CoChat
Sign Up
Cited by 168Artificial Intelligence in Healthcare and EducationEthics and Social Impacts of AIExplainable Artificial Intelligence (XAI)

Bias in artificial intelligence for medical imaging: fundamentals, detection, avoidance, mitigation, challenges, ethics, and prospects

Open access
Koçak, Burak, Ponsiglione, Andrea, Stanzione, Arnaldo, Bluethgen, Christian, Santinha, João, Ugga, Lorenzo, Huisman, Merel, Klontzas, Michail E. +2 more·2024·Diagnostic and Interventional Radiology
DOI 10.4274/dir.2024.242854PMC PMC11880872
Open this paper in CoChat — free account

Abstract

Although artificial intelligence (AI) methods hold promise for medical imaging-based prediction tasks, their integration into medical practice may present a double-edged sword due to bias (i.e., systematic errors). AI algorithms have the potential to mitigate cognitive biases in human interpretation, but extensive research has highlighted the tendency of AI systems to internalize biases within their model. This fact, whether intentional or not, may ultimately lead to unintentional consequences in the clinical setting, potentially compromising patient outcomes. This concern is particularly important in medical imaging, where AI has been more progressively and widely embraced than any other medical field. A comprehensive understanding of bias at each stage of the AI pipeline is therefore essential to contribute to developing AI solutions that are not only less biased but also widely applicable. This international collaborative review effort aims to increase awareness within the medical imaging community about the importance of proactively identifying and addressing AI bias to prevent its negative consequences from being realized later. The authors began with the fundamentals of bias by explaining its different definitions and delineating various potential sources. Strategies for detecting and identifying bias were then outlined, followed by a review of techniques for its avoidance and mitigation. Moreover, ethical dimensions, challenges encountered, and prospects were discussed.

AI snapshot

Written from the paper’s full text.

TL;DR

This international collaborative review explains bias in medical imaging artificial intelligence across the full pipeline, from definitions and sources to detection, avoidance, mitigation, ethics, challenges, and prospects. It argues that proactive identification and mitigation are needed, while acknowledging that completely bias-free systems are unrealistic.

Key findings

  • Bias is framed as systematic error or distance between prediction and truth, with sources spanning study design/dataset, modeling, and deployment, including demographic imbalance, covariate shift, annotation/reference standard bias, data leakage, deployment bias, concept drift, automation bias, and algorithmic aversion.
  • Detection strategies include dataset evaluation, unsupervised analysis, subgroup statistical comparison, visualization, independent code review, testing on unseen populations, and explainability analysis; named tools include PROBAST with four bias categories, QUADAS-2 with 14 questions, the under-development PROBAST-AI, TRIPOD-AI, FUTURE-AI, CLEAR, and METRICS.
  • Mitigation is divided into preprocessing, during-training, and post-processing approaches; explainable AI methods include perturbation-based methods such as occlusion, LIME, and SHAP, and backpropagation-based methods such as saliency maps and CAM/Grad-CAM.
  • Ethical and regulatory discussion cites the WHO’s six key ethical principles for AI in healthcare and the European Union’s AI Act, including Article 10 data governance for high-risk systems and Article 61 post-market surveillance, with full effect in 2026; challenges include limited benchmark diversity and the finding that fewer than approximately 6% of published radiology AI articles partially or completely share experimental data.

Methods

  • Design: international collaborative narrative review covering fundamentals, detection, avoidance, mitigation, challenges, ethics, and prospects, explicitly without detailed technical discussion.
  • Data or sample: no original dataset or systematic search is described; the paper synthesizes prior literature and references, with figures, tables, and collective recommendations in Table 3.
  • Main analysis: conceptual and descriptive synthesis; no quantitative meta-analysis or primary statistical analysis is reported.

Limitations

  • Stated: the review does not provide a detailed technical discussion; the authors state that expecting completely bias-free systems may be unrealistic and that complete inclusivity of all data types and sources remains unattainable.
  • Stated challenges: ambiguities in interpreting results, limited diversity in benchmark datasets, subjectivity in detecting bias, and constraints related to data sharing and patient privacy.
  • Evident: as a narrative review, no systematic search strategy, inclusion criteria, or quantitative synthesis is reported in the text; recommendations are expert-based rather than empirically tested here.

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

Integrity signals

Registry

Crossref · 10.4274/dir.2024.242854

Cites

144 of 145 references verified1 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 74 cited sources checked · 0 supported · 0 flagged

  • Premisenone given

    The choice of metrics used to estimate model performance could introduce bias if those selected do not match the information needed.

    “Another aspect to carefully consider is the choice of metrics used to estimate the model’s performance, which could introduce bias if those selected do not match the information needed.”
    Quote in the paperNot a substantial claim
  • Premisecitation

    The choice of data preprocessing techniques such as normalization, augmentation, or cropping can influence the model's output by emphasizing certain features over others.

    “Additionally, the choice of data preprocessing techniques, such as normalization, augmentation, or cropping, can also influence the model’s output by emphasizing certain features over others.28”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28] Tajbakhsh N Jeyaseelan L Li Q Chiang JN Wu Z Ding X Embracing imperfect datasets: a review of deep learning solutions for medical image segmentation. Med Image Anal 2020 63 101693 10.1016/j.media.2020.101693 32289663 source

      Not checked in the free run.

  • Premisenone given

    Improper utilization by end-users can arise when systems are used to analyze images from anatomical districts or imaging modalities different from those they were trained and validated with.

    “Additionally, improper utilization by end-users can also arise when utilizing systems to analyze images from anatomical districts or imaging modalities that differ from those they have been trained and validated with-for example, employing abdominal computed tomography images instead of abdominal magnetic resonance images.”
    Quote in the paperNot a substantial claim
  • Interpretationreasoning

    Variations in equipment, protocols, and standards can cause AI systems to become overfitted to training data characteristics, reducing generalizability and effectiveness in different settings.

    “Such variations can cause AI systems to become overfitted to the characteristics specific to the data they were trained on, reducing their generalizability and effectiveness when deployed in different settings.”
    Quote in the paperSubstantial claim
  • Premisecitation

    In medical imaging, deployment bias can manifest when an algorithm designed for segmentation tasks is used by human operators as a detection tool instead.

    “In medical imaging, this bias can manifest when an algorithm designed for segmentation tasks is utilized by human operators, whether intentionally or inadvertently, as a detection tool instead.37”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [37] Ueda D Kakinuma T Fujita S et al Fairness of artificial intelligence in healthcare: review and recommendations. Jpn J Radiol 2024 42 1 3 15 37540463 PMC10764412 source

      Not checked in the free run.

  • Classified as probably not a substantial claim. 39
  • May combine two claims. 14

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingown data

    Medical publication trends indicate a growing interest in bias in AI.

    “Medical publication trends indicate a growing interest in bias in AI (Figure 1).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Some reference standards are highly accurate but costly and poorly available, whereas others neglect intermediate findings or are operator-dependent, potentially reducing label applicability and reliability.

    “Some could be highly accurate but also costly and poorly available, whereas others could neglect intermediate findings or be operator-dependent,27 potentially reducing label applicability and reliability.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [27] Chen PC Mermel CH Liu Y Evaluation of artificial intelligence on a reference standard based on subjective interpretation. Lancet Digit Health 2021 3 11 e693 e695 10.1016/S2589-7500(21)00216-8 34561202 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Behavioral bias can lead to skewed data cohorts, incomplete information, heightened uncertainty in outcomes, and potential dismissal of algorithm-assisted medical advice.

    “In particular, this bias can lead to skewed data cohorts, incomplete information, heightened uncertainty in outcomes, and potential dismissal of algorithm-assisted medical advice.40”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [40] Olteanu A Castillo C Diaz F Kıcıman E Social data: biases, methodological pitfalls, and ethical boundaries. Front Big Data 2019 2 13 10.3389/fdata.2019.00013 33693336 PMC7931947 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Concept drift arises when the correlation between input variables and output predictions evolves due to fluctuations in data such as variations in image acquisition hardware or protocols, shifts in disease prevalence, or advancements in gold-standard technologies.

    “Specifically, it arises when the correlation between input variables, such as images, and output predictions, such as diagnoses, evolves due to fluctuations in data, such as variations in image acquisition hardware or protocols, shifts in disease prevalence, or advancements in gold-standard technologies.38”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] Roland T Böck C Tschoellitsch T et al Domain shifts in machine learning based Covid-19 diagnosis from blood tests. J Med Syst 2022 46 5 23 35348909 PMC8960704 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Bias in machine learning and predictive modeling is intrinsically tied to the concept of variance.

    “The concept of bias in machine learning (ML) research and more generally in the field of predictive modeling is intrinsically tied to the concept of variance.14”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [14] Geman S Bienenstock E Doursat R Neural networks and the bias/variance dilemma. Neural Comput 1992 4 1 1 58 10.1162/neco.1992.4.1.1 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Uncertainty bias encompasses the influence of uncertainty on decision-making stemming from AI/ML models.

    “Uncertainty bias encompasses the influence of uncertainty on decision-making stemming from AI/ML models.39”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [39] Drukker K Chen W Gichoya J et al Toward fairness in artificial intelligence for medical image analysis: identification and mitigation of potential biases in the roadmap from data collection to model deployment. J Med Imaging(Bellingham) 2023 10 6 061104 10.1117/1.jmi.10.6.061104 37125409 PMC10129875 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Noise, bias, and variance are three components that lead to model performance degradation and negatively affect generalization to new data.

    “From a mathematical point of view, noise (the joint probability distribution between training and test/inference samples), bias, and variance are the three components that lead to model performance degradation and negatively affect generalization to new data.17”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [17] Guan X Burton H Bias-variance tradeoff in machine learning: theoretical formulation and implications to structural engineering applications. Structures 2022 46 17 30 10.1016/j.istruc.2022.10.004 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Behavioral bias pertains to potential distortions in user behavior seen across various platforms, contexts, or datasets.

    “Behavioral bias pertains to the potential distortions in user behavior seen across various platforms, contexts, or datasets.39”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [39] Drukker K Chen W Gichoya J et al Toward fairness in artificial intelligence for medical image analysis: identification and mitigation of potential biases in the roadmap from data collection to model deployment. J Med Imaging(Bellingham) 2023 10 6 061104 10.1117/1.jmi.10.6.061104 37125409 PMC10129875 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    Bias and variance are interdependent, and reducing one typically increases the other.

    “However, it should be made clear that these two entities are interdependent, and reducing one (e.g., variance) typically comes at the expense of increasing the other (i.e., bias), which gives birth to the concept of a bias-variance tradeoff.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Deployment bias emerges when there is a misalignment between the envisioned purpose of a system or algorithm and its actual application.

    “A deployment bias emerges when there is a misalignment between the envisioned purpose of a system or algorithm and its actual application.36”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36] Rajpurkar P Chen E Banerjee O Topol EJ AI in health and medicine. Nat Med 2022 28 1 31 38 10.1038/s41591-021-01614-0 35058619 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Code review can be carried out by an independent experienced coder or auditor for bias detection.

    “This starts with a code review that can be carried out by an independent experienced coder/auditor.49”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [49] Murphy-Hill E Jaspan CN Egelman C Cheng L The Pushback effects of race, ethnicity, gender, and age in code review. Commun ACM 2022 65 3 52 57 10.1145/3474097 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Strategies at the data processing stage may include data augmentation or data filtering to create more representative training datasets.

    “Apart from collecting more data, strategies at the data processing stage may include the creation of more representative training datasets by data augmentation (e.g., by specifically adding underrepresented examples to the data through additional sampling or data generation) or data filtering (e.g., actively undersampling or filtering out undesirable or non-representative samples).68”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [68] Gallegos IO, Rossi RA, Barrow J, et al. Bias and Fairness in large language models: a survey. Published online 2023. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    The composition of the involved teams can influence the AI's propensity for bias.

    “The composition of the involved teams can influence the AI’s propensity for bias.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Incorporating ethical considerations early into the design phase can create AI systems that are less likely to perpetuate or amplify biases.

    “By incorporating the above-mentioned considerations early into the design phase, developers can create systems that are less likely to perpetuate or amplify biases.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The way data are presented to the model during training and how model parameters are updated can introduce bias into the model.

    “The way data are presented to the model during training (affected by the problem formulation and the labeling methodology) and how model parameters are updated (defined through training setup including the objective function) can introduce bias into the model.13, 68”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [13] Obermeyer Z Powers B Vogeli C Mullainathan S Dissecting racial bias in an algorithm used to manage the health of populations. Science 2019 366 6464 447 453 10.1126/science.aax2342 31649194 source

      Not checked in the free run.

    • cites [68] Gallegos IO, Rossi RA, Barrow J, et al. Bias and Fairness in large language models: a survey. Published online 2023. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Bias mitigation algorithms can be categorized according to whether they explicitly or implicitly address bias by accessing or not accessing the bias variables during training.

    “Additionally, algorithms can be categorized according to whether they explicitly or implicitly address bias by accessing or not accessing the bias variables during training.77”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [77] Shrestha R, Kafle K, Kanan C. An investigation of critical issues in bias mitigation techniques. In:IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE; Waikoloa, HI, USA: 2022:2512-2523.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Guiding principles such as transparency, fairness, non-maleficence, and respect for privacy are widely included in recommendations and position papers and can help prevent bias.

    “Guiding principles, such as transparency, fairness, non-maleficence, and respect for privacy from the outset, are widely included in recommendations and position papers and can help to prevent bias (Figure 9).67”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [67] Jobin A Ienca M Vayena E The global landscape of AI ethics guidelines. Nat Mach Intell 2019 1 9 389 399

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Preprocessing bias mitigation algorithms are independent of the model and can be used in a black-box setting by altering the data distribution to increase model fairness.

    “Additionally, these are independent of the model and can be used in a black-box setting by altering the data distribution to increase model fairness.76”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [76] Feldman T, Peake A. End-to-end bias mitigation: removing gender bias in deep learning. Published online June 20, 2021. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Bias mitigation algorithms can be divided into three types according to the phase in which they are applied: preprocessing, during model training, or after model training.

    “Bias mitigation algorithms can be divided into three types according to the phase in which they are applied: in a preprocessing phase, during model training, or after model training.76”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [76] Feldman T, Peake A. End-to-end bias mitigation: removing gender bias in deep learning. Published online June 20, 2021.

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

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Re-sampling and re-weighting algorithms focus on rebalancing the class distribution by adjusting the sample probability or loss weight for majority and minority samples.

    “Re-sampling and re-weighting algorithms focus on rebalancing the class distribution by adjusting the sample probability/loss weight for majority/minority samples.78, 79, 80, 81, 82, 83”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [78] Feldman M, Friedler SA, Moeller J, Scheidegger C, Venkatasubramanian S. Certifying and removing disparate impact. In:Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD ’15. Association for Computing Machinery; 2015:259-268. source

      Not checked in the free run.

    • cites [79] Cui Y, Jia M, Lin TY, Song Y, Belongie S. Class-balanced loss based on effective number of samples. In:IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA, USA: 2019:9260-9269. source

      Not checked in the free run.

    • cites [80] Zou Y, Yu Z, Kumar BVKV, Wang J. Domain adaptation for semantic segmentation via class-balanced self-training. Published online October 25, 2018. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The PROBAST group has initiated the process of developing an AI-specific version of PROBAST called PROBAST-AI, which is still under development.

    “Nonetheless, the current version of PROBAST is not suitable for the evaluation of ML studies, and this is the reason that the PROBAST group has initiated the process of developing an AI-specific version of PROBAST called PROBAST-AI, which is still under development.61”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [61] Collins GS Dhiman P Andaur Navarro CL et al Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ Open 2021 11 7 e048008 10.1136/bmjopen-2020-048008 PMC8273461 34244270 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    AI/ML models can be confidently wrong, meaning they may yield incorrect outcomes with a high level of certainty.

    “Consider this scenario: AI/ML models can be “confidently wrong,” meaning they may yield incorrect outcomes with a high level of certainty.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Premisecitation

    Transparent and comprehensive documentation of the AI system's design choices, data sources, and assumptions made during development is crucial and can help spot sources of bias before, during, and after training.

    “Transparent and comprehensive documentation of the AI system’s design choices, data sources, and any assumptions made during development (e.g., through model cards)75 is crucial and can help spot sources of bias before, during, and after training.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [75] Mitchell M, Wu S, Zaldivar A, et al. Model cards for model reporting. In:Proceedings of the Conference on Fairness, Accountability, and Transparency. ACM; 2019:220-229.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Premisereasoning

    Teams diverse in gender, ethnicity, culture, and professional background bring a wide array of perspectives that can help identify and eliminate potential biases early in the development process.

    “Teams that are diverse in terms of gender, ethnicity, culture, and professional background bring a wide array of perspectives to the table, which can help identify and eliminate potential biases early in the development process.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    XAI methods can be categorized into two main groups: perturbation-based and backpropagation-based explanations.

    “In general, XAI methods can be categorized into two main groups: perturbation-based and backpropagation-based explanations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Algorithmic aversion refers to a phenomenon where clinicians or healthcare professionals exhibit reluctance or skepticism toward relying on AI algorithms for making diagnostic or treatment decisions in medical imaging.

    “Algorithmic aversion refers to a phenomenon where clinicians or healthcare professionals exhibit reluctance or skepticism toward relying on AI algorithms for making diagnostic or treatment decisions in medical imaging.46”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [46] Cadario R Longoni C Morewedge CK Understanding, explaining, and utilizing medical artificial intelligence. Nat Hum Behav 2021 5 12 1636 1642 34183800 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Reporting guidelines such as FUTR-AI and TRIPOD-AI can assist authors of AI manuscripts in reporting their studies according to the Fairness principle, promoting the identification of bias sources.

    “Reporting guidelines, such as the Fairness Universality Traceability Usability Robustness Explainability-AI and TRIPOD-AI, can assist authors of AI manuscripts in reporting their studies according to the Fairness principle, promoting the identification of bias sources.58, 63, 64”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [58] Klontzas ME Gatti AA Tejani AS Kahn CE Jr AI reporting guidelines: how to select the best one for your research. Radiol Artif Intell 2023 5 3 e230055 10.1148/ryai.230055 37293341 PMC10245184 source

      Not checked in the free run.

    • cites [63] Lekadir K, Osuala R, Gallin C, et al. FUTURE-AI: Guiding principles and consensus recommendations for trustworthy artificial intelligence in medical imaging. Published online October 30, 2023.

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

    • cites [64] Collins GS Moons KGM Dhiman P et al TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024 385 7996 e078378 10.1038/d41586-024-00173-x 38626948 PMC11019967 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    Several challenges in handling bias can be related to ambiguities in interpreting results, limited diversity in benchmark datasets, and the subjectivity of detecting bias.

    “However, there are several challenges in handling bias that can be related to ambiguities in interpreting results, limited diversity in benchmark datasets, and the subjectivity of detecting bias.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Automation bias refers to the tendency of individuals to rely excessively on automated systems and to disregard or underutilize their own judgment or critical thinking skills.

    “Automation bias refers to the tendency of individuals to rely excessively on automated systems, such as AI algorithms, and to disregard or underutilize their own judgment or critical thinking skills.42”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [42] Khera R Simon MA Ross JS Automation Bias and Assistive AI: Risk of Harm From AI-driven clinical decision support. JAMA 2023 330 23 2255 2257 10.1001/jama.2023.22557 38112824 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    By identifying and addressing biases, AI medical imaging systems can improve healthcare outcomes and equity.

    “By identifying and addressing biases, these systems can improve healthcare outcomes and equity.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Medical imaging AI systems are sensitive to biases.

    “Understanding that medical imaging AI systems are sensitive to biases is key for their effective real-world integration into clinical practice.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Expecting completely bias-free AI systems may be unrealistic.

    “Before going into detail, it should be acknowledged that expecting completely bias-free systems may be unrealistic.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Complete inclusivity of all data types and sources remains an unattainable goal in model development.

    “Despite the aspiration for unbiased AI, complete inclusivity of all data types and sources remains an unattainable goal in model development.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Article 61 of the EU AI Act mandates that developers of high-risk AI systems establish ongoing, systematic post-market surveillance mechanisms.

    “Concerning monitoring, Article 61 of the legislation mandates that developers of high-risk AI systems establish ongoing, systematic post-market surveillance mechanisms.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Lack of diversity among developers and researchers can worsen AI bias issues, as teams may unconsciously favor perspectives similar to their own.

    “Additionally, the lack of diversity among developers and researchers can worsen these issues, as teams may unconsciously favor perspectives similar to their own.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Subjectivity in the detection of bias can be related to personal interpretation and individual perspectives related to the identification of the bias itself.

    “Subjectivity in the detection of bias can be related to personal interpretation and individual perspectives related to the identification of the bias itself.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Article 10 of the EU AI Act mandates that for high-risk AI systems, training, validation, and testing datasets must be relevant, representative, error-minimal, and complete for their intended use.

    “Article 10 mandates that for high-risk AI systems, training, validation, and testing datasets must be relevant, representative, error-minimal, and complete for their intended use.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    The European Union's AI Act is a pioneering legal framework aimed at regulating AI use, particularly in high-risk applications such as healthcare as defined in Article 6.

    “The recent European Union’s AI Act serves as a pioneering legal framework aimed at regulating AI use, particularly in high-risk applications such as healthcare (as defined in Article 6).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    AI software trained for adult fracture detection is at risk of erroneous results if applied in a pediatric population.

    “For example, AI software trained for adult fracture detection is at risk of erroneous results if applied in a pediatric population.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Premisecitation

    Publicly accessible benchmarks are essential for comparison for AI models and represent a crucial element of open science.

    “Publicly accessible benchmarks are essential for comparison for AI models and represent a crucial element of open science.122”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [122] Bell LC Shimron E Sharing Data Is Essential for the Future of AI in Medical Imaging. Radiol Artif Intell 2024 6 1 e230337 10.1148/ryai.230337 38231036 PMC10831510 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Although sharing data is essential for developing robust AI tools, patient privacy when collecting medical information can pose significant challenges.

    “However, although sharing data is essential for developing robust AI tools, patient privacy when collecting medical information can pose significant challenges.124”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [124] Ferrara E Fairness and bias in artificial intelligence: a brief survey of sources, impacts, and mitigation strategies. Sci 2023 6 1 3 10.3390/sci6010003 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The 2020 Dutch case against the "system risk indication" tool violated privacy laws and wrongly identified innocent people as fraud suspects.

    “The 2020 Dutch case against the “system risk indication” tool, which violated privacy laws and wrongly identified innocent people as fraud suspects, underscores the impact of such biases.128”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [128] SyRI legislation in breach of European Convention on Human Rights. Accessed April 11, 2024.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Different sources of bias-including the data source, algorithm, and users-can contribute to the generation of bias, making identification more cumbersome.

    “Moreover, different sources of bias can contribute to the generation of bias, including the data source, algorithm, and users, which makes the identification more cumbersome.124”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [124] Ferrara E Fairness and bias in artificial intelligence: a brief survey of sources, impacts, and mitigation strategies. Sci 2023 6 1 3 10.3390/sci6010003 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisereasoning

    The EU AI Act is set to fully take effect in 2026 and governs the development, deployment, and use of AI, ensuring safety, transparency, and adherence to ethical standards across the EU.

    “Set to fully take effect in 2026, the act governs the development, deployment, and use of AI, ensuring safety, transparency, and adherence to ethical standards across the EU.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    WHO principles 2 and 5 address bias by mandating that AI tools prioritize human well-being, safety, and public interest.

    “The WHO principles 2 and 5 address bias, mandating that AI tools prioritize human well-being, safety, and public interest.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [127] Ethics and Governance of Artificial Intelligence for Health: WHO guidance. Published online 2021.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The World Health Organization framework poses six key ethical principles for AI in healthcare: protect autonomy; promote human well-being, human safety, and the public interest; ensure transparency, explainability, and intelligibility; foster responsibility and accountability; ensure inclusiveness and equity; and promote AI that is responsive and sustainable.

    “The World Health Organization (WHO) poses six key ethical principles for AI in healthcare in their framework (Figure 9): (1) protect autonomy, (2) promote human well-being, human safety, and the public interest, (3) ensure transparency, explainability, and intelligibility, (4) foster responsibility and accountability, (5) ensure inclusiveness and equity, and (6) promote AI that is responsive and s”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [127] Ethics and Governance of Artificial Intelligence for Health: WHO guidance. Published online 2021.

      Not checked in the free run.

46 of 46

  • The paper relies on this without support.

    AI has been more progressively and widely embraced in medical imaging than in any other medical field.

    “This concern is particularly important in medical imaging, where AI has been more progressively and widely embraced than any other medical field.”

  • The paper relies on this without support.

    The general sources and types of bias in machine learning apply to medical imaging AI.

    “This section focuses on the most common sources of bias that medical imaging professionals, particularly radiologists, may encounter.”

  • The authors state this as an assumption.

    AI models are biased.

    “Even though AI models are assumed to be biased, AI-based bias auditing tools may be leveraged to help mitigate bias.132, 133”
3 of 3

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Incorrect estimation can lead to misinformed decisions, compromising patient safety and treatment efficacy.

    “Incorrect estimation can lead to misinformed decisions, compromising patient safety and treatment efficacy32,33.”

    From Probabilistic medical predictions of large language models (2024) · cites this paper as [32]

Cites

Flagged first, then by citations

19 of the 164 references couldn’t be matched to a known paper.

Research works printed but not linked: 11

  1. 21. Gong J, Currano RM, Sirkin D, Yeung S, Holsinger FC. NICE: Four Human-Centered AI principles for bridging the AI-to-clinic translational gap. In: 2021. Accessed April 8, 2024. [CrossRef]
  2. 74. Hooker S, Moorosi N, Clark G, Bengio S, Denton E. Characterising bias in compressed models. Published online 2020. [CrossRef]
  3. Waikoloa, HI, USA: 2022:2512-2523. [CrossRef]
  4. 92. Ramakrishnan S, Agrawal A, Lee S. Overcoming language priors in visual question answering with adversarial regularization. Published online November 8, 2018. [CrossRef]
  5. 96. Arjovsky M, Bottou L, Gulrajani I, Lopez-Paz D. Invariant risk minimization. Published online March 27, 2020. [CrossRef]
  6. 101. Pezeshki M, Kaba SO, Bengio Y, Courville A, Precup D, Lajoie G. Gradient starvation: a learning proclivity in neural networks. Published online November 24, 2021. [CrossRef]
  7. 106. Lohia PK, Ramamurthy KN, Bhide M, Saha D, Varshney KR, Puri R. Bias mitigation post-processing for individual and group fairness. Published online December 14, 2018. [CrossRef]
  8. 108. Zeiler MD, Fergus R. Visualizing and understanding convolutional networks. Published online November 28, 2013. [CrossRef]
  9. 112. Dabkowski P, Gal Y. Real time image saliency for black box classifiers. Published online May 22, 2017. [CrossRef]
  10. 127. Ethics and Governance of Artificial Intelligence for Health: WHO guidance. Published online 2021. [CrossRef]
  11. 128. SyRI legislation in breach of European Convention on Human Rights. Accessed April 11, 2024. [CrossRef]

Cited by

Sorted by citations

Related

Authors
Koçak, B., Ponsiglione, A., Stanzione, A., Bluethgen, C., Santinha, J., Ugga, L., Huisman, M., Klontzas, M. E., Cannella, R., & Cuocolo, R.
Affiliations
  • Sağlık Bilimleri Üniversitesi (TR)
  • University of Naples Federico II (IT)
  • Champalimaud Foundation (PT)
  • Radboud University Medical Center (NL)
  • University of Crete (GR)
  • University of Palermo (IT)
  • University of Salerno (IT)
Corresponding author
Koçak, Burak
ORCID
Venue
Diagnostic and Interventional Radiology
Publisher
Galenos Yayinevi
Published
2024-07-02
Volume
31
Issue
2
Pages
75-88
ISSN
  • 1300-4360 (ISSN-L)
  • 1305-3825
  • 1305-3612
DOI
10.4274/dir.2024.242854
PMC
PMC11880872
OpenAlex
W4400252889
Semantic Scholar
a42554ce8fc8a3c3b6235abac522bc2a27158281
Cited by
168
References
164
Citations by year
2026: 66 · 2025: 93 · 2024: 9
FWCI
3.83
Type
journal article
Language
en
Open access
Licence
cc-by-nc
Locations
  • Diagnostic and Interventional Radiology (publishedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • Zurich Open Repository and Archive (University of Zurich) (acceptedVersion, cc-by-nc)
  • Radboud Repository (Radboud University) (submittedVersion, other-oa)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Archivio Istituzionale della Ricerca - Università degli Studi della Campania "Luigi Vanvitelli" (submittedVersion)
  • Nova Science Publishers (Nova Science Publishers, Inc.) (submittedVersion, other-oa)
  • Zurich Open Repository and Archive (University of Zurich) (submittedVersion, cc-by-nc)
  • Zurich Open Repository and Archive (University of Zurich)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
Keywords
algorithmic bias, diagnostic imaging, bias detection, bias mitigation, bias avoidance, ethical implications of AI, clinical implementation of AI, patient outcomes
MeSH terms
Algorithms, Artificial Intelligence, Diagnostic Imaging, Humans, Bias

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 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

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