Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

CXRAgent claims that routing every tool output through an evidence-driven validator, then adaptively assembling specialist teams, yields state-of-the-art chest X-ray interpretation on three public benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 08:17 UTC pith:W5SICU2D

load-bearing objection A sensible multi-stage agent for CXR interpretation with an interesting validator and adaptive team modes, but the claim of state-of-the-art rests on point estimates with no error bars and several margins are within sampling noise. the 3 major comments →

arxiv 2510.21324 v2 pith:W5SICU2D submitted 2025-10-24 cs.AI cs.MA

CXRAgent: Director-Orchestrated Multi-Stage Reasoning for Chest X-Ray Interpretation

Jinhui Lou , Yan Yang , Zhou Yu , Zhenqi Fu , Weidong Han , Qingming Huang , Jun Yu This is my paper
classification cs.AI cs.MA
keywords chest X-ray interpretationmedical AI agentmultimodal large language modelevidence-driven validationtool orchestrationadaptive team collaborationradiology report generationmedical visual question answering
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that the best way to build a chest X-ray interpretation agent is not to aggregate tool outputs directly, but to route every tool result through an evidence-driven validator before any diagnosis is formed. It proposes CXRAgent, in which a multimodal LLM acts as a director: it calls specialized analysis tools, asks a validator to state supporting and refuting visual evidence for each tool claim, assigns a confidence rating, then chooses how much specialist collaboration the case needs. On three public CXR benchmarks the paper reports top accuracies—67.0% on a comprehensive benchmark, 75.6% on a visual-question-answering benchmark, and a top entity-level score on report generation. A sympathetic reader should care because the approach directly targets the problem that makes prior agents fragile: when tools disagree, a system needs a principled way to decide which finding is grounded in the actual image.

Core claim

On its own terms, the paper's discovery is that tool-mediated diagnosis can be made substantially more reliable by interposing a validation layer between tool invocation and reasoning. The Evidence-driven Validator (EDV) reformats each tool's conclusion, identifies image features that support it, identifies image features that contradict it, and attaches a qualitative confidence rating. The director then selects among four collaboration strategies—skip, relay, dispatch, or probe—and assembles a virtual specialist team matched to the case. In the reported ablations, each addition (tools, then EDV, then team collaboration) improves accuracy, and the full system yields stated gains over both mo

What carries the argument

The central object is the Evidence-driven Validator (EDV), a tool-agnostic module that, for each diagnostic statement from any tool, produces a reformatted conclusion, supportive visual findings, refuting visual findings, and a qualitative confidence rating. The paper uses EDV as the gate that resolves conflicting tool outputs and creates a unified evidence format; the director then maps case complexity to one of four team collaboration modes (Skip, Relay, Dispatch, Probe) before a final evidence-backed synthesis.

Load-bearing premise

The load-bearing premise is that the same multimodal LLM that directs the case can also genuinely check tools' claims against the image rather than just rationalizing its own prior leanings; if that premise fails, the evidence-validation stage adds formatting without real grounding.

What would settle it

A reader could rerun the ablation on only the test cases where the invoked tools disagree: if the full CXRAgent (with EDV) is not more accurate than the pipeline with tools but without EDV on exactly those cases, then the claimed conflict-resolution mechanism is not doing the work. A second check is to replace EDV's visual-evidence statements with evidence from an independent, frozen vision module; unchanged accuracy would show that the benefit comes from output normalization, not grounding.

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

If this is right

  • If a tool flags a finding with no supporting image evidence, the system can downgrade or discard it before it reaches the final diagnosis, reducing error propagation from individual tools.
  • Because the same evidence format is used for every tool, the agent can incorporate new CXR-analysis tools without retraining the coordinator.
  • The Skip/Relay/Dispatch/Probe strategy means the system spends more computation on ambiguous cases and less on clear ones, making deployment cost adaptive.
  • The final diagnosis is accompanied by supportive and refuting evidence, so a clinician can inspect why a finding was accepted or rejected.
  • The reported gains across disease identification, VQA, and report generation imply the same pipeline generalizes to tasks of different output formats.

Where Pith is reading between the lines

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

  • A testable extension is to replace EDV's self-reported evidence with evidence produced by an independent, frozen vision module; if accuracy does not drop, the claimed visual grounding could be mostly format normalization.
  • The same director-orchestrated template could transfer to other imaging modalities because neither the tool set nor the team logic is X-ray-specific; a CT or pathology instantiation would be a natural stress test.
  • The reported average accuracies do not distinguish easy from hard cases; logging per-case strategy choices would verify whether Skip is selected for genuinely unambiguous cases and Probe for genuinely ambiguous ones.
  • The confidence ratings produced by EDV are qualitative, so a clinician-facing deployment would need calibration of those ratings against actual error rates, especially for rare findings.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces CXRAgent, a director-orchestrated, multi-stage agent for chest X-ray interpretation. A multimodal LLM (GPT-4o or Qwen-VL-Max) serves as the central director that (1) invokes CXR-analysis tools via ReAct, (2) validates and normalizes tool outputs through an Evidence-driven Validator (EDV), (3) selects one of four collaboration modes (Skip, Relay, Dispatch, Probe) and recruits a specialist team, and (4) synthesizes a final evidence-backed diagnosis from team outputs and contextual memory. The system is evaluated on CheXbench, Medical-CXR-VQA, and MIMIC-CXR report generation, with comparisons against foundation models and the MedRAX agent. The paper claims state-of-the-art overall accuracy of 67.0% on CheXbench, 75.6% on Medical-CXR-VQA, and RaTEScore 0.513 on MIMIC-CXR report generation, with ablation studies attributing gains to tools, EDV, and team collaboration.

Significance. If the claims are confirmed, CXRAgent would be a meaningful advance in agentic medical image analysis, addressing the real problem of conflicting tool outputs and rigid single-pipeline agents. The paper's strengths include a clear conceptual design, a novel evidence-driven validation component, a clinically inspired adaptive team mechanism, and experiments across multiple CXR tasks with code/data links. However, the central state-of-the-art claim is currently supported only by point estimates from small random subsets without confidence intervals or significance tests, and the EDV's visual grounding is not independently verified. These issues are addressable but are load-bearing for the paper's main contribution, so substantial revision is required before the claims can be accepted.

major comments (3)
  1. [IV-A/IV-B, Tables I–III] The central SOTA claim is based on single point estimates from small random subsets. The paper states n=594/256/150 for Medical-CXR-VQA and n=400 for MIMIC-CXR; CheXbench subset sizes are not given. No confidence intervals, standard deviations, or significance tests are reported. With n=150, the View difference 83.3 vs 80.6 has a standard error of roughly 3 percentage points, so the 2.7-point margin is within sampling noise; the Presence margin (70.2 vs 67.5, n=594) is about 1 SE. The RaTEScore difference 0.513 vs 0.510 on n=400 is a 0.003 gap. Please report confidence intervals (e.g., bootstrap or exact binomial), paired significance tests (McNemar for accuracy; bootstrap for RaTEScore), specify CheXbench subset sizes, and disclose seeds/randomness. Without this, the 'state-of-the-art' headline is not statistically supported.
  2. [III-A, Fig. 2] EDV is the main novelty, but the supportive/refuting evidence and confidence ratings are generated by the same director LLM that requested the tool outputs. The paper does not provide any independent mechanism verifying that the generated evidence corresponds to actual image features; a plausible alternative is that EDV paraphrases the director's prior beliefs and assigns confidence accordingly. This weakens the causal interpretation of the ablation gains in Tables IV–VII. Please validate EDV evidence against annotated image regions or human review on a sample, or at least ablate with an externally grounded/evidence-checking component.
  3. [IV-C, Tables IV–VII] Ablation results are presented as point estimates with no variance, and several configurations show non-monotonic behavior: Table V shows VQA Rad-Restruct dropping from 60.8 to 55.7 with tools and jumping to 66.1 with EDV; Table IV shows BDC CheXpert dropping from 61.8 to 60.5 in the full model; Table V VC shows full model 74.0 vs. baseline 79.0. Without error bars or multiple seeds, the claims that 'the full model delivers optimal performance' and that components 'synergistically' combine are not established. Report per-configuration confidence intervals or at least repeated-run variability for the main ablation tables.
minor comments (6)
  1. [IV-A, Table I] The 'Overall' row in Table I appears to be an unweighted average across tasks with differing numbers of classes and likely different sample sizes. Clarify how 'Overall' is computed and report per-task N for CheXbench.
  2. [IV-A, Table III] LLMScore is implemented with GPT-4o as the judge while GPT-4o also serves as a director, creating a risk of judge bias toward GPT-4o-generated reports. Use a second judge or a reference-based metric (e.g., RadGraph-F1) as a sensitivity analysis.
  3. [III-B/Algorithm 1] The team recruitment function R(I,q,s) and the strategy-selection criterion for Skip/Relay/Dispatch/Probe are underspecified. Provide the prompt templates or decision rules used to choose the strategy, team size, roles, and missions; without this, the method is difficult to reproduce exactly.
  4. [IV-D] The two case studies are selective and do not provide average evidence. State explicitly that they are illustrative examples, not an evaluation of typical behavior.
  5. [Abstract] Typo: 'Code and data are valuable at this link' should be 'available at this link'.
  6. [Conclusion] The paper lacks a limitations section. Given the statistical and grounding concerns above, please add a discussion of these limitations and their implications for clinical deployment.

Circularity Check

0 steps flagged

No circularity: benchmark results come from external evaluations, not from fitted or self-defined quantities.

full rationale

CXRAgent's reported gains are measured on external benchmarks (CheXbench, Medical-CXR-VQA, MIMIC-CXR) using fixed pretrained tool models and standard metrics (accuracy, RaTEScore, GPT-4o-based LLMScore). No parameter is fitted to the target benchmark labels and then renamed as a prediction; the architecture is a prompting/orchestration framework with no trainable components tuned on the evaluation sets. The main performance claims therefore rest on externally observed accuracy numbers rather than on an equation that equates the output to an input. Two overlaps exist but do not constitute circularity in the derivation sense: (1) EDV's supporting/refuting evidence and confidence ratings are produced by the same multimodal LLM that serves as director (Section III-A), which is a visual-grounding reliability concern rather than a case where the benchmark answer is defined in terms of the confidence estimate; (2) GPT-4o is used both as a director and as the LLMScore judge, which may introduce evaluation bias but does not fit or define any reported result. The related self-citations [16]-[18] are background literature, not load-bearing justifications of the CXRAgent design or its state-of-the-art claim. Statistical concerns about small sample sizes and missing confidence intervals, while legitimate, are experimental-validity issues and are outside the circularity definition used here.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No numeric free parameters are fitted; the system is defined by prompts, tool sets, and strategy choices. The central claim rests on the capability of proprietary LLM directors, on LLM-generated 'visual evidence' being treated as grounding, and on benchmark labels/subset representativeness.

axioms (4)
  • domain assumption The director multimodal LLM (GPT-4o or Qwen-VL-Max) reliably performs ReAct planning, team recruitment, and final diagnosis synthesis.
    Sections III-A to III-C assume the director chooses correct tools, strategies, and team roles; no separate evaluation of planning quality is provided.
  • domain assumption EDV's supportive/refuting evidence and confidence assessments accurately reflect image content.
    Section III-A: EDV is implemented as LLM-generated statements; no independent vision grounding or verification is described.
  • domain assumption Benchmark labels in CheXbench, Medical-CXR-VQA, and MIMIC-CXR are correct and the random subsets are representative.
    Section IV-A: evaluation relies on sampled test subsets of 150-594 cases without stratification or error bars.
  • domain assumption RaTEScore and LLMScore (GPT-4o) are valid measures of report quality.
    Section IV-A uses these metrics; LLMScore is generated by GPT-4o, the same model family as one director, creating potential evaluator bias.

pith-pipeline@v1.3.0-alltime-deepseek · 15190 in / 11378 out tokens · 103375 ms · 2026-08-04T08:17:30.844695+00:00 · methodology

0 comments
read the original abstract

Chest X-ray (CXR) plays a pivotal role in clinical diagnosis, and a variety of task-specific and foundation models have been developed for automatic CXR interpretation. However, these models often struggle to adapt to new diagnostic tasks and complex reasoning scenarios. Recently, LLM-based agent models have emerged as a promising paradigm for CXR analysis, enhancing model's capability through tool coordination, multi-step reasoning, and team collaboration, etc. However, existing agents often rely on a single diagnostic pipeline and lack mechanisms for assessing tools' reliability, limiting their adaptability and credibility. To this end, we propose CXRAgent, a director-orchestrated, multi-stage agent for CXR interpretation, where a central director coordinates the following stages: (1) Tool Invocation: The agent strategically orchestrates a set of CXR-analysis tools, with outputs normalized and verified by the Evidence-driven Validator (EDV), which grounds diagnostic outputs with visual evidence to support reliable downstream diagnosis; (2) Diagnostic Planning: Guided by task requirements and intermediate findings, the agent formulates a targeted diagnostic plan. It then assembles an expert team accordingly, defining member roles and coordinating their interactions to enable adaptive and collaborative reasoning; (3) Collaborative Decision-making: The agent integrates insights from the expert team with accumulated contextual memories, synthesizing them into an evidence-backed diagnostic conclusion. Experiments on various CXR interpretation tasks show that CXRAgent delivers strong performance, providing visual evidence and generalizes well to clinical tasks of different complexity. Code and data are valuable at this \href{https://github.com/laojiahuo2003/CXRAgent/}{link}.

Figures

Figures reproduced from arXiv: 2510.21324 by Jinhui Lou, Jun Yu, Qingming Huang, Weidong Han, Yan Yang, Zhenqi Fu, Zhou Yu.

Figure 1
Figure 1. Figure 1: Comparison between prior agent models and our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the CXRAgent. The central director coordinates the following key stages: (1) ingesting user queries, (2) orchestrating CXR-analysis tools with outputs normalized and validated by the Evidence-driven Validator (EDV) to ensure visually grounded reliability, (3) making diagnostic plans to adaptively assemble specialized expert teams for collaborative reasoning, and (4) synthesizing evidence-backed… view at source ↗
Figure 3
Figure 3. Figure 3: Case study of CXRAgent versus state-of-the-art methods on report generation. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Case study of CXRAgent versus state-of-the-art methods on visual question answer. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Multi-Modal Multi-Agent Reinforcement Learning for Radiology Report Generation

    cs.CV 2026-02 unverdicted novelty 6.0

    MARL-Rad trains region-specific and global agents with reinforcement learning on clinical rewards to produce more accurate radiology reports than prior methods on MIMIC-CXR and IU X-ray datasets.

Reference graph

Works this paper leans on

40 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Interactive and explainable region-guided radiology report generation,

    T. Tanida, P. M ¨uller, G. Kaissis, and D. Rueckert, “Interactive and explainable region-guided radiology report generation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023, pp. 7433–7442

    2023

  2. [2]

    R2GenGPT: Radiology report generation with frozen llms,

    Z. Wang, L. Liu, L. Wang, and L. Zhou, “R2GenGPT: Radiology report generation with frozen llms,”Meta-Radiology, vol. 1, no. 3, p. 100033, 2023

    2023

  3. [3]

    MAIRA-2: grounded radiology report generation,

    S. Bannuret al., “MAIRA-2: grounded radiology report generation,” arXiv preprint arXiv:2406.04449, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    A clinically accessible small multimodal radiology model and evaluation metric for chest x-ray findings,

    Z. Chaveset al., “A clinically accessible small multimodal radiology model and evaluation metric for chest x-ray findings,”Nature Commu- nications, vol. 16, no. 1, p. 3108, 2025

    2025

  5. [5]

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

    C. Wu, X. Zhang, Y . Zhang, H. Hui, Y . Wang, and W. Xie, “Towards generalist foundation model for radiology by leveraging web-scale 2D&3D medical data,”Nature Communications, vol. 16, no. 1, p. 7866, 2025

    2025

  6. [6]

    CheXagent: Towards a foundation model for chest x-ray interpretation,

    Z. Chenet al., “CheXagent: Towards a foundation model for chest x-ray interpretation,” inAAAI 2024 Spring Symposium on Clinical Foundation Models, 2024

    2024

  7. [7]

    A fully open AI foundation model applied to chest radiography,

    D. Ma, J. Pang, M. B. Gotway, and J. Liang, “A fully open AI foundation model applied to chest radiography,”Nature, pp. 1–11, 2025

    2025

  8. [8]

    Medgemma technical report,

    A. Sellergrenet al., “Medgemma technical report,” 2025. [Online]. Available: https://arxiv.org/abs/2507.05201

    Pith/arXiv arXiv 2025

  9. [9]

    MedRAX: medical reasoning agent for chest x-ray,

    A. Fallahpour, J. Ma, A. Munim, H. Lyu, and B. Wang, “MedRAX: medical reasoning agent for chest x-ray,” inProceedings of the 42nd International Conference on Machine Learning (ICML), 2025

    2025

  10. [10]

    MDAgents: An adaptive collaboration of LLMs for medical decision-making,

    Y . Kimet al., “MDAgents: An adaptive collaboration of LLMs for medical decision-making,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  11. [11]

    MMedAgent: Learning to use medical tools with multi- modal agent,

    B. Liet al., “MMedAgent: Learning to use medical tools with multi- modal agent,” inFindings of the Association for Computational Linguis- tics: EMNLP 2024, Nov. 2024, pp. 8745–8760

    2024

  12. [12]

    Torchxrayvision: A library of chest x-ray datasets and models,

    J. P. Cohenet al., “Torchxrayvision: A library of chest x-ray datasets and models,” inInternational Conference on Medical Imaging with Deep Learning, 2022, pp. 231–249

    2022

  13. [13]

    A structure-aware relation network for thoracic diseases detection and segmentation,

    J. Lianet al., “A structure-aware relation network for thoracic diseases detection and segmentation,”IEEE Transactions on Medical Imaging, vol. 40, no. 8, pp. 2042–2052, 2021

    2042

  14. [14]

    A generative model uses healthy and diseased image pairs for pixel-level chest x-ray pathology localization,

    K. Dong, Y . Cheng, K. He, and J. Suo, “A generative model uses healthy and diseased image pairs for pixel-level chest x-ray pathology localization,”Nature Biomedical Engineering, 2025

    2025

  15. [15]

    Segment anything in medical images,

    J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,”Nature Communications, vol. 15, no. 1, p. 654, 2024

    2024

  16. [16]

    Consistency conditioned memory augmented dynamic diagnosis model for medical visual question answering,

    T. Yu, B. Ge, S. Wang, Y . Yang, Q. Huang, and J. Yu, “Consistency conditioned memory augmented dynamic diagnosis model for medical visual question answering,”IEEE Journal of Biomedical and Health Informatics, vol. 29, no. 2, pp. 1357–1370, 2025

    2025

  17. [17]

    Fine-grained adaptive visual prompt for generative medical visual question answering,

    T. Yu, Z. Tong, J. Yu, and K. Zhang, “Fine-grained adaptive visual prompt for generative medical visual question answering,” inthe AAAI Conference on Artificial Intelligence, 2025, pp. 9662–9670

    2025

  18. [18]

    Token-Mixer: Bind image and text in one embedding space for medical image reporting,

    Y . Yanget al., “Token-Mixer: Bind image and text in one embedding space for medical image reporting,”IEEE Transactions on Medical Imaging, vol. 43, no. 11, pp. 4017–4028, 2024

    2024

  19. [19]

    LLaV A-Med: Training a large language-and-vision assis- tant for biomedicine in one day,

    C. Liet al., “LLaV A-Med: Training a large language-and-vision assis- tant for biomedicine in one day,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 28 541–28 564

    2023

  20. [20]

    MedFILIP: medical fine-grained language-image pre- training,

    X. Lianget al., “MedFILIP: medical fine-grained language-image pre- training,”IEEE Journal of Biomedical and Health Informatics, 2025

    2025

  21. [21]

    Med-flamingo: a multimodal medical few-shot learner,

    M. Mooret al., “Med-flamingo: a multimodal medical few-shot learner,” inMachine Learning for Health (ML4H). PMLR, 2023, pp. 353–367

    2023

  22. [22]

    Lingshu: A generalist foundation model for unified multimodal medical understanding and reasoning,

    L. Teamet al., “Lingshu: A generalist foundation model for unified multimodal medical understanding and reasoning,” 2025. [Online]. Available: https://arxiv.org/abs/2506.07044

    Pith/arXiv arXiv 2025

  23. [23]

    A foundation model for chest x-ray interpretation with grounded reasoning via online reinforcement learning,

    Q. Linet al., “A foundation model for chest x-ray interpretation with grounded reasoning via online reinforcement learning,” 2025. [Online]. Available: https://arxiv.org/abs/2509.03906

    arXiv 2025

  24. [24]

    AgentMD: Empowering language agents for risk pre- diction with large-scale clinical tool learning,

    Q. Jinet al., “AgentMD: Empowering language agents for risk pre- diction with large-scale clinical tool learning,”Nature Communications, vol. 16, no. 1, p. 9377, 2025

    2025

  25. [25]

    CoD, towards an interpretable medical agent using chain of diagnosis,

    J. Chenet al., “CoD, towards an interpretable medical agent using chain of diagnosis,” inFindings of the Association for Computational Linguistics: ACL 2025, Vienna, Austria, Jul. 2025, pp. 14 345–14 368

    2025

  26. [26]

    Biomni: A general-purpose biomedical ai agent,

    K. Huanget al., “Biomni: A general-purpose biomedical ai agent,” bioRxiv, pp. 2025–05, 2025

    2025

  27. [27]

    MedAgents: Large language models as collaborators for zero-shot medical reasoning,

    X. Tanget al., “MedAgents: Large language models as collaborators for zero-shot medical reasoning,” inICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024

    2024

  28. [28]

    ClinicalAgent: clinical trial multi-agent system with large language model-based reasoning,

    L. Yue, S. Xing, J. Chen, and T. Fu, “ClinicalAgent: clinical trial multi-agent system with large language model-based reasoning,” inPro- ceedings of the 15th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics, 2024, pp. 1–10

    2024

  29. [29]

    DoctorAgent-RL: A multi- agent collaborative reinforcement learning system for multi-turn clinical dialogue,

    Y . Feng, J. Wang, L. Zhou, and Y . Li, “DoctorAgent-RL: A multi- agent collaborative reinforcement learning system for multi-turn clinical dialogue,”arXiv preprint arXiv:2505.19630, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    End-to-end agentic RAG system training for traceable diagnostic reasoning,

    Q. Zhenget al., “End-to-end agentic RAG system training for traceable diagnostic reasoning,”arXiv preprint arXiv:2508.15746, 2025

    arXiv 2025

  31. [31]

    Integration of multi-source medical data for medical diagnosis question answering,

    Q. Penget al., “Integration of multi-source medical data for medical diagnosis question answering,”IEEE Transactions on Medical Imaging, vol. 44, no. 3, pp. 1373–1385, 2025

    2025

  32. [32]

    MAM: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration,

    Y . Zhou, L. Song, and J. Shen, “MAM: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration,” in Findings of the Association for Computational Linguistics: ACL 2025, Vienna, Austria, Jul. 2025, pp. 25 319–25 333

    2025

  33. [33]

    CT-Agent: a multimodal- llm agent for 3D CT radiology question answering,

    Y . Mao, W. Xu, Y . Qin, and Y . Gao, “CT-Agent: a multimodal- llm agent for 3D CT radiology question answering,”arXiv preprint arXiv:2505.16229, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Patho-AgenticRAG: Towards multimodal agentic retrieval-augmented generation for pathology vlms via reinforcement learning,

    W. Zhanget al., “Patho-AgenticRAG: Towards multimodal agentic retrieval-augmented generation for pathology vlms via reinforcement learning,”arXiv preprint arXiv:2508.02258, 2025

    arXiv 2025

  35. [35]

    Prompt-agent-driven integration of foundation model priors for low-count pet reconstruction,

    X. Xieet al., “Prompt-agent-driven integration of foundation model priors for low-count pet reconstruction,”IEEE Transactions on Medical Imaging, pp. 1–1, 2025

    2025

  36. [36]

    ReAct: Synergizing reasoning and acting in language models,

    S. Yaoet al., “ReAct: Synergizing reasoning and acting in language models,” inInternational Conference on Learning Representations (ICLR), 2023

    2023

  37. [37]

    Interpretable medical image visual question answering via multi-modal relationship graph learning,

    X. Huet al., “Interpretable medical image visual question answering via multi-modal relationship graph learning,”Medical Image Analysis, vol. 97, p. 103279, 2024

    2024

  38. [38]

    MIMIC-CXR: A large publicly available database of labeled chest radiographs,

    A. E. W. Johnsonet al., “MIMIC-CXR: A large publicly available database of labeled chest radiographs,”Scientific Data, vol. 6, p. 317, 2019

    2019

  39. [39]

    MedVLM-R1: Incentivizing Medical Reasoning Capability of Vision-Language Models (VLMs) via Reinforcement Learning,

    J. Panet al., “MedVLM-R1: Incentivizing Medical Reasoning Capability of Vision-Language Models (VLMs) via Reinforcement Learning,” inproceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, vol. LNCS 15966, September 2025

    2025

  40. [40]

    RaTEScore: A metric for radiology report generation,

    W. Zhao, C. Wu, X. Zhang, Y . Zhang, Y . Wang, and W. Xie, “RaTEScore: A metric for radiology report generation,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 15 004–15 019

    2024