbyeol Cho, Gyubok Lee, Sunjun Kweon, Jungwoo Oh, Lei JI, Eric Chang, Tackeun Kim, et al. Mimic-ext-mimic-cxr-vqa: A complex, diverse, and large-scale visual question answering dataset for chest x-ray images, 2024

    2024

  31. [39]

    Slake: A semantically-labeled knowledge-enhanced dataset for medical visual question answering

    Bo Liu, Li-Ming Zhan, Li Xu, Lin Ma, Yan Yang, and Xiao-Ming Wu. Slake: A semantically-labeled knowledge-enhanced dataset for medical visual question answering. In 2021 IEEE 18th international symposium on biomedical imaging (ISBI), pages 1650–1654. IEEE, 2021. 12

    2021

  32. [40]

    Expert knowledge-aware image difference graph representation learning for difference-aware medical visual question answering

    Xinyue Hu, Lin Gu, Qiyuan An, Mengliang Zhang, Liangchen Liu, Kazuma Kobayashi, Tatsuya Harada, Ronald M Summers, and Yingying Zhu. Expert knowledge-aware image difference graph representation learning for difference-aware medical visual question answering. InProceedings of th...

    2023

  33. [41]

    Chest imagenome dataset

    Joy Wu, Nkechinyere Agu, Ismini Lourentzou, Arjun Sharma, Joseph Paguio, Jasper Seth Yao, Ed- ward Christopher Dee, William Mitchell, Satyananda Kashyap, Andrea Giovannini, et al. Chest imagenome dataset. Physio Net, 2021

    2021

  34. [42]

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

    Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria De La Iglesia-Vaya. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797, 2020

    2020

  35. [43]

    The national lung screening trial: overview and study design

    National Lung Screening Trial Research Team. The national lung screening trial: overview and study design. Radiology, 258(1):243–253, 2011

    2011

  36. [44]

    Ms-cxr-t: Learning to exploit temporal structure for biomedical vision-language processing, 2023

    Shruthi Bannur, Stephanie Hyland, Qianchu Liu, Fernando Pérez-García, Max Ilse, Daniel Coelho de Cas- tro, Benedikt Boecking, Harshita Sharma, Kenza Bouzid, Anton Schwaighofer, et al. Ms-cxr-t: Learning to exploit temporal structure for biomedical vision-language processing, 2023

    2023

  37. [45]

    Ms-cxr: Making the most of text semantics to improve biomedical vision-language processing

    Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel Coelho de Castro, Anton Schwaighofer, Stephanie Hyland, Maria Teodora Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez Valle, et al. Ms-cxr: Making the most of text semantics to improve biomedical vision-language...

    doi

  38. [46]

    Medical-cxr-vqa dataset: A large-scale llm-enhanced medical dataset for visual question answering on chest x-ray images

    Xinyue Hu et al. Medical-cxr-vqa dataset: A large-scale llm-enhanced medical dataset for visual question answering on chest x-ray images

  39. [47]

    Preparing a collection of radiology examinations for distribution and retrieval

    Dina Demner-Fushman, Marc D Kohli, Marc B Rosenman, Sonya E Shooshan, Laritza Rodriguez, Sameer Antani, George R Thoma, and Clement J McDonald. Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Associat...

    2016

  40. [48]

    Rsna pneumonia detection challenge

    MD Anouk Stein, Carol Wu, Chris Carr, George Shih, Jamie Dulkowski, kalpathy, Leon Chen, Luciano Prevedello, MD Marc Kohli, Mark McDonald, Peter, Phil Culliton, Safwan Halabi MD, and Tian Xia. Rsna pneumonia detection challenge. https://kaggle.com/competitions/ rsna-pneumonia-...

    2018

  41. [49]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  42. [50]

    The unified medical language system (umls): integrating biomedical terminology

    Olivier Bodenreider. The unified medical language system (umls): integrating biomedical terminology. Nucleic acids research, 32(suppl_1):D267–D270, 2004

    2004

  43. [51]

    Medcot: Medical chain of thought via hierarchical expert

    Jiaxiang Liu, Yuan Wang, Jiawei Du, Joey Tianyi Zhou, and Zuozhu Liu. Medcot: Medical chain of thought via hierarchical expert. arXiv preprint arXiv:2412.13736, 2024

    2024 arXiv

  44. [52]

    Bertscore: Evaluating text generation with bert

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019

    1904 arXiv

  45. [53]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022

    2022

  46. [54]

    Roformer: Enhanced transformer with rotary position embedding

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024

    2024

  47. [55]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024

    2024 arXiv

  48. [56]

    Radgraph: Extracting clinical entities and relations from radiology reports

    Saahil Jain, Ashwin Agrawal, Adriel Saporta, Steven QH Truong, Du Nguyen Duong, Tan Bui, Pierre Chambon, Yuhao Zhang, Matthew P Lungren, Andrew Y Ng, et al. Radgraph: Extracting clinical entities and relations from radiology reports. arXiv preprint arXiv:2106.14463, 2021

    2021 arXiv

  49. [57]

    textblob documentation

    Steven Loria et al. textblob documentation. Release 0.15, 2(8):269, 2018

    2018

  50. [58]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81, 2004

    2004

  51. [59]

    Ng, Curtis P

    Feiyang Yu, Mark Endo, Rayan Krishnan, Ian Pan, Andy Tsai, Eduardo Pontes Reis, Eduardo Kaiser Ururahy Nunes Fonseca, Henrique Min Ho Lee, Zahra Shakeri Hossein Abad, Andrew Y . Ng, Curtis P. Langlotz, Vasantha Kumar Venugopal, and Pranav Rajpurkar. Evaluating progress in auto...

    2022

  52. [60]

    Interpreting Chest X-rays Like a Radiologist: A Benchmark with Clinical Reasoning

    Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 3...

    2020

  53. [61]

    Right-sided mild pleural effusion is again seen largely unchanged

  54. [62]

    Left-sided ground glass opacity which has slightly improved consistent with improving pulmonary edema

  55. [63]

    Endotracheal tube in appropriate position, about 6 cm from the carina

  56. [64]

    as- sociated with

    Incidental note of right lateral pleural calcification which is better seen on ct imaging. S1’s Info SN’s Info Sent-1 Sent-N LLM Fine-grained Information Extraction Framework Medical Terms Input [Material] Chest X-ray Report Medical Terms [Appendix C.2] [Product] Entities and ...

  57. [65]

    Match: Entities must match the abnormalities and foreign bodies described in the X-ray report

  58. [66]

    Exclude Descriptors and Anatomy: Do not include size, degree, location, numerical measurements, or anatomy unless they are part of a specific abnormality term

  59. [67]

    {0}”, do you think that “{1}

    No Inference: Only include entities explicitly mentioned in the report. First, output the analysis in the following format: Analysis: [Output analysis of all possible abnormalities and foreign bodies descriptions based on the definitions and rules; if not found, state that cle...

  60. [68]

    Location is not part of the entity name (“{0}”)

  61. [69]

    Do not infer an answer that is not explicitly mentioned in the report

  62. [70]

    Location: None Prompt for severity attributes extraction

    Location must be detailed and complete with anatomy, location, and other direction and distance information if possible. Location: None Prompt for severity attributes extraction. We employ the following prompt to extract the severity attributes of an entity: Severity Attribute...

  63. [71]

    Do not infer severity information from the report if it is not explicitly mentioned. 24

  64. [72]

    The severity must be from the provided severity list

  65. [73]

    The severity must directly link to the entity in the report

  66. [74]

    and”, “or

    Connection words such as “and”, “or” can transfer the severity information. First, you need to analyze the severity information for each entity: Analysis: [Output the analysis of the severity information for each entity in one sentence, strictly using the report’s language. If...

  67. [75]

    Do not infer trend information from the report if it is not explicitly mentioned

  68. [76]

    The trend must be from the provided trend list

  69. [77]

    The trend must directly link to the entity in the report

  70. [78]

    and”, “or

    Connection words such as “and”, “or” can transfer the trend information. First, you need to analyze the trend information for each entity: Analysis: [Output the analysis of the trend information for each entity in one sentence, strictly using the report’s language. If no trend...

  71. [79]

    Do not infer trait information from the report if it is not explicitly mentioned

  72. [80]

    The trait must be from the provided trait list

  73. [81]

    The trait must directly link to the entity in the report

  74. [82]

    and”, “or

    Connection words such as “and”, “or” can transfer the trait information. First, you need to analyze the trait information for each entity: Analysis: [Output the analysis of the trait information for each entity in one sentence, strictly using the report’s language. If no trait...

  75. [83]

    Only focus on {0}’s description, exclude the information of other abnormalities and foreign bodies

  76. [84]

    Do not infer the information from the report if it is not explicitly mentioned

  77. [85]

    We found a severe pleural effusion in the left hemithorax

    The description must be detailed and complete, including the shape, size, density, and other characteristics if possible. Description: [Answer one sentence about the detailed and complete description of {0} using the report’s language] Stop immediately after outputting the res...

  78. [86]

    Correct the error or conflict information

  79. [87]

    Polish the unspecific location. 26

  80. [88]

    If there are no valid and correct location specified, state that clearly

  81. [89]

    Focus on the location of {0} only

  82. [90]

    {0}”. The extracted traits information is: “{1}

    If the location can be merged, please summarize them into one word or a sentence, such as left and right should be merged. Output in the following format with <answer> and <think> tags: <think> Thinking process to analyze and refine the content in one sentence; If no valid and...

  83. [91]

    Correct the error or conflicting traits to find the suitable traits

  84. [92]

    Polish the unspecific traits with more specific located information

  85. [93]

    If there are no valid traits specified, state that clearly

  86. [94]

    Focus on the traits of {0} only

  87. [95]

    {0}”. The extracted severity information is: “{1}

    Merge the traits with the same location information into a combined word. Output in the following format with <answer> and <think> tags: <think> Thinking process to analyze and refine the content in one concise and short sentence; If no valid and correct traits specified, clea...

  88. [96]

    Correct the error or conflicting severity to find the suitable severity. 27

  89. [97]

    Polish the unspecific severity with more specific located information

  90. [98]

    Focus on the severity of {0} only

  91. [99]

    {0}”. The extracted trend information is: “{1}

    Merge the severity with the same location information into a combined word. Output in the following format with <answer> and <think> tags: <think> Thinking process to analyze and refine the content in one concise and short sentence; If no valid and correct severity specified, ...

  92. [100]

    Correct the error or conflicting trend to find the suitable trend

  93. [101]

    Polish the unspecific trend with more specific located information

  94. [102]

    Focus on the trend of {0} only

  95. [103]

    {0}”. The extracted description is: “{1}

    Merge the trend with the same location information into a combined word. Output in the following format with <answer> and <think> tags: <think> Thinking process to analyze and refine the content in one concise and short sentence; If no valid and correct trend specified, clearl...

  96. [104]

    Correct the error or conflicting description to find the suitable description

  97. [105]

    Polish or ignore the unspecific description

  98. [106]

    What are the pathological relationships among the findings?

    Focus on the description of {0} only. 28 Output in the following format with <answer> and <think> tags: <think> Thinking process to analyze and refine the content in one concise and short sentence; If no valid and correct description specified, clearly state this. </think> <an...

  99. [107]

    Exclude entities without relationships or only with relationships to itself

  100. [108]

    Do not change the facts or infer relationships from context

  101. [109]

    What are the X-ray issues regarding assessment difficulty, patient posture, image quality, and technical issues?

    Only the direct relationships between the findings are considered. Output in the following format: <think> Analysis adhering to the rules and report in one sentence; If no relationships mentioned, explicitly state this. </think> <answer> Strictly using the report’s language an...

  102. [110]

    Exclude any information about “{1}”

  103. [111]

    Only include information related to assessment difficulty, patient posture, image quality, and technical issues

  104. [112]

    What are the recommendations for further CT/MR/X-ray imaging, clinical correla- tion, requirements, follow-up, uncertain diagnosis examination, or treatment in this study?

    Do not infer issues from the report or change the facts if they are not explicitly men- tioned. Output in the following format: <think> Analysis adhering to the rules and report in one sentence; If no quality issues mentioned, explicitly state this. </think> <answer> A simple ...

  105. [113]

    Focus on the recommendation information for further CT/MR/X-ray imaging, clinical correlation, requirements, follow-up, uncertain diagnosis examination, or treatment

  106. [114]

    If a CT/MR is mentioned in the report, it implies a CT/MR is recommended

  107. [115]

    Output in the following format: <think> Analysis adhering to the rules and report in one sentence; If no recommendations mentioned, explicitly state this

    Do not infer if it is not explicitly mentioned. Output in the following format: <think> Analysis adhering to the rules and report in one sentence; If no recommendations mentioned, explicitly state this. </think> <answer> A simple sentence that fully details the recommendations...

  108. [116]

    Guidelines: • The answer NA means that the abstract and introduction do not include the claims made in the paper

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: Yes, the abstract accurately reflects the paper’s contributions: introducing the CXRTrek multi-stage VQA dataset to em...

  109. [117]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We have discussed the limitations of our work, and then further discussed the future direction of our work in the Appendix Section D. Guidelines: • T...

  110. [118]

    However, relevant existing theories or concepts that underpin the work are appropriately referenced and cited

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [NA] Justification: This paper does not propose new theories or theoretical results; hence, this item is not appl...

  111. [119]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  112. [120]

    Guidelines: • The answer NA means that paper does not include experiments requiring code

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: Not open yet, but the paper ...

  113. [121]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental setting/details Question: Does the paper specify all the training and test details ( e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: The paper provides detailed ...

  114. [122]

    However, we observe stable and consistent performance trends across all evaluation stages and datasets, which supports the reliability of our results

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer:[No] Justification: Error bars are not reported. However, we observe sta...

  115. [123]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The paper provides deta...

  116. [124]

    Guidelines: • The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: This research fully conforms to the NeurIPS Code of Ethics. Guidelines: • The a...

  117. [125]

    Guidelines: • The answer NA means that there is no societal impact of the work performed

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? 35 Answer: [Yes] Justification: The broader impacts have been discussed in the Appendix Section E. Guidelines: • The answer NA means t...

  118. [126]

    This responsible approach is reinforced by strict adherence to source data use agreements and a planned tiered access mechanism for dissemination

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [Yes] Justification: A core safegua...

  119. [127]

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: Creators of all ...

  120. [128]

    Docu- mentation will be provided with their release upon acceptance

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [No] Justification: Our dataset (CXRTrek) and model (CXRTrekNet) are not yet released. Docu- mentation will be provided with their releas...

  121. [129]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  122. [130]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  123. [131]

    Answer: [Yes] Justification: The paper thoroughly describes the use of the relevant LLM in Sections 3, 4, and 5

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.