pith. sign in

arxiv: 2606.13211 · v1 · pith:BV4YLVL6new · submitted 2026-06-11 · 💻 cs.AI

Hallucination in Medical Imaging AI: A Cross-Modality Analytical Framework for Taxonomy, Detection, and Mitigation under Regulatory Constraints

Omar Alshahrani , Muzammil Behzad This is my paper

Pith reviewed 2026-06-27 06:59 UTC · model grok-4.3

classification 💻 cs.AI
keywords hallucinationmedical imagingfoundation modelsAI regulationFDAtaxonomymitigationcross-modality
0
0 comments X

The pith

General-purpose foundation models outperform medical-specialized models on hallucination benchmarks because narrow fine-tuning introduces overfitting-induced confabulation.

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

The paper synthesizes peer-reviewed studies, benchmarks, and FDA guidance across five imaging modalities to unify hallucination taxonomies, compare model types, and map mitigations to regulatory requirements. It establishes that general-purpose foundation models produce fewer fabricated anatomical structures, missed findings, and incorrect measurements than models specialized through narrow domain fine-tuning. A sympathetic reader would care because such errors directly affect biopsy decisions, staging, and treatment planning. The review also shows that radiologist oversight is still required for a high percentage of AI outputs and that combining physics-informed constraints, Chain-of-Thought prompting, and human-in-the-loop checks addresses distinct failure modes.

Core claim

Three taxonomic frameworks together cover the imaging pipeline in a way no single framework does alone. General-purpose foundation models outperform medical-specialized models on hallucination-specific benchmarks, indicating that narrow domain fine-tuning can introduce overfitting-induced confabulation. Physics-informed architectural constraints, Chain-of-Thought prompting, and human-in-the-loop safeguards each address different failure modes and remain effective when combined. All findings map to the FDA's Total Product Lifecycle and Predetermined Change Control Plan frameworks, which treat hallucination management as a lifecycle obligation rather than a pre-deployment checklist.

What carries the argument

A cross-modality analytical framework that unifies three existing taxonomic approaches to analyze hallucination taxonomy, etiology, detection, and mitigation across the full imaging pipeline.

If this is right

  • Unified taxonomies from the three frameworks together cover the entire imaging pipeline where no single framework suffices.
  • General-purpose models should be considered as base architectures to avoid the confabulation introduced by narrow fine-tuning.
  • Mitigation strategies must be combined because each targets distinct failure modes such as anatomical fabrication or incorrect laterality.
  • Hallucination management is a continuous obligation under FDA lifecycle frameworks rather than a one-time pre-deployment check.
  • Human oversight remains essential, as a high percentage of AI-generated flags require expert correction before clinical use.

Where Pith is reading between the lines

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

  • The finding on fine-tuning may extend to other high-stakes domains where broad pretraining could reduce domain-specific confabulation better than narrow specialization.
  • Regulators may need to require disclosure of training scale and protocol details to prevent confounded comparisons in future model approvals.
  • Future benchmarks could isolate the effect of fine-tuning scale by holding data volume constant while varying domain specificity.
  • The reliance on human-in-the-loop suggests that regulatory pathways may need to define required levels of radiologist review rather than aiming for full autonomy.

Load-bearing premise

The benchmarks and studies synthesized are representative across modalities and comparisons between general-purpose and medical-specialized models are not confounded by differences in training data scale or evaluation protocols.

What would settle it

A controlled study that equalizes training data volume and evaluation protocols across model types and still finds medical-specialized models producing fewer hallucinations than general-purpose ones would falsify the central performance claim.

Figures

Figures reproduced from arXiv: 2606.13211 by Muzammil Behzad, Omar Alshahrani.

Figure 1
Figure 1. Figure 1: FIGURE 1 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIGURE 2 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIGURE 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIGURE 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIGURE 6 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIGURE 5 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIGURE 7 [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIGURE 8 [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIGURE 11 [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
read the original abstract

AI systems are being deployed across medical imaging faster than their failure modes are understood. At this point in time, the failure of greatest clinical concern is hallucination: clinically plausible but factually incorrect outputs, including fabricated anatomical structures, missed findings, incorrect laterality, and invented measurements in generated reports, with direct consequences, for example, for biopsy decisions, staging, and treatment planning. This structured narrative synthesizes peer-reviewed studies, benchmark datasets, and FDA regulatory guidance across five imaging modalities to produce a cross-modality analysis of hallucination taxonomy, etiology, detection, and mitigation. Specifically, we address three questions in this study: (1) how can existing taxonomies be unified across modalities?, (2) how do medical-specialized foundation models hallucinate less than general-purpose ones?, and (3) which mitigation strategies are effective and compatible with FDA lifecycle oversight? We note that three taxonomic frameworks together cover the imaging pipeline in a way no single framework does alone. We also highlight that general-purpose foundation models outperform medical-specialized models on hallucination-specific benchmarks, indicating that narrow domain fine-tuning can introduce overfitting-induced confabulation. At the same time, the oversight of radiologists remains essential; for instance, a very high percentage of of AI-generated flags required expert correction before clinical use. Physics-informed architectural constraints, Chain-of-Thought prompting, and human-in-the-loop safeguards each address different failure modes and is effective when combined. All findings are mapped to the FDA's Total Product Lifecycle and Predetermined Change Control Plan frameworks, which treat hallucination management as a lifecycle obligation rather than a pre-deployment checklist.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. This manuscript is a structured narrative synthesis of peer-reviewed studies, benchmarks, and FDA guidance on hallucination (clinically plausible but factually incorrect outputs) in medical imaging AI across five modalities. It unifies three taxonomic frameworks to cover the imaging pipeline, addresses three questions on taxonomy unification, model performance differences, and mitigation strategies compatible with regulatory oversight, claims that general-purpose foundation models outperform medical-specialized ones on hallucination benchmarks (suggesting overfitting from narrow fine-tuning), notes the necessity of radiologist oversight, and maps findings to FDA Total Product Lifecycle and Predetermined Change Control Plan frameworks.

Significance. If the synthesized comparisons and mappings hold after addressing confounders, the work would offer a useful cross-modality taxonomy unification and regulatory-aligned mitigation overview for a high-stakes clinical failure mode, potentially informing safer AI deployment in imaging.

major comments (2)
  1. [Abstract] Abstract, research question (2): The question presupposes that 'medical-specialized foundation models hallucinate less than general-purpose ones,' yet the highlighted finding states the opposite (general-purpose models outperform specialized ones, implying specialized models hallucinate more due to overfitting). This internal contradiction between the framing question and the central claim requires explicit resolution.
  2. [Abstract] Abstract, paragraph on model comparison: The claim that general-purpose foundation models outperform medical-specialized models on hallucination-specific benchmarks (with attribution to overfitting-induced confabulation) is stated without any quantitative synthesis, tables of results, error bars, or discussion of controls for confounders such as model scale, training data volume, or non-identical evaluation protocols. As a narrative synthesis, the manuscript does not demonstrate that the reviewed studies isolate domain specialization as the causal factor.
minor comments (2)
  1. [Abstract] Abstract: 'a very high percentage of of AI-generated flags' contains a duplicated word.
  2. [Abstract] Abstract, final sentence: 'Physics-informed architectural constraints, Chain-of-Thought prompting, and human-in-the-loop safeguards each address different failure modes and is effective when combined' has a subject-verb agreement error ('is' should be 'are').

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review. The comments highlight an important inconsistency in the abstract phrasing and raise valid points about the evidentiary basis for the model comparison claim in a narrative synthesis. We address each below and commit to targeted revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract, research question (2): The question presupposes that 'medical-specialized foundation models hallucinate less than general-purpose ones,' yet the highlighted finding states the opposite (general-purpose models outperform specialized ones, implying specialized models hallucinate more due to overfitting). This internal contradiction between the framing question and the central claim requires explicit resolution.

    Authors: We agree that the current wording of research question (2) creates an unintended presupposition that does not match the manuscript's central finding. The question was drafted to explore comparative hallucination behavior but was phrased in a way that assumes the direction of the effect. We will revise it to: '(2) how do hallucination rates in medical-specialized foundation models compare to those in general-purpose ones?' This removes the presupposition and aligns the framing directly with the reported observation that general-purpose models outperform specialized ones. The revised question will be reflected consistently in the abstract and introduction. revision: yes

  2. Referee: [Abstract] Abstract, paragraph on model comparison: The claim that general-purpose foundation models outperform medical-specialized models on hallucination-specific benchmarks (with attribution to overfitting-induced confabulation) is stated without any quantitative synthesis, tables of results, error bars, or discussion of controls for confounders such as model scale, training data volume, or non-identical evaluation protocols. As a narrative synthesis, the manuscript does not demonstrate that the reviewed studies isolate domain specialization as the causal factor.

    Authors: As a structured narrative synthesis, the manuscript aggregates and interprets patterns reported in the existing peer-reviewed literature rather than conducting a new quantitative meta-analysis with statistical controls. We therefore cannot claim that the reviewed studies isolate domain specialization as the sole causal factor, and we acknowledge that confounders such as model scale, training data volume, and heterogeneous evaluation protocols limit causal inference. To address this, we will add an explicit limitations subsection discussing these confounders and the interpretive boundaries of narrative synthesis. We will also insert a summary table collating the key studies, their reported hallucination metrics, and noted methodological differences to increase transparency. These additions strengthen the presentation without altering the manuscript's primary scope as a cross-modality taxonomy and regulatory mapping exercise. revision: partial

Circularity Check

0 steps flagged

No circularity: synthesis rests on external peer-reviewed studies and FDA guidance

full rationale

The paper is a structured narrative synthesis referencing external peer-reviewed studies, benchmark datasets, and FDA regulatory guidance across modalities. The central claim that general-purpose foundation models outperform medical-specialized models on hallucination benchmarks is attributed to synthesized external sources rather than any derivation, equation, or fitted parameter internal to the paper. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the provided text. The work is self-contained against external benchmarks and does not reduce its conclusions to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a literature synthesis paper. No mathematical derivations, new empirical fits, or postulated entities are introduced in the abstract.

pith-pipeline@v0.9.1-grok · 5836 in / 1083 out tokens · 16463 ms · 2026-06-27T06:59:05.049349+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

45 extracted references · 8 canonical work pages

  1. [1]

    On hallucinations in AI-generated content for nuclear medicine imaging (the DREAM Report),

    M. Xia, R. Bayerlein, Y . Chemli et al., “On hallucinations in AI-generated content for nuclear medicine imaging (the DREAM Report),”J. Nucl. Med., published online Nov. 6, 2025, PMCID: PMC11844622

    2025

  2. [2]

    Medical hallucination in foundation models and their impact on healthcare

    Y . Kim, H. Jeong, S. Chen et al., “Medical hallucination in foundation models and their impact on healthcare,”medRxiv2025.02.28.25323115 (also arXiv 2503.05777). DOI: 10.1101/2025.02.28.25323115

    work page doi:10.1101/2025.02.28.25323115 2025

  3. [3]

    A taxonomy of machine hallucina- tion in radiology,

    F. J. Brooks and M. A. Anastasio, “A taxonomy of machine hallucina- tion in radiology,”Radiology: Artif. Intell., DOI: 10.1148/ryai.250203

    work page doi:10.1148/ryai.250203

  4. [4]

    Detecting and evaluating medical hallucinations in large vision language models,

    J. Chen, D. Yang, T. Wu et al., “Detecting and evaluating medical hallucinations in large vision language models,” arXiv 2406.10185, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    J., Madotto, A., and Fung, P

    Z. Ji, N. Lee, R. Frieske et al., “Survey of hallucination in natural language generation,”ACM Comput. Surv., vol. 55, no. 12, Art. no. 248, 2023. DOI: 10.1145/3571730

    work page doi:10.1145/3571730 2023

  6. [6]

    On hallucinations in tomographic image reconstruction,

    S. Bhadra, V . A. Kelkar, F. J. Brooks, and M. A. Anastasio, “On hallucinations in tomographic image reconstruction,”IEEE Trans. Med. Imag., vol. 40, no. 11, pp. 3249–3260, Nov. 2021. DOI: 10.1109/TMI.2021.3089456

    work page doi:10.1109/tmi.2021.3089456 2021

  7. [7]

    A survey on hallucination in large language models: principles, taxonomy, challenges, and open questions,

    L. Huang, W. Yu, W. Ma et al., “A survey on hallucination in large language models: principles, taxonomy, challenges, and open questions,”ACM Trans. Inf. Syst., 2024, arXiv 2311.05232

    Pith/arXiv arXiv 2024

  8. [8]

    Deep learning-based algorithm for detection and characterization of COVID-19 pneumonia in chest X-rays and CT images,

    Y . Sim, M. J. Chung, E. Kotter et al., “Deep learning-based algorithm for detection and characterization of COVID-19 pneumonia in chest X-rays and CT images,”JAMA Netw. Open, vol. 4, no. 12, Art. no. e2141096, 2021

    2021

  9. [9]

    Radiologist diagnostic sensitivity with versus without AI assistance for chest radiograph findings,

    J. S. Ahn, S. Ebrahimian, S. McDermott et al., “Radiologist diagnostic sensitivity with versus without AI assistance for chest radiograph findings,”JAMA Netw. Open, vol. 5, no. 8, Art. no. e2229289, 2022

    2022

  10. [10]

    AI double-reading triage across 25,104 consecutive chest radiographs,

    L. Topff, R. van der Sluijs, H. Laue et al., “AI double-reading triage across 25,104 consecutive chest radiographs,”Eur. Radiol., vol. 34, no. 9, pp. 5876–5885, 2024, PMC11364654

    2024

  11. [11]

    AI improves nodule detection on chest radiographs in a health screening population: a randomized controlled trial,

    J. G. Nam, E. J. Hwang, J. Kim et al., “AI improves nodule detection on chest radiographs in a health screening population: a randomized controlled trial,”Radiology, vol. 307, no. 2, Art. no. e221894, 2023. DOI: 10.1148/radiol.221894

    work page doi:10.1148/radiol.221894 2023

  12. [12]

    A physics-informed deep learning model for MRI brain motion correction,

    M. Safari, X. Yang, Z. Eidex et al., “A physics-informed deep learning model for MRI brain motion correction,” PMCID PMC11844622, arXiv 2502.09296, 2025

    arXiv 2025

  13. [13]

    Hallucination of multimodal large language models: a survey,

    Z. Bai, P. Wang, T. Xiao et al., “Hallucination of multimodal large language models: a survey,” arXiv 2404.18930, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Diagnostic accuracy and failure mode analysis of a deep learning algorithm for the detection of cervical spine fractures,

    A. F. V oter, M. E. Larson, J. W. Garrett, and J. P. Yu, “Diagnostic accuracy and failure mode analysis of a deep learning algorithm for the detection of cervical spine fractures,”Amer. J. Neuroradiol., vol. 42, no. 7, pp. 1298–1305, 2021. DOI: 10.3174/ajnr.A7179

    work page doi:10.3174/ajnr.a7179 2021

  15. [15]

    Do as AI say: susceptibility in deployment of clinical decision-aids,

    S. Gaube, H. Suresh, M. Raue et al., “Do as AI say: susceptibility in deployment of clinical decision-aids,”NPJ Digit. Med., vol. 4, Art. no. 31, 2021. DOI: 10.1038/s41746-021-00385-9

    work page doi:10.1038/s41746-021-00385-9 2021

  16. [16]

    FDA perspective on the regulation of artificial intelligence in health care and biomedicine,

    H. J. Warraich, T. Tazbaz, and R. M. Califf, “FDA perspective on the regulation of artificial intelligence in health care and biomedicine,” JAMA, 2024. DOI: 10.1001/jama.2024.21451

    work page doi:10.1001/jama.2024.21451 2024

  17. [17]

    Total product lifecycle con- siderations for generative AI-enabled devices,

    U.S. Food and Drug Administration, “Total product lifecycle con- siderations for generative AI-enabled devices,” Executive Summary for the Digital Health Advisory Committee, Nov. 20, 2024. [Online]. Available: https://www.fda.gov/media/184273/download

    2024

  18. [18]

    Artificial intelligence-enabled device software functions: lifecycle management and marketing sub- mission recommendations,

    U.S. Food and Drug Administration, “Artificial intelligence-enabled device software functions: lifecycle management and marketing sub- mission recommendations,” Draft Guidance, Jan. 7, 2025. Docket FDA-2025-D-0070. VOLUME , 11 Omar Alshahrani and Muzammil Behzad: Hallucination in Medical Imaging AI

    2025

  19. [19]

    Predetermined change control plans (PCCP) for machine learning-enabled devices: final guidance,

    U.S. Food and Drug Administration, “Predetermined change control plans (PCCP) for machine learning-enabled devices: final guidance,” Dec. 2024

    2024

  20. [20]

    A survey on deep learning in medical image analysis,

    G. Litjens, T. Kooi, B. E. Bejnordi et al., “A survey on deep learning in medical image analysis,”Med. Image Anal., vol. 42, pp. 60–88, 2017

    2017

  21. [21]

    AI in health and medicine,

    P. Rajpurkar, E. Chen, O. Banerjee, and E. J. Topol, “AI in health and medicine,”Nat. Med., vol. 28, pp. 31–38, 2022

    2022

  22. [22]

    Predicting the future—big data, machine learning, and clinical medicine,

    Z. Obermeyer and E. J. Emanuel, “Predicting the future—big data, machine learning, and clinical medicine,”N. Engl. J. Med., vol. 375, pp. 1216–1219, 2016

    2016

  23. [23]

    High-performance medicine: the convergence of human and artificial intelligence,

    E. J. Topol, “High-performance medicine: the convergence of human and artificial intelligence,”Nat. Med., vol. 25, pp. 44–56, 2019

    2019

  24. [24]

    On the interpretability of artificial intelligence in radiology: challenges and opportunities,

    M. Reyes, R. Meier, S. Pereira et al., “On the interpretability of artificial intelligence in radiology: challenges and opportunities,”Ra- diology: Artif. Intell., vol. 2, Art. no. e190043, 2020

    2020

  25. [25]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,

    C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,”Nat. Mach. Intell., vol. 1, pp. 206–215, 2019

    2019

  26. [26]

    Implementing machine learning in health care—addressing ethical challenges,

    D. S. Char, N. H. Shah, and D. Magnus, “Implementing machine learning in health care—addressing ethical challenges,”N. Engl. J. Med., vol. 378, pp. 981–983, 2018

    2018

  27. [27]

    Key challenges for delivering clinical impact with artificial intelligence,

    C. J. Kelly, A. Karthikesalingam, M. Suleyman et al., “Key challenges for delivering clinical impact with artificial intelligence,”BMC Med., vol. 17, Art. no. 195, 2019

    2019

  28. [28]

    Deep learning in medical image analysis,

    D. Shen, G. Wu, and H.-I. Suk, “Deep learning in medical image analysis,”Annu. Rev. Biomed. Eng., vol. 19, pp. 221–248, 2017

    2017

  29. [29]

    AI in medical imaging informatics: current challenges and future directions,

    A. S. Panayides, A. Amini, N. D. Filipovic et al., “AI in medical imaging informatics: current challenges and future directions,”IEEE J. Biomed. Health Informat., vol. 24, no. 7, pp. 1837–1857, 2020

    2020

  30. [30]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar et al., “Attention is all you need,” inAdv. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  31. [31]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. IEEE Conf. Comput. Vision Pattern Recognit. (CVPR), 2016

    2016

  32. [32]

    Deep learning,

    Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,”Nature, vol. 521, pp. 436–444, 2015

    2015

  33. [33]

    Generative adver- sarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza et al., “Generative adver- sarial networks,”Commun. ACM, vol. 63, pp. 139–144, 2020

    2020

  34. [34]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv 1409.1556, 2014

    Pith/arXiv arXiv 2014

  35. [35]

    U-Net: convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: convolutional net- works for biomedical image segmentation,” inProc. MICCAI, 2015

    2015

  36. [36]

    An image is worth 16×16words: transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov et al., “An image is worth 16×16words: transformers for image recognition at scale,” arXiv 2010.11929, 2020

    Pith/arXiv arXiv 2010

  37. [37]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder et al., “Language models are few-shot learners,” inAdv. Neural Inf. Process. Syst., vol. 33, pp. 1877–1901, 2020

    1901

  38. [38]

    Learning transferable visual models from natural language supervision (CLIP),

    A. Radford, J. W. Kim, C. Hallacy et al., “Learning transferable visual models from natural language supervision (CLIP),” inProc. ICML, 2021

    2021

  39. [39]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson et al., “On the opportunities and risks of foundation models,” arXiv 2108.07258, 2021

    Pith/arXiv arXiv 2021

  40. [40]

    Review of deep learning algorithms and architectures,

    A. Shrestha and A. Mahmood, “Review of deep learning algorithms and architectures,”IEEE Access, vol. 7, pp. 53040–53065, 2019

    2019

  41. [41]

    A guide to deep learning in healthcare,

    A. Esteva, A. Robicquet, B. Ramsundar et al., “A guide to deep learning in healthcare,”Nat. Med., vol. 25, pp. 24–29, 2019

    2019

  42. [42]

    MRI motion artifact detection and correction using AI: systematic review and meta-analysis,

    K. Pawar, J. Fripp, and J. Dowling, “MRI motion artifact detection and correction using AI: systematic review and meta-analysis,” PMC, 2025

    2025

  43. [43]

    Uncertainty quantification for machine learning in healthcare: a survey,

    M. Abdar, F. Pourpanah, S. Hussain et al., “Uncertainty quantification for machine learning in healthcare: a survey,”Neurocomputing, vol. 461, pp. 243–268, 2021

    2021

  44. [44]

    Fusion of medical imaging and electronic health records using deep learning: systematic review and implementation guidelines,

    S.-C. Huang, A. Pareek, S. Seyyedi, I. Banerjee, and M. P. Lungren, “Fusion of medical imaging and electronic health records using deep learning: systematic review and implementation guidelines,”NPJ Digit. Med., vol. 3, Art. no. 136, 2020

    2020

  45. [45]

    MedHallBench: a benchmark for hallucination detection in med- ical VLMs with reinforcement learning-assisted annotation,

    “MedHallBench: a benchmark for hallucination detection in med- ical VLMs with reinforcement learning-assisted annotation,” arXiv 2412.18947, 2024. 12 VOLUME ,

    arXiv 2024