Pith. sign in

REVIEW 3 major objections 5 minor 41 references

RADAR: Enhancing Radiology Report Generation with Supplementary Knowledge Injection

T0 review · 3 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 This paper claims that radiology report generation improves when a draft LLM report is filtered by agreement with an expert image classifier and then augmented with retrieved findings for what remains, with gains over prior multimodal…

desk verdict A well-engineered combination of expert filtering and retrieval for radiology reports, worth engaging despite a possible retrieval leak and an uncharacterized expert model. read the letter →

arxiv 2505.14318 v2 pith:L2Z7SIC2 submitted 2025-05-20 cs.CV cs.CL

Wenjun Hou , Yi Cheng , Kaishuai Xu , Heng Li , Yan Hu , Wenjie Li , Jiang Liu This is my paper
classification cs.CVcs.CL
keywords radiologyreportgenerationlargelanguagemodelsretrieval-augmentedknowledgeinjectionchestX-rayclinicalaccuracyobservationclassificationhallucinationreduction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a radiology report generator becomes more clinically accurate when the model's own draft is treated as a proposal rather than a final answer. The proposed framework, RADAR, first generates a report, then uses an independently trained expert classifier to mark which proposed observations are trustworthy: only findings present in both the draft and the classifier's output ($O_{\checkmark} = O_I \cap O_R$) survive. It then retrieves top-$K$ similar reference reports, discards anything those surviving findings already cover, and feeds the remaining supplementary findings back into the LLM to compose the final report. Across MIMIC-CXR, CheXpert-Plus, and IU-Xray, the authors report that this two-stage scheme beats prior multimodal LLMs on language quality and clinical metrics, with relative gains above 6% on four CheXpert observation metrics.

What carries the argument

The load-bearing mechanism is the agreement intersection between two independent observers. An expert model, built from a Swin Transformer V2 image encoder and a clinical BERT text encoder and trained on CheXpert observation labels, predicts a set of observations $O_I$; the LLM's draft report is converted into semi-structured findings by CheXbert annotation, giving $O_R$; and the framework keeps only $O_{\checkmark} = O_I \cap O_R$ as trustworthy internal knowledge. A retrieval step then scores candidate reference reports by KL-divergence between their normalized observation probability vectors and the current image's, takes the top-$K$ reports, and removes every sentence that does not correspond to an observation outside $O_{\checkmark}$, producing the supplementary findings. Both sets are concatenated into the clinical context for a second generation pass, and during training the target report is prefixed by its own observation list so the model first identifies observations and then writes the findings.

What would settle it

Measure the expert classifier's per-observation recall on a radiologist-labeled subset, then compare it with RADAR's final-report recall for the same observations. If final-report recall tracks classifier recall on rare findings such as Fracture, the agreement filter is setting the ceiling; a sharper experiment would be to disable one observation (for example, force the classifier's Edema output to zero on a random half of the test set) and check whether RADAR's Edema recall drops correspondingly, which would confirm that the filter, not the retrieved supplement, controls what reaches the report.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that hallucination in LLM radiology reporting is best treated by a split of labor between what the model already knows and what it still needs. The observations the LLM proposes are kept only where they agree with an expert vision–language classifier's output; the observations the classifier flags but the draft missed are recovered from retrieved reference reports whose content does not duplicate the surviving findings. This converts generation from a single pass into two passes, with the second pass consuming both preliminary and supplementary findings as context, and with the report's observation list included in the training target so the model learns to state observations explicitly before writing findings. The ablations support the claim that each piece carries weight: removing the agreement filter, the supplementary retrieval, or the observation-identification target each lowers clinical accuracy, while replacing the agreement filter with plain retrieval-augmented generation leaves a relative gap of more than 6% on four CheXpert metrics.

Load-bearing premise

The whole framework rests on the expert classifier being accurate enough that a finding it misses is truly absent: whenever the classifier overlooks a real abnormality, the agreement intersection silently removes that correct finding from the final report, and nothing later in the pipeline restores it.

Editorial extensions

If this is right

  • If RADAR's results hold, retrieval-augmented medical report generation is improved by filtering the model's own output against an independent classifier before injecting retrieved knowledge, rather than by injecting retrieved knowledge unconditionally.
  • A clinician can read the expert model's label set as a contract: the final report can only discuss observations that survive the agreement filter, so any finding the expert classifier cannot see is excluded by construction.
  • The two-pass design decouples language quality from factual grounding, letting the same LLM verify its own draft and then compose the final text from the verified findings.
  • The observation-identification training target alone, without any retrieval, raises CheXpert macro-F1 noticeably, which suggests that asking the model to list observations before writing findings is itself a useful inductive bias.

Reading between the lines

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

  • A direct way to separate the two contributions is to run RADAR with the expert model's output replaced by ground-truth radiologist labels: the residual gap would measure how much the expert classifier's own errors, rather than the framework's design, cost in clinical accuracy.
  • Because final-report recall is capped by the expert model's recall, rare findings such as Fracture (per-observation F1 of 0.309 in the paper's appendix) are where the agreement filter is most likely to delete true positives, so a sensitivity analysis varying the classifier's decision threshold on rare classes would show whether the filter should be asymmetric per observation.
  • The retrieval stage uses KL-divergence over observation probability vectors, but the paper does not compare this to retrieval over image embeddings or over the draft report's text, so it remains untested whether observation-space retrieval is the key to the supplementary findings' usefulness or merely adequate.
  • A control that replaces retrieved reports with randomly selected reports would reveal whether the second stage helps by adding specific missing knowledge or simply by giving the model a longer, more structured context to condition on.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes RADAR, a framework for radiology report generation that combines an MLLM's internal knowledge with externally retrieved knowledge. The method first generates a draft report, uses a trained expert model (Swin Transformer V2 + BioClinicalBERT) to classify observations, retains the intersection of the draft's observations and the expert's observations (O✓ = OI ∩ OR) as high-confidence internal knowledge, retrieves top-K similar reports from the MIMIC-CXR training set via KL divergence on expert probabilities, extracts only observations not in O✓ as supplementary findings, and finally generates the report from the aggregated context. Experiments on MIMIC-CXR, CheXpert-Plus, and IU X-ray report improvements over prior LLM baselines in lexical and clinical metrics, with relative gains above 6% on several CheXpert metrics over the backbone.

Significance. The idea of systematically separating an LLM's already-learned knowledge from genuinely supplementary retrieved knowledge is well motivated and practically relevant. The paper includes a public code release, multi-dataset evaluation, and ablations of the main components (PF, SF, OI). If the reported results are robust, the framework could be a useful step toward reducing redundant retrieval and improving clinical accuracy in report generation. However, the central validation gaps—particularly the lack of any evaluation of the expert model that drives both the agreement filter and the retrieval—mean that the strength of the claims is currently disproportionate to the evidence.

major comments (3)
  1. [§3.1 (Eq. 4), §5.2 (Fig. 4)] The expert model's per-observation accuracy is never reported, yet the agreement filter O✓ = OI ∩ OR is the load-bearing mechanism for the claimed clinical improvements. Because the filter can only delete observations, any true finding that the expert model misses is removed from the preliminary findings and can only be recovered if the supplementary retrieval happens to contain it. Since the expert model is trained on CheXbert-derived labels from the same MIMIC-CXR reports, its error profile is entirely uncharacterized. A concrete test is to report the expert model's precision, recall, and AUROC on the MIMIC-CXR test set (or a human-annotated subset), and to analyze how often its false negatives cause correct findings in the LLM draft to be deleted. Without this, the SOTA claim in Table 1 rests on an unvalidated component.
  2. [§5.1, Tables 1–3] The SOTA claims rely almost entirely on baseline results cited from the literature, rather than on re-running the baselines under a consistent evaluation protocol. Differences in training data, prompt templates, metric implementations, and CheXpert uncertainty labeling (e.g., 'Uncertain as Positive' vs 'Uncertain as Negative') make the reported margins, such as the 'smallest gain over the second-best model is 2.1%', difficult to interpret. The evaluation on CheXpert-Plus uses only 62 validation samples (Section 4.1); without confidence intervals or significance tests, the ranking in Table 3 is not statistically meaningful. The authors should provide error bars and ideally re-implement or at least re-evaluate the most directly comparable baselines (e.g., Libra, MAIRA-2) under the same conditions.
  3. [§3.2 (Eq. 5), Table 6] The retrieval similarity is computed using the same expert model that is used for filtering, so errors in the expert model propagate to both stages. The paper does not specify the threshold used to binarize p(Oi) into OI, nor does it report the expert model's calibration or recall. This makes the method difficult to reproduce and the impact of expert errors on the final report impossible to assess. A sensitivity analysis (e.g., varying the binarization threshold, or corrupting the expert probabilities) would help establish whether the reported gains are robust to the expert model's imperfections.
minor comments (5)
  1. [§2.2] The conversion of reports into semi-structured knowledge via CheXbert labels is described only briefly; it is unclear how sentences with multiple observations or mixed positive/negative statements are split and labeled.
  2. [§4.1/Table 3] The CheXpert-Plus evaluation set contains only 62 samples, but the paper does not mention this limitation in the Limitations section; the small sample size should be acknowledged and reflected in the strength of the claims.
  3. [Table 1] The layout of Table 1 is confusing in places (e.g., the MAIRA-1 row contains multiple consecutive values for different metrics without clear column alignment); the table should be reformatted for readability.
  4. [§3.2] The paper states that 'Top-2 reports are selected as knowledge' (Section 4.4), but no ablation over K is provided; since the paper lists 'top-K retrieved reports' as a free parameter, a sensitivity analysis would strengthen the work.
  5. [Throughout] The name of the framework is inconsistently capitalized ('RADAR' in the title, 'RADAR' in the abstract and elsewhere); the authors should standardize the spelling.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: RADAR's pipeline is a standard supervised retrieval-augmented generation setup evaluated on held-out test sets, with no fitted parameter or self-citation chain forcing the claimed results.

full rationale

RADAR does not derive its predictions from its own inputs by construction. The expert model is trained with binary cross-entropy on MIMIC-CXR training labels (Section 3.1), while the clinical metrics are computed on held-out test references using the same CheXpert/CheXbert taxonomy; this is standard supervised learning, not fitting a parameter to the evaluation target. The agreement filter O✓ = OI ∩ OR is a design choice for selecting which draft observations to trust, and it does not define the evaluation metric. The final report is generated by an LLM from PF and SF context and evaluated as text, so the expert model's labels are not directly the output. The retrieval stage (Section 3.2) uses Oδ = O − O✓ but still requires the LLM to verbalize retrieved sentences correctly, and the paper's own Case B (Figure 4) shows an expert false negative being rescued by retrieval, demonstrating that the pipeline can succeed even when the expert misses a finding. The self-citations (Hou et al. 2023a,b; Hou et al. 2024) appear only in related-work, baseline, and contrastive statements; none is load-bearing for RADAR's central mechanism. The unmeasured per-observation accuracy of the expert model is a genuine correctness and robustness risk, but it is not circular reasoning, so the circularity score is 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method introduces no new theoretical entities or physical postulates. It relies on standard domain assumptions and one hand-chosen constant.

free parameters (1)
  • top-K retrieved reports = 2
    The number of supplementary reports retrieved in Stage II is set to 2 without a sensitivity study; the final result depends on this choice.
assumptions (3)
  • domain assumption CheXpert 14-category taxonomy is a sufficient set of observations for filtering and supplementing report knowledge
    All filtering and retrieval operate on these 14 categories; any clinically important finding outside this set cannot be handled.
  • domain assumption CheXbert annotations of radiology reports are reliable labels for training the expert model and evaluating clinical accuracy
    The expert model is trained on CheXbert-derived labels (Section 2.2) and the clinical metrics are also computed against CheXpert labels, so label noise enters both training and evaluation.
  • domain assumption The expert model's predicted observation probabilities are meaningful for KL-divergence retrieval across samples
    Section 3.2 uses the normalized expert probabilities to rank similar patients; no calibration or retrieval-specific validation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RADAR: Enhancing Radiology Report Generation with Supplementary Knowledge Injection." pith.science (2026). https://pith.science/paper/L2Z7SIC2

@misc{pith2026250514318,
  author       = {Pith},
  title        = {Pith review of: RADAR: Enhancing Radiology Report Generation with Supplementary Knowledge Injection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L2Z7SIC2}},
  note         = {Machine review of arXiv:2505.14318}
}
read the original abstract

Large language models (LLMs) have demonstrated remarkable capabilities in various domains, including radiology report generation. Previous approaches have attempted to utilize multimodal LLMs for this task, enhancing their performance through the integration of domain-specific knowledge retrieval. However, these approaches often overlook the knowledge already embedded within the LLMs, leading to redundant information integration. To address this limitation, we propose Radar, a framework for enhancing radiology report generation with supplementary knowledge injection. Radar improves report generation by systematically leveraging both the internal knowledge of an LLM and externally retrieved information. Specifically, it first extracts the model's acquired knowledge that aligns with expert image-based classification outputs. It then retrieves relevant supplementary knowledge to further enrich this information. Finally, by aggregating both sources, Radar generates more accurate and informative radiology reports. Extensive experiments on MIMIC-CXR, CheXpert-Plus, and IU X-ray demonstrate that our model outperforms state-of-the-art LLMs in both language quality and clinical accuracy.

Figures

Figures reproduced from arXiv: 2505.14318 by the authors.

Figure 1
Figure 1. A motivating example. The report directly [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the RADAR. In Preliminary Findings, only sentences that reach agreement are retained, whereas in Supplementary Findings, only sentences that supplement the Preliminary Findings are preserved. LLM, we first feed the input (X, Xp, and C) into RADAR to generate a report Yˆ : Yˆ = argmax Yˆ ∈Y Y T t=1 MLLM(X, Xp, C, yˆ<t), where Y represents the set of possible reports. Note that exact maximization is intrac… view at source ↗
Figure 3
Figure 3. Comparisons among BACKBONE+RAG, BACKBONE+FP+SF, and RADAR on six clinical met￾rics. CXR dataset, across both lexical and clinical met￾rics. Furthermore, RADAR surpasses the baseline that is trained on CHEXPERT PLUS alone as well as the one trained on a combination of both datasets. These results demonstrate the strong generalization ability of RADAR across different datasets. Ad￾ditionally, RADAR significantly outpe… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Two cases generated by RADAR, where false positive observation appears in the PF of case A and false negative observation shows in the PF of case B. Analysis of RADAR versus RAG. To evaluate the effectiveness of knowledge integration in RADAR, we conduct experiments co…
Figure 5
Figure 5. Figure 5: Error case generated by RADAR, where spans and spans indicate incorrect and correct observations. age external knowledge to recover missing findings, thereby improving factual completeness. Error Analysis. We conduct an error analysis to gain deeper insights, as shown …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 20 canonical work pages

  1. [3]

    Preprint, arXiv:2406.04449

    Maira-2: Grounded radi- ology report generation. Preprint, arXiv:2406.04449. Shruthi Bannur, Stephanie Hyland, Qianchu Liu, Fer- nando Pérez-García, Maximilian Ilse, Daniel C. Cas- tro, Benedikt Boecking, Harshita Sharma, Kenza Bouzid, Anja Thieme, Anton Schwaighofer, Maria Wetscherek, Matthew P. Lungren, Aditya Nori, Javier Alvarez-Valle, and Ozan Oktay

    arXiv

  2. [5]

    Preprint, arXiv:2405.19538

    Chexpert plus: Augment- ing a large chest x-ray dataset with text radiology reports, patient demographics and additional image formats. Preprint, arXiv:2405.19538. Juan Manuel Zambrano Chaves, Shih-Cheng Huang, Yanbo Xu, Hanwen Xu, Naoto Usuyama, Sheng Zhang, Fei Wang, Yujia Xie, Mahmoud Khademi, Ziyi Yang, Hany Awadalla, Julia Gong, Houdong Hu, Jianwei Y...

    arXiv

  3. [6]

    Preprint, arXiv:2403.08002

    Towards a clinically accessi- ble radiology foundation model: open-access and lightweight, with automated evaluation. Preprint, arXiv:2403.08002. Zhihong Chen, Yaling Shen, Yan Song, and Xi- ang Wan

    arXiv

  4. [9]

    Preprint, arXiv:2401.12208

    Chexagent: Towards a foundation model for chest x-ray interpre- tation. Preprint, arXiv:2401.12208. Francesco Dalla Serra, Chaoyang Wang, Fani Deli- gianni, Jeff Dalton, and Alison O’Neil

    arXiv

  5. [10]

    In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4891–4904, Singapore

    Con- trollable chest X-ray report generation from longitu- dinal representations. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4891–4904, Singapore. Association for Computa- tional Linguistics. Dina Demner-Fushman, Marc D Kohli, Marc B Rosen- man, Sonya E Shooshan, Laritza Rodriguez, Sameer Antani, George R Thoma, and Cl...

    work page 2023

  6. [11]

    BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, V olume1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Asso- ciation for Computational Linguistics. Wenjun...

    work page 2019

  7. [12]

    In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9043–9056, Miami, Florida, USA

    ICON: Improving inter- report consistency in radiology report generation via lesion-aware mixup augmentation. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9043–9056, Miami, Florida, USA. As- sociation for Computational Linguistics. Wenjun Hou, Yi Cheng, Kaishuai Xu, Wenjie Li, and Jiang Liu. 2023a. RECAP: Towards precise...

    work page 2024

  8. [13]

    Preprint, arXiv:2311.13668

    Maira-1: A spe- cialised large multimodal model for radiology report generation. Preprint, arXiv:2311.13668. Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn L. Ball, Katie S. Shpan- skaya, Jayne Seekins, David A. Mong, Safwan S. Ha- labi, Jesse K. Sandberg, Ricky Jones, David...

    arXiv

Show all 41 references
  1. [14]

    CoRR, abs/2106.14463

    Radgraph: Extracting clinical enti- ties and relations from radiology reports. CoRR, abs/2106.14463. Haibo Jin, Haoxuan Che, Yi Lin, and Hao Chen

    arXiv

  2. [15]

    Preprint, arXiv:2308.12604

    Promptmrg: Diagnosis-driven prompts for medical report generation. Preprint, arXiv:2308.12604. Baoyu Jing, Pengtao Xie, and Eric P. Xing

    arXiv

  3. [17]

    arXiv preprint arXiv:1901.07042

    Mimic-cxr-jpg, a large pub- licly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042. Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Nau- mann, Hoifung Poon, and Jianfeng Gao. 2023a. Llava-med: Training a l...

    1901 arXiv

  4. [18]

    Hybrid retrieval-generation reinforced agent for medical image report generation. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 1537–1547. Chin-Yew Lin

    2018

  5. [19]

    In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 13753–13762

    Exploring and distilling posterior and prior knowledge for radiology report generation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 13753–13762. Computer Vision Foundation / IEEE. Haotian Liu, Chunyuan Li, Qingyang...

    2021

  6. [20]

    Preprint, arXiv:2304.08485

    Visual instruction tuning. Preprint, arXiv:2304.08485. Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo

    arXiv

  7. [21]

    In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1235–1243, On- line

    Learning to generate clinically coherent chest X-ray reports. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1235–1243, On- line. Association for Computational Linguistics. Yasuhide Miura, Yuhao Zhang, Emily Tsai, Curtis Lan- glotz, and Dan Jurafsky

    2020

  8. [22]

    In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5288–5304, Online

    Improving factual completeness and consistency of image-to-text radi- ology report generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5288–5304, Online. Associatio...

    2021

  9. [24]

    In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2824–2832, Punta Cana, Dominican Re- public

    Progressive transformer-based generation of radiology reports. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2824–2832, Punta Cana, Dominican Re- public. Association for Computational Linguistics. OpenAI

    2021

  10. [25]

    Preprint, arXiv:2408.16213

    M4cxr: Explor- ing multi-task potentials of multi-modal large lan- guage models for chest x-ray interpretation. Preprint, arXiv:2408.16213. Chantal Pellegrini, Ege Özsoy, Benjamin Busam, Nas- sir Navab, and Matthias Keicher

    arXiv

  11. [26]

    Preprint, arXiv:2311.18681

    Radialog: A large vision-language model for radiology report generation and conversational assistance. Preprint, arXiv:2311.18681. Han Qin and Yan Song

    arXiv

  12. [27]

    In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 448–458

    Reinforced cross- modal alignment for radiology report generation. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 448–458. Association for Computational Linguistics. Vignav Ramesh, Nathan Andrew Chi, and Pranav R...

    2022

  13. [28]

    Preprint, arXiv:2210.06340

    Improving radiology report generation systems by removing hallucinated references to non- existent priors. Preprint, arXiv:2210.06340. Mercy Ranjit, Gopinath Ganapathy, Ranjit Manuel, and Tanuja Ganu

    arXiv

  14. [29]

    Preprint, arXiv:2305.03660

    Retrieval augmented chest x-ray report generation using openai gpt models. Preprint, arXiv:2305.03660. Karan Singhal, Shekoofeh Azizi, Tao Tu, S. Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, Perry Payne, Martin Senev...

    arXiv

  15. [30]

    Preprint, arXiv:2212.13138

    Large language models encode clinical knowledge. Preprint, arXiv:2212.13138. Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Pa- reek, Andrew Ng, and Matthew Lungren

    arXiv

  16. [31]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1500–1519, Online

    Com- bining automatic labelers and expert annotations for accurate radiology report labeling using BERT. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1500–1519, Online. Association for Computa- tional Linguistics. Xia...

    2020

  17. [33]

    In Proceedings of the 23rd Workshop on Biomedical Natural Language Processing, pages 440–448, Bangkok, Thailand

    XrayGPT: Chest radiographs summarization using large med- ical vision-language models. In Proceedings of the 23rd Workshop on Biomedical Natural Language Processing, pages 440–448, Bangkok, Thailand. As- sociation for Computational Linguistics. Zhanyu Wang, Lingqiao Liu, Lei W...

    arXiv

  18. [34]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online

    Trans- formers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Y...

    2020

  19. [35]

    Preprint, arXiv:2308.02463

    Towards generalist foundation model for radiology by leveraging web-scale 2d&3d medical data. Preprint, arXiv:2308.02463. Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, Shrikant Kendre, Jieyu Zha...

    arXiv

  20. [36]

    Preprint, arXiv:2408.08872

    xgen-mm (blip- 3): A family of open large multimodal models. Preprint, arXiv:2408.08872. Benjamin Yan, Ruochen Liu, David Kuo, Subathra Adithan, Eduardo Reis, Stephen Kwak, Vasan- tha Venugopal, Chloe O’Connell, Agustina Saenz, Pranav Rajpurkar, and Michael Moor

  21. [37]

    In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14676–14688, Singapore

    Style- aware radiology report generation with RadGraph and few-shot prompting. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14676–14688, Singapore. Association for Computational Linguistics. Shuxin Yang, Xian Wu, Shen Ge, Shaohua Kevin Zhou, ...

    2023

  22. [38]

    CoRR, abs/2112.15009

    Knowledge matters: Radiology report generation with general and specific knowl- edge. CoRR, abs/2112.15009. 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, An...

    arXiv

  23. [39]

    Preprint, arXiv:2303.15343

    Sigmoid loss for language image pre-training. Preprint, arXiv:2303.15343. Xi Zhang, Zaiqiao Meng, Jake Lever, and Edmond S. L. Ho

    arXiv

  24. [40]

    Preprint, arXiv:2411.19378

    Libra: Leveraging temporal im- ages for biomedical radiology analysis. Preprint, arXiv:2411.19378. Hong-Yu Zhou, Subathra Adithan, Julián Nicolás Acosta, Eric J. Topol, and Pranav Rajpurkar

    arXiv

  25. [41]

    Preprint, arXiv:2405.07988

    A generalist learner for multifaceted medical image interpretation. Preprint, arXiv:2405.07988. A Appendix A.1 Full List of Specialists In addition to specialist baselines in Table 1, the fol- lowing baselines are included: R2G EN (Chen et al., 2020), R2G ENCMN (Chen et al., 2...

    2020 arXiv

  26. [2018]

    In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, V olume1: Long Papers, pages 2577–2586

    On the automatic generation of medical imaging re- ports. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, V olume1: Long Papers, pages 2577–2586. Associa- tion for Computational Lingu...

    2018

  27. [2019]

    Preprint, arXiv:1904.03323

    Publicly available clinical bert embeddings. Preprint, arXiv:1904.03323. Satanjeev Banerjee and Alon Lavie

    1904 arXiv

  28. [2020]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1439–1449, Online

    Generating radiology reports via memory-driven transformer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1439–1449, Online. Association for Computational Linguistics. Zhihong Chen, Maya Varma, Jean-Benoit Delbrouck, M...

    2020

  29. [2021]

    Cross-modal memory net- works for radiology report generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (V olume 1: Long Papers), Virt...

    2021

  30. [2022]

    In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Online

    Factual accuracy is not enough: Planning consistent description order for radiology report generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Online. Association for Com- putational Linguistics. Farhad Nooralahzadeh, Nicolas ...

    2022

  31. [2023]

    Preprint, arXiv:2301.04558

    Learning to exploit temporal structure for biomedical vision- language processing. Preprint, arXiv:2301.04558. Pierre Chambon, Jean-Benoit Delbrouck, Thomas Sounack, Shih-Cheng Huang, Zhihong Chen, Maya Varma, Steven QH Truong, Chu The Chuong, and Curtis P. Langlotz

    arXiv

  32. [2024]

    Preprint, arXiv:2404.14219

    Phi-3 technical report: A highly capable language model locally on your phone. Preprint, arXiv:2404.14219. Emily Alsentzer, John R. Murphy, Willie Boag, Wei- Hung Weng, Di Jin, Tristan Naumann, and Matthew B. A. McDermott

    arXiv

  33. [2025]

    Fact-aware multimodal re- trieval augmentation for accurate medical radiol- ogy report generation. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (V olume1: Long Papers)...

    2025

Pith tools

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