Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

ReXVQA: A Large-scale Visual Question Answering Benchmark for Generalist Chest X-ray Understanding

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 A chest X-ray benchmark with roughly 696,000 questions built from radiology reports finds that the medical model MedGemma answers multiple-choice questions more accurately than three radiology residents, including the best resident, on a…

desk verdict A genuinely large and carefully constructed chest X-ray VQA benchmark, but the model-beats-residents headline is not supported by the current design. read the letter →

arxiv 2506.04353 v1 pith:QQYKMCIF submitted 2025-06-04 cs.CV cs.AIcs.CEcs.CLcs.LG

Ankit Pal , Jung-Oh Lee , Xiaoman Zhang , Malaikannan Sankarasubbu , Seunghyeon Roh , Won Jung Kim , Meesun Lee , Pranav Rajpurkar This is my paper · ORCID
classification cs.CVcs.AIcs.CEcs.CLcs.LG
keywords chestX-rayvisualquestionansweringradiologyreasoningbenchmarkmultimodallargelanguagemodelsMedGemmahumanreaderstudynegationdetectiondifferentialdiagnosismedicalAIevaluation
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

ReXVQA is a chest X-ray benchmark built to test whether multimodal AI models can do the reasoning a radiologist actually does, not just flag pathologies but judge presence, location, negation, differential diagnosis, and geometry. The paper generates roughly 696,000 multiple-choice questions from about 160,000 studies with paired radiology reports, validates a fraction of them with radiologists, and evaluates eight multimodal large language models. Its central result is that MedGemma, a four-billion-parameter medical vision-language model, reaches 83.24% overall accuracy, and in a 200-case reader study scores 83.84% against 77.27% for the best of three radiology residents. The authors read this as the first demonstration that an AI system can consistently exceed expert human evaluation in chest X-ray interpretation, and they present the benchmark as a new standard testbed for generalist radiological AI.

What carries the argument

The argument is carried by a three-layer dataset pipeline that turns paired chest X-rays and radiology reports into multiple-choice questions: GPT-4o first converts each report into structured bullet points, then into MCQs using prompts refined over twelve rounds of radiologist feedback; structural and content validators, a ClinicalBERT protected-health-information screen, and a cosine-similarity diversity filter clean the output; and a stratified sample of questions goes to board-certified radiologists for expert review. The five-category cognitive taxonomy, negation, presence, differential diagnosis, location, and geometry, defines what the benchmark measures, and the 200-case reader study with three residents provides the human comparison.

What would settle it

Randomly sample about five hundred questions from the roughly ninety percent of ReXVQA that received no expert review, have board-certified radiologists answer them from the X-ray images and original reports independently, and score the models against that adjudicated ground truth. If the corrected labels change which model ranks first or push MedGemma below the best resident, the paper's central claim about exceeding expert human evaluation is refuted.

Watch

Extended reading notes

Core claim

ReXVQA is offered as the largest chest X-ray visual question answering benchmark: about 696,000 four-option multiple-choice questions paired with 160,000 studies, generated from radiology reports rather than fixed templates. The paper's central finding is that the medical vision-language model MedGemma achieves 83.24% overall accuracy on the private test set and 83.84% on 200 randomly sampled cases, compared with 77.27% for the best of three radiology residents, which the authors call the first instance where AI consistently surpasses expert human evaluation in chest X-ray interpretation. The benchmark is built to test five radiological reasoning skills, presence assessment, location analysis, negation detection, differential diagnosis, and geometric reasoning, and the paper reports category-level and anatomical breakdowns showing MedGemma leading in most of them.

Load-bearing premise

The benchmark's ground-truth answers come from radiology reports that GPT-4o turns into multiple-choice questions, and the paper assumes those reports are accurate and complete descriptions of the X-rays and that the generated questions preserve that accuracy; only a small fraction of the questions were checked by radiologists.

Editorial extensions

If this is right

  • Model rankings on ReXVQA reflect five distinct reasoning skills rather than a single classification score, so a high overall number requires competence in negation, localization, and differential diagnosis together.
  • Because the same report-derived questions are released with public leaderboards and a private test set, future models can be compared on the same clinical-reasoning tasks without seeing the test answers.
  • If the reported accuracy holds up under independent adjudication, it would mean a four-billion-parameter medical model can outperform the measured performance of radiology residents on this style of multiple-choice chest X-ray interpretation.
  • The task and anatomical breakdowns identify where generalist models still lag, notably differential diagnosis and geometric reasoning, and thereby point to the skills next-generation medical AI systems need to improve.

Reading between the lines

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

  • The 'AI beats experts' headline is currently a claim about MCQ accuracy on report-derived ground truth; whether it extends to clinical reading of real X-rays requires a study where the reference standard is set by independent multi-reader adjudication rather than by the same reports used to write the questions.
  • Since only about ten percent of questions were expert-reviewed, auditing the unreviewed ninety percent, especially negation questions where the initial review found multiple valid answers, could change model rankings and is the cheapest test of the benchmark's validity.
  • If LLMs are better at multiple-choice exam-taking than at producing free-text interpretations, the MCQ format may overstate their clinical reasoning ability; the paper's own suggestion of open-ended questions is the direct way to check this.
  • The moderate human-model agreement reported in the paper suggests that even when models match or beat residents on accuracy, their reasoning paths may differ, so accuracy alone should not be read as clinical trust.
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 presents ReXVQA, a large-scale visual question answering benchmark for chest X-ray interpretation, comprising approximately 696,000 multiple-choice questions paired with 160,000 chest X-ray studies derived from the ReXGradient-160K dataset. Questions are generated by converting radiology reports into bullet points and then into MCQs using GPT-4o, with a validation pipeline including structural checks, compliance screening, and limited expert review. The authors evaluate eight multimodal LLMs and report that MedGemma achieves 83.24% overall accuracy on the private test set. They also conduct a reader study with three radiology residents on 200 cases, reporting that MedGemma (83.84%) outperforms the best resident (77.27%), and interpret this as the first instance where AI consistently exceeds expert human evaluation in chest X-ray interpretation.

Significance. If the results hold, ReXVQA would be a valuable large-scale resource for evaluating generalist chest X-ray models, with a public leaderboard, fine-grained task splits, and coverage of multiple reasoning types. The evaluation of eight models provides useful comparative data for the community, and the release of a large curated MCQ benchmark is a contribution in itself. However, the headline claim that AI surpasses expert human evaluation is not fully supported by the current evidence, because the reader study's ground truth is derived from reports rather than images and the validation of image-answerability is limited. The benchmark itself may be useful for training and evaluation, but the human-comparison conclusion needs additional support before it can be considered established.

major comments (4)
  1. [Section 6.1 / 3.3.1 / 3.3.3] The reader study's validity is threatened by the provenance of the ground truth. Section 3.3.1 states that questions are generated exclusively from radiology reports, and Section 6.1 compares human readers, who see only the image and the question, against AI models using the same report-derived answers as ground truth. Chest X-ray reports commonly contain information that is not present in the image, such as clinical history, comparison with prior examinations, and laboratory data. The 200-case image-alignment audit (Section 3.3.3) found only one misalignment, but that audit was not necessarily performed on the same 200 cases used in the reader study, and the subsequent 300-question validation set is also not explicitly matched to the reader-study cases. If even a small fraction of the reader-study questions are not answerable from the image alone, human accuracy would be systematically lowered. Since the headline claim depends on the 6.6-percentage-point gap between MedGemma (83.84%) and the best resident (77.27%), the authors should either report an independent audit of the exact 200 reader-study questions confirming that each is answerable from the image alone, or re-perform the reader study on a subset of questions with verified image-answerability.
  2. [Abstract / Section 7] The claim that MedGemma's performance 'represents the first instance where AI consistently exceeds expert human evaluation in chest X-ray interpretation' is overbroad. Prior studies (e.g., Tiu et al., 2022, cited in the paper) have reported AI achieving expert-level or superior performance on specific chest X-ray tasks. Furthermore, the evidence here is a single reader study with three residents and 200 cases; the term 'consistently' is not supported by the data. I recommend limiting the claim to the specific comparison in this study, e.g., 'MedGemma outperformed the three participating residents on this benchmark.'
  3. [Abstract / Section 3.5 / Table 1] The dataset size is reported inconsistently: the abstract says approximately 696,000 questions, Section 3.5 text sums to 695,663, and Table 1 sums to 694,841. In addition, the validation set count is given as 40,878 in the text but 40,858 in Table 1, and the public test set count differs (40,826 in text vs 40,557 in Table 1). These discrepancies must be reconciled, and a single authoritative set of counts should be used throughout the paper.
  4. [Section 3.3.3] The expert validation of the benchmark is limited in scale. Only 100 questions per 1,000 are reviewed, and the initial review of 120 questions found 10.8% with multiple valid answers and 3.3% with hallucinated content. While the authors state these issues were fixed, the subsequent validation of 300 questions with 2 errors is small relative to the approximately 41,000 private test questions and is not tied to the specific cases used in the reader study. This leaves uncertainty about the error rate in the final benchmark, which directly affects both model rankings and the human-AI comparison. A larger independent validation of the test set, or at least an error-rate estimate with confidence intervals, would strengthen the paper.
minor comments (5)
  1. [Section 3.3.2] Equation (3) defines diversity using only a semantic term, despite the text stating that both semantic and structural features are used; the structural term is missing or undefined.
  2. [Section 3.3.3] The citation to Warner et al. (2024) for 'sequential checks' appears to refer to the ModernBERT paper and is likely a citation error; the intended method reference should be checked.
  3. [Section 4.2 / Section 6.1] The text states that models receive all available images for multi-view studies, but it is not clear whether the human readers also received all views; this should be reported to ensure a fair comparison.
  4. [Figure 2] The caption says the bar chart shows overall accuracy, but the figure includes error bars; the source of these error bars (e.g., bootstrap or standard error) is not described in the text.
  5. [Various] There are numerous typographical errors, including 'T ask' in Figure 3, 'L VMs' in Figure 1 caption, 'T rain' in Table 1 headers, 'F ormat' in Section 3.3.2, and inconsistent spacing in author names (e.g., 'W on Jung Kim').

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ReXVQA's model-vs-human result is an empirical evaluation against fixed report-derived labels, not a derivation that reduces to its inputs.

full rationale

The paper's central claims are benchmark construction and comparative evaluation; neither is a prediction derived from fitted parameters. The most plausible circularity candidate is the report-only generation pipeline (Section 3.3: 'our pipeline utilizes radiology reports exclusively') combined with the reader study that scores residents against these labels. But this is not circular: the labels are fixed external data, and the model and human readers are both scored on the same questions. The outcome is contingent (LLaVA scores 24.75%, MedGemma 83.84%), so it is not forced by construction. The initial expert review finding 10.8% multiple-valid-answer questions and the small 200-question image-alignment audit are validity caveats about the benchmark, not definitional identities, and the paper itself discloses only limited external validation. The source dataset ReXGradient-160K is cited from overlapping authors, but it is a public, externally inspectable resource rather than an unverified theorem invoked to forbid alternatives. No step in the paper reduces to its own inputs by the paper's equations or definitions.

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

The benchmark construction depends on several thresholds and domain assumptions. The free parameters (diversity threshold, compliance threshold, quality threshold) control which questions remain in the dataset. The key domain assumptions are that radiology reports accurately reflect the images and that GPT-4o faithfully converts them into clinically valid MCQs.

free parameters (3)
  • diversity threshold tau_diversity = 0.9
    Empirically set based on expert evaluation of question similarity; used in Eq. (4) to filter redundant questions.
  • compliance threshold theta_c = unspecified
    Binary threshold in Eq. (6) for PHI/HIPAA compliance filtering; its value is not reported.
  • quality threshold for difficulty filtering = unspecified
    Used to remove questions judged too elementary or lacking radiological specificity; exact value not given.
assumptions (4)
  • domain assumption Radiology reports are accurate and complete reflections of the chest X-ray findings
    The entire dataset generation pipeline (Section 3.3.1) converts reports into bullets and MCQs; if reports contain errors, the benchmark answers are wrong.
  • domain assumption GPT-4o faithfully converts report content into clinically valid MCQs without introducing hallucinations
    The authors report hallucination issues in initial review (3.3% in the first 120 questions) and assume the refined prompts fixed this.
  • domain assumption The five cognitive task categories (presence, location, negation, differential diagnosis, geometry) adequately capture core radiological reasoning
    Selected based on prior VQA literature and radiologist consultation, but no formal justification that these are sufficient or exhaustive.
  • domain assumption A sample of 100 questions per 1,000 generated is representative for expert validation
    Expert review only covers about 10% of the dataset; the assumption is that the sampled questions reflect the quality of the rest.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReXVQA: A Large-scale Visual Question Answering Benchmark for Generalist Chest X-ray Understanding." pith.science (2026). https://pith.science/paper/QQYKMCIF

@misc{pith2026250604353,
  author       = {Pith},
  title        = {Pith review of: ReXVQA: A Large-scale Visual Question Answering Benchmark for Generalist Chest X-ray Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QQYKMCIF}},
  note         = {Machine review of arXiv:2506.04353}
}
read the original abstract

We present ReXVQA, the largest and most comprehensive benchmark for visual question answering (VQA) in chest radiology, comprising approximately 696,000 questions paired with 160,000 chest X-rays studies across training, validation, and test sets. Unlike prior efforts that rely heavily on template based queries, ReXVQA introduces a diverse and clinically authentic task suite reflecting five core radiological reasoning skills: presence assessment, location analysis, negation detection, differential diagnosis, and geometric reasoning. We evaluate eight state-of-the-art multimodal large language models, including MedGemma-4B-it, Qwen2.5-VL, Janus-Pro-7B, and Eagle2-9B. The best-performing model (MedGemma) achieves 83.24% overall accuracy. To bridge the gap between AI performance and clinical expertise, we conducted a comprehensive human reader study involving 3 radiology residents on 200 randomly sampled cases. Our evaluation demonstrates that MedGemma achieved superior performance (83.84% accuracy) compared to human readers (best radiology resident: 77.27%), representing a significant milestone where AI performance exceeds expert human evaluation on chest X-ray interpretation. The reader study reveals distinct performance patterns between AI models and human experts, with strong inter-reader agreement among radiologists while showing more variable agreement patterns between human readers and AI models. ReXVQA establishes a new standard for evaluating generalist radiological AI systems, offering public leaderboards, fine-grained evaluation splits, structured explanations, and category-level breakdowns. This benchmark lays the foundation for next-generation AI systems capable of mimicking expert-level clinical reasoning beyond narrow pathology classification. Our dataset will be open-sourced at https://huggingface.co/datasets/rajpurkarlab/ReXVQA

Figures

Figures reproduced from arXiv: 2506.04353 by the authors.

Figure 1
Figure 1. Sample from the ReXVQA dataset, where human radiologists correctly identified mild scarring in the left lung base (correct answer B), while three state-of-the-art LVMs (Gemini, Qwen-2.5, and Phi-3.5) provided incorrect assessments, misiden￾tifying the condition as pleural effusion or consolidation. 2023). However, they remain fundamentally limited to a fixed set of predetermined labels and cannot engage in the broad… view at source ↗
Figure 2
Figure 2. Performance comparison of AI models and human readers across 200 random sampled cases. The bar chart shows overall accuracy (%) for eight AI models (Eagle2, Gemini, Janus, LLaVA, Phi35, Qwen2VL, Qwen25VL, and MedGemma) and three human readers. thentic clinical reasoning patterns rather than artificial templates [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Expert-Guided Medical MCQ Generation Pipeline: We propose a three￾layer approach combining computational processes and expert oversight for cre￾ating high-quality radiology MCQs. {Oi1, Oi2, Oi3, Oi4}. The task requires models to analyze both the visual input Ii and the textual question Qi to select the correct answer(s) from the option set. The ground truth label for each data point is defined as y ∈ R 1 where y i =… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Hierarchical Taxonomy of Chest X-Ray Categories. This expert-validated classification system, developed in collaboration with radiologists, organizes chest X-ray findings into five major domains: Clinical Assessment, Respiratory System, Cardiovascular, Medical Devices,…
Figure 5
Figure 5. Figure 5: Interrater Agreement Analysis Across Different Medical Assessment Tasks. We plot a heatmap showing Cohen’s Kappa coefficients for interrater agreement be￾tween eight AI models and three radiologist readers across five medical imaging assessment tasks and a combined ana…
Figure 6
Figure 6. Figure 6: The radiology image tagging platform interface used for expert an￾notation. The platform displays a portable chest X-ray with associated meta￾data, categorization fields, and multiple-choice assessment options. The interface includes structured fields for capturing fin…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.CV 2026-08 conditional novelty 6.0 of 10

    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, V...

Reference graph

Works this paper leans on

29 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Hasan, Vivek Datla, Joey Liu, Dina Demner-Fushman, and Henning M \"u ller

    Asma Ben Abacha, Sadid A. Hasan, Vivek Datla, Joey Liu, Dina Demner-Fushman, and Henning M \"u ller. Vqa-med: Overview of the medical visual question answering task at imageclef 2019. In Conference and Labs of the Evaluation Forum, 2019. URL https://api.semanticscholar.org/CorpusID:198489641

    work page 2019

  2. [2]

    Hasan, Dina Demner-Fushman, and Henning M \"u ller

    Asma Ben Abacha, Vivek Datla, Sadid A. Hasan, Dina Demner-Fushman, and Henning M \"u ller. Overview of the vqa-med task at imageclef 2020: Visual question answering and generation in the medical domain. In Conference and Labs of the Evaluation Forum, 2020. URL https://api.semanticscholar.org/CorpusID:225073930

    work page 2020

  3. [3]

    Qwen-vl: An open-source vision-language model by alibaba

    Alibaba DAMO Academy. Qwen-vl: An open-source vision-language model by alibaba. 2024. URL https://modelscope.cn/models/damo/qwen_vl_base/summary

    work page 2024

  4. [5]

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

    Seongsu Bae, Daeun Kyung, Jaehee Ryu, Eunbyeol Cho, Gyubok Lee, Sunjun Kweon, Jungwoo Oh, Lei Ji, Eric Chang, Tackeun Kim, and Edward Choi. Mimic-ext-mimic-cxr-vqa: A complex, diverse, and large-scale visual question answering dataset for chest x-ray images, 2024 b . URL https://doi.org/10.13026/deqx-d943

    doi:10.13026/deqx-d943 2024

  5. [6]

    Ehrxqa: A multi-modal question answering dataset for electronic health records with chest x-ray images

    Seongsu Bae, Daeun Kyung, Jaehee Ryu, Eunbyeol Cho, Gyubok Lee, Sunjun Kweon, Jungwoo Oh, Lei Ji, Eric Chang, Tackeun Kim, et al. Ehrxqa: A multi-modal question answering dataset for electronic health records with chest x-ray images. Advances in Neural Information Processing Systems, 36, 2024 c

    work page 2024

  6. [7]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  7. [8]

    Janus-pro: Unified multimodal understanding and generation with data and model scaling

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-pro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025

    arXiv 2025

  8. [9]

    Gemini pro vision: Google's multimodal model for complex reasoning

    Google. Gemini pro vision: Google's multimodal model for complex reasoning. 2024. URL https://ai.google.com/research/gemini-pro-vision

    work page 2024

Show all 29 references
  1. [10]

    Medgemma hugging face

    Google. Medgemma hugging face. https://huggingface.co/collections/google/medgemma-release-680aade845f90bec6a3f60c4, 2025. Accessed: [Insert Date Accessed, e.g., 2025-05-20]

    2025

  2. [11]

    Zeng, and Long H

    Sergey Goryachev, Margarita Sordo, Qing T. Zeng, and Long H. Ngo. Implementation and evaluation of four different methods of negation detection. 2007. URL https://api.semanticscholar.org/CorpusID:11699116

    2007

  3. [12]

    An, Mengliang Zhang, Liangchen Liu, Kazuma Kobayashi, Tatsuya Harada, Ronald M

    Xinyue Hu, Lin Gu, Qi A. 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. Proceedings of the ...

    2023

  4. [13]

    Medical-Diff-VQA: A Large-Scale Medical Dataset for Difference Visual Question Answering on Chest X-Ray Images (version 1.0.1)

    Xinyue Hu, Lin Gu, Qiyuan An, Mengliang Zhang, liangchen liu, Kazuma Kobayashi, Tatsuya Harada, Ronald Summers, and Yingying Zhu. Medical-Diff-VQA: A Large-Scale Medical Dataset for Difference Visual Question Answering on Chest X-Ray Images (version 1.0.1) . https://doi.org/10...

    doi

  5. [14]

    Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman

    Jason J. Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images. 2018 a . URL https://api.semanticscholar.org/CorpusID:53712941

    2018

  6. [15]

    Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman

    Jason J. Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. Descriptor : A dataset of clinically generated visual questions and answers about radiology images. 2018 b . URL https://api.semanticscholar.org/CorpusID:53712941

    2018

  7. [16]

    Eagle 2: Building post-training data strategies from scratch for frontier vision-language models

    Zhiqi Li, Guo Chen, Shilong Liu, Shihao Wang, Vibashan VS, Yishen Ji, Shiyi Lan, Hao Zhang, Yilin Zhao, Subhashree Radhakrishnan, et al. Eagle 2: Building post-training data strategies from scratch for frontier vision-language models. arXiv preprint arXiv:2501.14818, 2025

    2025 arXiv

  8. [17]

    Gemex: A large-scale, groundable, and explainable medical vqa benchmark for chest x-ray diagnosis

    Bo Liu, Kevin Yingyin Zou, Li-Ming Zhan, Zexin Lu, Xiaoyu Dong, Yidi Chen, Chengqiang Xie, Jiannong Cao, Xiao-Ming Wu, and Huazhu Fu. Gemex: A large-scale, groundable, and explainable medical vqa benchmark for chest x-ray diagnosis. ArXiv, abs/2411.16778, 2024. URL https://api...

    2024 arXiv

  9. [18]

    Krumholz, Jure Leskovec, Eric J

    Michael Moor, Oishi Banerjee, Zahra F H Abad, Harlan M. Krumholz, Jure Leskovec, Eric J. Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616: 0 259--265, 2023. URL https://api.semanticscholar.org/CorpusID:258083369

    2023

  10. [19]

    Clinicalbert fine-tuned for medical note de-identification

    OBI\_Lab. Clinicalbert fine-tuned for medical note de-identification. https://huggingface.co/obi/deid_bert_i2b2, 2022. Model fine-tuned on the I2B2 2014 dataset for PHI detection

    2022

  11. [20]

    Gemini goes to med school: Exploring the capabilities of multimodal large language models on medical challenge problems & hallucinations

    Ankit Pal and Malaikannan Sankarasubbu. Gemini goes to med school: Exploring the capabilities of multimodal large language models on medical challenge problems & hallucinations. In Clinical Natural Language Processing Workshop, 2024. URL https://api.semanticscholar.org/CorpusI...

    2024

  12. [21]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), ...

    2019 arXiv

  13. [22]

    Phi-2 vision: Microsoft's vision-language model

    Microsoft Research. Phi-2 vision: Microsoft's vision-language model. 2024. URL https://www.microsoft.com/en-us/research/project/phi-2-vision/

    2024

  14. [23]

    Llava 1.5: Open-source vision-language model

    LLaVA Team. Llava 1.5: Open-source vision-language model. 2024. URL https://github.com/haotian-liu/LLaVA

    2024

  15. [24]

    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 0 (12): 0 1399--1406, 2022

    2022

  16. [25]

    Interactive computer-aided diagnosis on medical image using large language models

    Sheng Wang, Zihao Zhao, Ouyang Xi, Tianming Liu, Qian Wang, and Dinggang Shen. Interactive computer-aided diagnosis on medical image using large language models. Communications Engineering, 3, 2024. URL https://api.semanticscholar.org/CorpusID:272692772

    2024

  17. [26]

    Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference

    Benjamin Warner, Antoine Chaffin, Benjamin Clavi \'e , Orion Weller, Oskar Hallstr \"o m, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli. Smarter, better, faster, longer: A modern bidirec...

    2024 arXiv

  18. [27]

    Medklip: Medical knowledge enhanced language-image pre-training for x-ray diagnosis

    Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Medklip: Medical knowledge enhanced language-image pre-training for x-ray diagnosis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21372--21383, 2023

    2023

  19. [28]

    Knowledge-enhanced visual-language pre-training on chest radiology images

    Xiaoman Zhang, Chaoyi Wu, Ya Zhang, Weidi Xie, and Yanfeng Wang. Knowledge-enhanced visual-language pre-training on chest radiology images. Nature Communications, 14 0 (1): 0 4542, 2023

    2023

  20. [29]

    Development of a large-scale medical visual question-answering dataset

    Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Development of a large-scale medical visual question-answering dataset. Communications Medicine, 4 0 (1): 0 277, 2024

    2024

  21. [30]

    Acosta, Josh Miller, Ouwen Huang, and Pranav Rajpurkar

    Xiaoman Zhang, Julián N. Acosta, Josh Miller, Ouwen Huang, and Pranav Rajpurkar. Rexgradient-160k: A large-scale publicly available dataset of chest radiographs with free-text reports. In arXiv:2505.00228v1, 2025

    2025 arXiv

Pith tools

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