CoChat
Sign Up
Cited by 1Multimodal Machine Learning ApplicationsArtificial Intelligence in Healthcare and EducationRadiomics and Machine Learning in Medical Imaging

MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation

Open access
Lin, Yi, Ding, Yihao, Wu, Yonghui, Peng, Yifan·2026·Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)
DOI 10.18653/v1/2026.acl-short.24PMC PMC13352495
Open this paper in CoChat — free accountOpen PDF

Abstract

Automated 3D radiology report generation often suffers from clinical hallucinations and a lack of the iterative verification found in human practice.While recent Vision-Language Models (VLMs) have advanced the field, they typically operate as monolithic "black-box" systems without the collaborative oversight characteristic of clinical workflows.To address these challenges, we propose MARCH (Multi-Agent Radiology Clinical Hierarchy), a multi-agent framework that emulates the professional hierarchy of radiology departments and assigns specialized roles to distinct agents.MARCH utilizes a Resident Agent for initial drafting with multiscale CT feature extraction, multiple Fellow Agents for retrieval-augmented revision, and an Attending Agent that orchestrates an iterative, stance-based consensus discourse to resolve diagnostic discrepancies.On the RadGenome-ChestCT dataset, MARCH significantly outperforms state-of-the-art baselines in both clinical fidelity and linguistic accuracy.Our work demonstrates that modeling human-like organizational structures enhances the reliability of AI in high-stakes medical domains.

AI snapshot

Written from the paper’s full text.

TL;DR

MARCH is a multi-agent radiology report generation framework that emulates a Resident–Fellow–Attending clinical hierarchy for 3D chest CT interpretation. On RadGenome-ChestCT, it is reported to outperform state-of-the-art baselines on linguistic and clinical-validity metrics.

Key findings

  • MARCH consistently outperforms all baseline methods across evaluation metrics on RadGenome-ChestCT, according to the paper’s comparison.
  • Ablation shows each component contributes, with the most notable performance drop when removing consensus-driven Finalization.
  • Clinical efficacy F1 across 18 abnormalities is superior to the Resident Agent baseline for most abnormalities, particularly “hiatal hernia” and “pericardial effusion.”
  • Sensitivity analysis across LLM backbones shows all MARCH variants outperform baselines, with marginal improvements across advanced LLMs.

Methods

  • Design: three-phase multi-agent framework: Resident Agent drafts from CT with multi-scale/region features; Retrieval Agents and Fellow Agents revise using image-, text-, and logits-based retrieval; Attending Agent orchestrates iterative stance-based consensus.
  • Data: RadGenome-ChestCT, 25,692 chest CT scans from 21,304 patients; official split of 24,128 training and 1,564 testing scans.
  • Implementation/analysis: GPT-4.1 and GPT-4o as Fellow and Attending backbones at temperature 0; Resident/Retrieval trained on one NVIDIA H100 with AdamW, lr 1e-5, batch size 1, 10 epochs (~40 hours); evaluated with BLEU, ROUGE-L, METEOR, and Clinical Efficacy F1 over 18 abnormalities using RadBERT-RoBERTa-4m.

Limitations

  • Stated: evaluation primarily uses GPT-series LLMs, so generalizability to open-source or domain-specific medical LLMs remains untested.
  • Stated: MARCH lacks a long-term memory mechanism for longitudinal patient history or learning from past diagnostic errors.
  • Stated: it operates fully autonomously without a hybrid human-agent interface; future work should test human-in-the-loop configurations.
  • Evident from text: the provided text references Tables 1–3 and Figures 2–3 but does not include their numeric values, so effect sizes cannot be verified from the text alone; evaluation is also limited to a single dataset.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.18653/v1/2026.acl-short.24

Cites

0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 5 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    On RadGenome-ChestCT, MARCH performs significantly better than state-of-the-art baselines on clinical fidelity and linguistic accuracy.

    “On the RadGenome-ChestCT dataset, MARCH significantly outperforms state-of-the-art baselines in both clinical fidelity and linguistic accuracy.”
    • Also stated in: § Comparison with State-of-the-Art Methods. · boosted
      “Across all evaluation metrics, MARCH consistently outperforms all baseline methods, demonstrating superior performance in generating high-quality and clinically accurate radiology reports.”
    • Also stated in: § Conclusions · boosted
      “Empirical evaluations demonstrate that MARCH significantly outperforms state-of-the-art models on both linguistic metrics and clinical fidelity, generates reports that reduce the risk of single-reader misinterpretation, and supports transparent, collaborative report generation.”
    Quote in the paperSubstantial claim
  • Premisecitation

    LLM and VLM approaches have difficulty detecting subtle pathological findings in sparse 3D data.

    “these approaches often exhibit clinical hallucinations, struggle to detect subtle pathological findings in sparse 3D data, and lack the iterative verification and cross-checking (Zhu et al., 2025).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Zhu et al., 2025

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Interpretationown data

    MARCH produces reports that lower the chance of misinterpretation by a single reader.

    “generates reports that reduce the risk of single-reader misinterpretation”
    Quote in the paperSubstantial claim
  • Premisecitation

    The hierarchical verification-driven workflow used in radiology improves the diagnostic accuracy of the final report.

    “This hierarchical, verification-driven workflow improves diagnostic accuracy of the final report”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Hill et al., 2017 source
  • Premisecitation

    LLM and VLM approaches for radiology report generation frequently produce clinical hallucinations.

    “these approaches often exhibit clinical hallucinations, struggle to detect subtle pathological findings in sparse 3D data, and lack the iterative verification and cross-checking (Zhu et al., 2025).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Zhu et al., 2025

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims. 7
  • Classified as probably not a substantial claim. 1

  • May combine two claims.

    Findingown data

    All MARCH variants using different LLM backbones consistently beat baseline methods, with only small gains from more advanced LLMs.

    “Table 3 shows all variants of MARCH consistently outperform baseline methods, with marginal performance improvements observed across advanced LLMs.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Every component of MARCH makes a significant contribution to overall performance, and the largest decrease occurs when the consensus-driven finalization is removed.

    “Table 2 indicates that each component significantly contributes to the overall performance, with the most notable drop observed when removing the consensus-driven Finalization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    MARCH outperforms the Resident Agent baseline on most of the 18 clinical abnormalities, especially minor ones like hiatal hernia and pericardial effusion.

    “Compared with the baseline method, Resident Agent, MARCH demonstrates superior performance across most abnormalities, particularly in detecting minor abnormalities such as “hiatal hernia” and “pericardial effusion”, indicating its enhanced ability to identify subtle clinical findings.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    MARCH's retrieval-augmented revision stage improves diagnostic grounding, coherence, and clinical fidelity.

    “We introduce a retrieval-augmented revision stage that utilizes image-, text-, and logit-based indices to improve diagnostic grounding, coherence, and clinical fidelity.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    MARCH supports transparent and collaborative report generation.

    “supports transparent, collaborative report generation”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Most existing automated report generation systems use end-to-end black-box generation and do not model multi-agent review.

    “By contrast, most existing automated report generation systems do not model this multi-agent review process, and instead rely on end-to-end, black-box generation (Wang et al., 2023a).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Wang et al., 2023a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    MARCH's design improves interpretability and reliability, allows seamless integration of domain-specific knowledge, and sets a new benchmark in 3D report generation.

    “This design not only enhances interpretability and reliability but also enables seamless integration of domain-specific knowledge, thereby setting a new benchmark in 3D report generation.”
    Quote in the paperSubstantial claim
7 of 7

  • The paper relies on this without support.

    The use of multi-agent systems for 3D radiology report generation is still underexplored.

    “Despite these advances, the application of multi-agent systems for 3D radiology report generation remains underexplored.”

  • The paper relies on this without support.

    In radiology, a Resident usually makes the initial interpretation and drafts a report, a Fellow independently reviews and may reinterpret it, and an Attending radiologist adjudicates and finalizes the report when discrepancies remain.

    “Typically, a Resident performs the initial interpretation and drafts a report, which is then independently reviewed and, if necessary, reinterpreted by a Fellow. When discrepancies persist, an Attending radiologist conducts a parallel read to adjudicate and finalize the report.”
2 of 2

Cites

The publisher hasn’t deposited this paper’s reference list.

Research works printed but not linked: 23

  1. Bai Fan, Du Yuxin, Huang Tiejun, Meng Max Q-H, and Zhao Bo. 2024. M3D: Advancing 3D medical image analysis with multi-modal large language models. arXiv [cs.CV].
  2. Banerjee Satanjeev and Lavie Alon. 2005. METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65– 72.
  3. Chen Kai, Qi Ji, Huo Jing, Tian Pinzhuo, Meng Fanyu, Yang Xi, and Gao Yang. 2025a. A self-evolving framework for multi-agent medical consultation based on large language models. In ICASSP 2025–2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1– 5. IEEE.
  4. Dang Yufan, Qian Chen, Luo Xueheng, Fan Jingru, Xie Zihao, Shi Ruijie, Chen Weize, Yang Cheng, Che Xiaoyin, Tian Ye, Xiong Xuantang, Han Lei, Liu Zhiyuan, and Sun Maosong. 2025. Multi-agent collaboration via evolving orchestration. In NeurIPS.
  5. Du Zhuoyun, Qian Chen, Liu Wei, Xie Zihao, Wang Yifei, Qiu Rennai, Dang Yufan, Chen Weize, Yang Cheng, Tian Ye, Xiong Xuantang, and Han Lei. 2025. Multi-agent collaboration via cross-team orchestration. In Findings of the Association for Computational Linguistics: ACL 2025, pages 10386– 10406, Stroudsburg, PA, USA. Association for Computational Linguistics.
  6. Fan Zhihao, Wei Lai, Tang Jialong, Chen Wei, Siyuan Wang, Wei Zhongyu, and Huang Fei. 2025. AI hospital: Benchmarking large language models in a multi-agent medical interaction simulator. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10183– 10213.
  7. Hamamci Ibrahim Ethem, Er Sezgin, and Menze Bjoern. 2024. CT2Rep: Automated radiology report generation for 3D medical imaging. In Medical Image Computing and Computer Assisted Intervention, Lecture notes in computer science, pages 476– 486. Springer Nature Switzerland, Cham.
  8. Hu Edward J, Wallis Phillip, Allen-Zhu Zeyuan, Li Yuanzhi, Wang Shean, Wang Lu, Chen Weizhu, and Others. 2021. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, pages 1– 26.
  9. Jiang Yue, Chen Jiawei, Yang Dingkang, Li Mingcheng, Wang Shunli, Wu Tong, Li Ke, and Zhang Lihua. 2025. Comt: Chain-of-medical-thought reduces hallucination in medical report generation. In ICASSP 2025–2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1– 5. IEEE.
  10. Johnson Alistair EW, Pollard Tom J, Greenbaum Nathaniel R, Lungren Matthew P, Deng Chih-ying, Peng Yifan, Lu Zhiyong, Mark Roger G, Berkowitz Seth J, and Horng Steven. 2019. MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042. PMC6908718 31831740 · arXiv
  11. Kim Yubin, Park Chanwoo, Jeong Hyewon, Chan Yik S, Xu Xuhai, McDuff Daniel, Lee Hyeonhoon, Ghassemi Marzyeh, Breazeal Cynthia, and Park Hae W. 2024. Mdagents: An adaptive collaboration of llms for medical decision-making. Advances in Neural Information Processing Systems, 37: 79410– 79452.
  12. Li Binxu, Yan Tiankai, Pan Yuanting, Luo Jie, Ji Ruiyang, Ding Jiayuan, Xu Zhe, Liu Shilong, Dong Haoyu, Lin Zihao, and 1 others. 2024. MMedAgent: Learning to use medical tools with multi-modal agent. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 8745– 8760.
  13. Liao Callie C, Liao Duoduo, and Gadiraju Sai Surya. 2025. AgentMaster: A multi-agent conversational framework using A2A and MCP protocols for multi-modal information retrieval and analysis. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 52– 72, Stroudsburg, PA, USA. Association for Computational Linguistics.
  14. Lin Chin-Yew. 2004. ROUGE: a package for automatic evaluation of summaries. In Text summarization branches out: Proceedings of the ACL-04 workshop, volume 8, pages 1– 8.
  15. Liu Chang, Tian Yuanhe, Chen Weidong, Song Yan, and Zhang Yongdong. 2024a. Bootstrapping large language models for radiology report generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18635– 18643.
  16. Liu Che, Wan Zhongwei, Wang Yuqi, Shen Hui, Wang Haozhe, Zheng Kangyu, Zhang Mi, and Arcucci Rossella. 2025. Argus: benchmarking and enhancing vision-language models for 3d radiology report generation. In Findings of the Association for Computational Linguistics: ACL 2025, pages 16448– 16460.
  17. Liu Rui, Li Mingjie, Zhao Shen, Chen Ling, Chang Xiaojun, and Yao Lina. 2024b. In-context learning for zero-shot medical report generation. In Proceedings of the 32nd ACM international conference on multimedia, pages 8721– 8730.
  18. Papineni Kishore, Roukos Salim, Ward Todd, and Zhu WeiJing. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02, pages 311– 318, USA. Association for Computational Linguistics.
  19. Touvron Hugo, Martin Louis, Stone Kevin, Albert Peter, Almahairi Amjad, Babaei Yasmine, Bashlykov Nikolay, Batra Soumya, Bhargava Prajjwal, Bhosale Shruti, Bikel Dan, Blecher Lukas, Ferrer Cristian Canton, Chen Moya, Cucurull Guillem, Esiobu David, Fernandes Jude, Fu Jeremy, Fu Wenyin, and 49 others. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv [cs.CL].
  20. Wang Zhanyu, Liu Lingqiao, Wang Lei, and Zhou Luping. 2023a. Metransformer: Radiology report generation by transformer with multiple learnable expert tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11558– 11567.
  21. Wang Zhanyu, Liu Lingqiao, Wang Lei, and Zhou Luping. 2023b. R2GenGPT: Radiology report generation with frozen LLMs. Meta-Radiology, 1(3): 100033.
  22. Wolf Thomas, Debut Lysandre, Sanh Victor, Chaumond Julien, Delangue Clement, Moi Anthony, Cistac Pierric, Rault Tim, Louf Remi, Funtowicz Morgan, and 1 others. 2020. Transformers: 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.
  23. Zhu Zhihong, Zhang Yunyan, Zhuang Xianwei, Zhang Fan, Wan Zhongwei, Chen Yuyan, QingqingLong QingqingLong, Zheng Yefeng, and Wu Xian. 2025. Can we trust ai doctors? a survey of medical hallucination in large language and large vision-language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 6748– 6769.

Cited by

Sorted by citations

Authors
Lin, Y., Ding, Y., Wu, Y., & Peng, Y.
Affiliations
  • Weill Cornell Medicine (US)
  • The University of Western Australia (AU)
  • University of Florida (US)
ORCID
Peng, Yifan
Venue
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)
Publisher
Association for Computational Linguistics
Published
2026
Volume
2026
Issue
v2 / short
Pages
273-285
DOI
10.18653/v1/2026.acl-short.24
PMC
PMC13352495
OpenAlex
W7166877470
Cited by
1
References
0
Citations by year
2026: 1
FWCI
10.34
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • Annual Meeting of the Association for Computational Linguistics (ACL) (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › Radiomics and Machine Learning in Medical Imaging
Keywords
automated radiology report generation, multi-agent systems, vision-language models, retrieval-augmented generation, chest CT, clinical hierarchy

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review