Ethical Considerations in Patient Privacy and Data Handling for AI in Cardiovascular Imaging and Radiology.

Ethical Considerations in Patient Privacy and Data Handling for AI in Cardiovascular Imaging and Radiology - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Imaging Inform Med . 2025 Sep 24;39(3):2004–2025. doi: 10.1007/s10278-025-01656-7 Ethical Considerations in Patient Privacy and Data Handling for AI in Cardiovascular Imaging and Radiology Saba Mehrtabar 1 , Ahmed Marey 2, ✉ , Anushka Desai 3 , Abdelrahman M Saad 2 , Vishal Desai 4 , Julian Goñi 5 , Basudha Pal 6 , Muhammad Umair 7 Author information Article notes Copyright and License information 1 Tabriz University of Medical Sciences, Tabriz, Iran 2 Faculty of Medicine, Alexandria University, Champollion street, Al Mesallah Sharq, Al Attarin, Alexandria Governorate, Alexandria, Egypt 3 Georgetown University School of Medicine, Washington, D.C. USA 4 Silver Edge Government Solutions, Columbia, MD USA 5 Science, Technology and Innovation Studies, University of Edinburgh, Edinburgh, UK 6 Department of Electrical and Computer Engineering, The Johns Hopkins University, Baltimore, MD USA 7 Department of Radiology, Columbia University, Irving Medical Center, New York, NY USA ✉ Corresponding author. Received 2025 Mar 26; Revised 2025 Aug 13; Accepted 2025 Aug 15; Collection date 2026 Jun. © The Author(s) under exclusive licence to Society for Imaging Informatics in Medicine 2025. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. PMC Copyright notice PMCID: PMC13230471  PMID: 40993477 Abstract The integration of artificial intelligence (AI) into cardiovascular imaging and radiology offers the potential to enhance diagnostic accuracy, streamline workflows, and personalize patient care. However, the rapid adoption of AI has introduced complex ethical challenges, particularly concerning patient privacy, data handling, informed consent, and data ownership. This narrative review explores these issues by synthesizing literature from clinical, technical, and regulatory perspectives. We examine the tensions between data utility and data protection, the evolving role of transparency and explainable AI, and the disparities in ethical and legal frameworks across jurisdictions such as the European Union, the USA, and emerging players like China. We also highlight the vulnerabilities introduced by cloud computing, adversarial attacks, and the use of commercial datasets. Ethical frameworks and regulatory guidelines are compared, and proposed mitigation strategies such as federated learning, blockchain, and differential privacy are discussed. To ensure ethical implementation, we emphasize the need for shared accountability among clinicians, developers, healthcare institutions, and policymakers. Ultimately, the responsible development of AI in medical imaging must prioritize patient trust, fairness, and equity, underpinned by robust governance and transparent data stewardship. Graphical Abstract Keywords: Patient privacy, Data security, Artificial intelligence (AI), Cardiovascular imaging, Radiology ethics, Informed consent, Data handling protocols, Blockchain in healthcare, Machine learning in medical imaging, Ethical and legal considerations Introduction Artificial intelligence (AI), first conceptualized in the 1950 s, has evolved substantially with the rise of deep learning and increased computational power [ 1 , 2 ]. While these technological advances have enabled breakthroughs in image analysis and automation, the widespread clinical adoption of AI remains uneven and often slow, due to challenges related to data privacy, bias, regulatory uncertainty, and clinician trust [ 1 , 3 ] In cardiovascular imaging, an area characterized by dynamic physiological data, complex anatomical structures, and high-volume imaging, AI has demonstrated significant potential to enhance diagnostic accuracy, improve workflow efficiency, and support personalized care [ 4 , 5 ]. However, many of the ethical and technical challenges explored in this review are not exclusive to cardiovascular imaging, and instead reflect broader concerns applicable to AI applications across radiology and medicine. Despite all advancements, AI in radiology remains complex, with inherent risks. The widespread use of AI systems introduces potential vulnerabilities, including the amplification of errors on a large scale and the reinforcement of existing data biases. One major challenge lies in the generalizability of AI models which mainly arises from domain shifts or data diversity [ 6 ]. Systems trained on data from specific institutions or populations often struggle to perform well in new or diverse clinical settings due to variations in imaging protocols, equipment, or patient demographics. This limitation can lead to inequities in care, particularly for underrepresented groups or rare conditions [ 7 ]. Moreover, ethical and social concerns span the entire AI lifecycle, from development to clinical implementation. As AI technology advances, thorough studies are necessary to determine the safest and most effective ways to integrate AI into healthcare [ 8 ]. The influence of AI is not only limited to medical image acquisition and analysis but also extends to clinical workflow, patient education, report generation, result dissemination, and even administrative tasks like billing [ 9 , 10 ]. Additionally, AI has the potential to alter professional relationships, patient engagement, and job markets, exacerbating existing disparities in resources [ 8 ]. Figure 1 depicts a possible workflow of how AI can be utilized in medical imaging and the potential cybersecurity vulnerabilities. Fig. 1. Open in a new tab Cybersecurity vulnerabilities of AI integration of medical imaging This figure illustrates a generalized workflow for integrating AI into medical imaging and highlights key cybersecurity vulnerabilities that may arise at different stages of data processing and interpretation. To ensure the ethical application of AI in radiology, it is essential to strike a balance between maximizing its benefits and safeguarding against potential harm, such as unethical data use or other adverse effects. AI development relies on accurately labeled radiological data, significantly increasing its value and raising concerns about commercial access [ 11 ]. Data sharing, a necessity for AI development, introduces further ethical considerations [ 12 ]. Radiologists have an ethical duty to use this data to benefit patients and advance the field while minimizing harm and avoiding bias [ 11 ]. Core ethical principles, including informed consent, data privacy, ownership, transparency, and objectivity, must guide AI’s use in healthcare [ 13 ]. Protecting patient data is of paramount importance, given the sensitivity of health information and the potential risks posed by unauthorized access or misuse. Maintaining patient trust by ensuring privacy and data security is crucial. This review adopts a narrative synthesis approach to explore key ethical concerns surrounding AI in radiology, with special emphasis on patient privacy, data management, informed consent, and regulatory frameworks. Our aim is to examine how emerging technologies intersect with ethical principles, clinical realities, and policy development, offering insights to guide future innovation and safeguard patient trust. This narrative review focuses on three central themes—patient privacy, data handling, and data ownership. Building on these themes, we examine the distributed responsibilities of AI ethics in clinical practice and examine ethical responses through the lenses of (a) transparency, (b) reliability and validity, and regulatory, and (c) legal frameworks. These responses are analyzed from both social and technical perspectives. Figure 2 summarizes the structure of this narrative review. Fig. 2. Open in a new tab Summary of the topics covered by the narrative review By critically assessing AI algorithms and their applications, the medical community can harness the benefits of AI while minimizing its ethical challenges [ 12 ]. This narrative review aims to explore the ethical concerns related to AI in cardiovascular radiology, specifically focusing on patient privacy, data handling, transparency, and reliability. Our goal is to encourage developers, clinicians, and stakeholders to foster AI innovation in ways that prioritize patient safety and equity while preventing harm. Patient Privacy Currently, there are more ethical issues than solutions in radiology, particularly in cardiovascular imaging, as most machine learning (ML) applications in medical imaging are still in their infancy. Here, we outline key ethical challenges and highlight areas for further discussion. In addition to identifying these issues, we also offer a series of recommendations aimed at addressing them. In healthcare, informed consent should extend beyond the immediate clinical encounter to include potential secondary uses of patient data, such as training and validating artificial intelligence systems. Clearly communicating these purposes ensures that patients retain control over how their information contributes to technological development, reinforcing principles of autonomy and trust in the clinician–patient relationship [ 14 ]. Regulatory approaches to health data governance vary significantly across jurisdictions, affecting how AI systems are trained, validated, and implemented. Table 1 provides a comparative overview of key frameworks from the European Union (GDPR), the USA (HIPAA), and China (PIPL), focusing on their handling of consent, data reuse, individual rights, and relevance to AI applications. Table 1. Comparison of health data protection frameworks relevant to AI in medical imaging Feature/region GDPR (EU) HIPAA (USA) PIPL (China) Consent requirement Explicit, informed, specific Not always required if data is de-identified Explicit, informed consent required Re-use of data Needs separate consent for new purposes Allowed under safe harbor de-identification Limited reuse without new consent Right to withdraw consent Yes, at any time Not explicitly covered Yes, but with certain conditions Right to be forgotten Yes No Yes Applies to retrospective data Yes, with re-consent needed Often exempt if de-identified Yes De-identification standards Strict (pseudonymization ≠ anonymization) Flexible (18 identifiers removed) Strict, similar to GDPR Penalties for violations Up to €20 million or 4% of global turnover Up to $1.5 million per violation Up to ¥50 million or 5% of revenue Open in a new tab Regulations such as the European General Data Protection Regulation (GDPR) require explicit, traceable consent for data reuse, especially for purposes beyond direct clinical care, such as AI model training [ 15 ]. These requirements present challenges for both academic and commercial AI research [ 13 ]. GDPR also emphasizes the traceability of healthcare decisions, which reinforces the need for interpretable AI systems [ 16 ]. In contrast, the US Health Insurance Portability and Accountability Act (HIPAA) provides more flexibility, permitting data use without explicit consent if the data is de-identified. However, HIPAA’s de-identification criteria are less stringent and may not adequately protect privacy in complex AI applications [ 17 ]. A major issue under GDPR is the challenge of applying informed consent retroactively to data collected prior to AI-specific regulations [ 18 ]. This limits the use of retrospective datasets for algorithm development, potentially introducing selection bias and reducing model generalizability. Additionally, patients retain the right to withdraw consent. However, once data is used to train AI models, especially in aggregated form, removal becomes technically infeasible. This raises ethical concerns about how patient autonomy can be upheld in AI development, especially when balancing individual rights with collective healthcare benefits. One potential solution is the development of AI models that are inherently designed to accommodate the removal of patient data without compromising their functionality [ 3 ]. China’s Personal Information Protection Law (PIPL) generally requires explicit, informed consent for personal information processing and for cross-border data transfers, especially for sensitive personal information, and also includes a right to withdraw consent. While PIPL allows for other legal bases, consent is often the most suitable basis for clinical trials and sensitive data processing. Like GDPR, PIPL emphasizes strict de-identification standards [ 17 , 19 ]. HIPAA generally requires individual authorization for the use of Protected Health Information (PHI) in research. However, an institutional review board (IRB) or privacy board can waive this requirement if certain safeguards are met, including minimal risk to privacy and the impracticality of conducting the research without the waiver. This waiver mechanism is applicable to retrospective data use for AI development [ 17 , 20 , 21 ]. This fragmented global regulatory landscape creates a significant barrier to the development of generalizable and robust AI models, which require training data from diverse continents and ethnic backgrounds. This “jurisdictional maze” means that what is permissible in one region may not be in another, leading to legal uncertainty, increased compliance costs, and potentially limiting the diversity of training data, thereby directly impacting the ability to mitigate bias and achieve global utility for AI. While de-identification may seem a proper and efficient approach to allow data sharing, it is not always as effective as it should be in protecting patient information. Data de-identified under HIPAA Safe Harbour standards remains vulnerable to re-identification. Re-identification refers to the process of linking ostensibly de-identified health records with external datasets containing identifiable information to accurately infer an individual’s identity. Although the HIPAA Safe Harbor standard requires the removal of 18 specified identifiers such as names, full addresses, and exact dates, it does not account for the potential of quasi-identifiers [ 22 ]. Quasi-identifiers are data attributes that, while not uniquely identifying on their own, can be combined with other information to distinguish or re-identify individuals. Examples include ZIP code, date of birth, gender, ethnicity, or rare clinical characteristics; when cross-referenced with other datasets, these variables can significantly narrow down or reveal an individual’s identity [ 23 ]. Common re-identification techniques include dataset linkage, correlation of overlapping variables across multiple databases, exploitation of rare attribute combinations, application of AI-driven pattern recognition to structured and unstructured data, and extraction of hidden metadata from files. Sweeney et al. (2017) demonstrated that even datasets fully compliant with Safe Harbor can be vulnerable: 3% of individuals were directly re-identified, and an additional 13% were narrowed to fewer than five possible matches by combining quasi-identifiers with publicly available demographic and geographic data [ 24 ]. Radiology faces uniquely acute privacy risks because images themselves can be re-identified. Multiple studies show that faces reconstructed from cranial MRI and even brain PET can be Matched to ordinary photographs with modern face-recognition software, with one experiment achieving over 80% identification accuracy from publicly available datasets. Defacing is the process of algorithmically removing or obscuring facial structures from medical images, and is a common de-identification step. However, these studies demonstrate that residual craniofacial features, skull contours, or partial facial anatomy can still be sufficient for recognition, illustrating the limits of Safe Harbor when quasi-identifiers persist in pixels and metadata [ 25 – 27 ]. Beyond pixels, DICOM headers routinely leak sensitive attributes if not rigorously scrubbed, and inconsistent tooling leaves gaps that adversaries can exploit [ 28 ]. One potential solution to the challenges of data removal and consent withdrawal is the development of AI models inherently designed to accommodate the removal of patient data without compromising their functionality [ 29 ]. To address these challenges more broadly, researchers are actively exploring solutions such as federated learning and differential privacy. These approaches allow AI models to train on decentralized data while preserving privacy [ 30 ]. Federated learning, for instance, ensures that raw data remains at its source, significantly reducing the risk of breaches during data transfer and potentially enabling partial data withdrawal without compromising model performance. Differential privacy adds calibrated noise to data before sharing, protecting individual identities while retaining overall data structure for meaningful AI training. Complementary strategies include strengthening informed consent procedures, improving transparency about data use, and clearly communicating the irreversibility of data once incorporated into model training. Transparency throughout the AI lifecycle can help reinforce trust and respect patients’ rights [ 31 ]. Open-access medical data platforms, such as Kaggle and The Cancer Imaging Archive (TCIA), are critical for developing standard procedures and reproducible results [ 32 ]. Publicly available libraries like the Digital Database for Screening Mammography (DDSM) and Optimam, which contain curated and labeled medical images, also play a significant role in training and validating deep learning algorithms [ 33 , 34 ]. A growing concern is the increasing use of health-related data collected outside of regulated healthcare systems, such as from smartwatches, fitness trackers, and mobile health apps. These devices continuously collect biometric data, which may be used by third parties for AI development [ 35 , 36 ]. However, such data often falls outside the jurisdiction of GDPR, HIPAA, or PIPL, raising serious questions about consent, ownership, and secondary use without oversight. For example, HIPAA protections only extend to health data collected, used, or disclosed by covered entities including hospitals, insurers, healthcare providers, and their business associates. Many AI-powered tools, such as consumer-facing symptom checkers, medical chatbots, mental health apps, or digital phenotyping platforms, are operated by companies that do not meet these definitions. Even if these tools collect identical sensitive health information to that collected by HIPAA-covered providers, the data may not be legally protected [ 37 ]. This “shadow data” represents a significant blind spot in current privacy regulations. While traditional medical data is increasingly protected, a vast and growing reservoir of highly personal health information is being collected and potentially used for AI development without the same consent, ownership, or oversight mechanisms [ 38 ]. This creates a loophole that could undermine the very principles of patient privacy and informed consent, indicating an urgent need for regulatory expansion to cover these emerging data sources and for greater public awareness about the secondary use of biometric and lifestyle data. An additional emerging risk is the growing reliance on synthetic health data which are artificially generated datasets intended to mimic real-world patient records for AI training and validation. This approach is often adopted to overcome barriers to accessing real clinical datasets, including regulatory restrictions, patient privacy concerns, and the need for larger, more diverse samples for algorithm development. While synthetic data is presented as a privacy-preserving alternative, it is not inherently risk-free. Generative models can overfit and inadvertently reproduce identifiable patient information, particularly when trained on small or unique datasets, enabling potential re-identification through linkage with external sources [ 39 ]. Recent work in medical imaging further demonstrates that advanced generative models, such as latent diffusion networks, can memorize and regenerate real patient images, unintentionally exposing sensitive content [ 40 ]. These risks exist even when the source data has been de-identified, as quasi-identifiers and rare clinical patterns can still be preserved and traced back to individuals. Without robust safeguards and clear regulatory guidance, synthetic data could become another vector for privacy breaches, undermining its intended protective role. Confidentiality Confidentiality in healthcare is rooted in the fiduciary duty that binds clinicians to act in their patients’ best interests [ 41 ]. The integration of AI into medical imaging introduces new challenges to this duty, as data are increasingly shared across institutions, vendors, and cloud platforms [ 41 ]. While regulations like GDPR in the European Union set strict boundaries for data use and re-identification, other frameworks such as HIPAA in the USA and PIPL in China adopt differing approaches to confidentiality enforcement. These international variations complicate cross-border AI research and underscore the need for universally aligned ethical standards [ 42 , 43 ]. To ensure effective patient data protection, institutions must move beyond general statements and implement concrete cybersecurity measures. These include AES-256 encryption for data at rest and in transit, multi-factor authentication, regular software updates, staff training to prevent phishing attacks, and adherence to international standards such as the NIST Cybersecurity Framework and ISO/IEC 27001 [ 44 , 45 ]. These protocols are essential for mitigating vulnerabilities that can arise throughout the AI data lifecycle, from acquisition and preprocessing to model deployment. From 2009 to 2019, healthcare data breaches in the USA exposed the records of more than 249 million individuals, with the Majority of these incidents occurring in the last 5 years of that period. This concentration of breaches underscores why healthcare remains a prime target for cyberattacks [ 46 ]. In radiology, picture archiving and communication systems (PACS) and connected modalities remain high-value targets. An analysis of US HHS breach records identified dozens of imaging-linked incidents over 2010–2020 and mapped them to preventable causes like exposed servers or weak email controls, while security reviews document real-world attack paths against PACS networks [ 47 ]. Without proactive and continuous investment in these advanced cybersecurity measures, the benefits of AI could be overshadowed by catastrophic data breaches, eroding patient trust and leading to severe penalties [ 48 ]. It is essential to ensure that qualified cybersecurity professionals are continuously engaged in monitoring and safeguarding patient data systems against potential breaches (Table 2 ). Table 2. Recommended cybersecurity protocols for AI in healthcare Category Protocol Key features/purpose Relevant standard/framework Data at rest/in transit AES-256 Encryption Protects sensitive data from unauthorized access during storage and transfer NIST Cybersecurity Framework, ISO/IEC 27001 Access control Multi-Factor Authentication (MFA) Adds layers of security beyond passwords to verify user identity NIST Cybersecurity Framework Role-Based Access Control (RBAC) Restricts access to data and systems based on user roles and responsibilities ISO/IEC 27001 System maintenance Regular Vulnerability Assessments Identifies and addresses security weaknesses in systems and applications NIST Cybersecurity Framework, ISO/IEC 27001 Software Patching Applies updates to fix known security vulnerabilities in software NIST Cybersecurity Framework Personnel Security Training for Staff Educates staff on best practices to prevent phishing, social engineering, and data mishandling ISO/IEC 27001 Audit and logging Tamper-Proof Audit Logs Records all data access and system activities for accountability and forensic analysis ISO/IEC 27001 Secure Data Storage Infrastructure Ensures physical and digital security of data storage facilities NIST Cybersecurity Framework, ISO/IEC 27001 AI-specific measures Model Access Restrictions Limits who can access and modify AI models and their parameters N/A (emerging best practice) Encrypted APIs Secures data exchange between AI models and other systems N/A (Emerging Best Practice) Compartmentalized Data Flows Isolates sensitive data segments to minimize breach impact N/A (Emerging Best Practice) Open in a new tab Data Handling Digital data in healthcare presents a dual challenge: while enabling significant advancements, it also increases the vulnerability of sensitive medical information to unauthorized access and misuse. AI systems, which fundamentally rely on vast amounts of data, create complex challenges around data sharing and security. Meeting those challenges requires hospital-level data-management capacity—petabyte-scale vendor-neutral archives, high-throughput networks, and dedicated data-engineering teams—investments that have run into multiple millions of dollars at recent US tertiary-center deployments. Data Security with State-of-the-Art AI The extensive reliance of AI systems on large datasets means that breaches can expose highly sensitive patient data, leading to severe consequences [ 32 ]. Notable examples include the 2020 Blackbaud ransomware attack and the cyberattack on Universal Health Services (UHS) in the USA in the same year, both of which resulted in the exposure of patient records, including protected health information (PHI) [ 49 ]. Adversarial Attacks Adversarial attacks are imperceptible, algorithm-crafted perturbations that fool a network, posing a serious security risk to imaging-AI systems. Gradient-based methods such as FGSM, PGD, and Carlini-Wagner can coerce a high-performing model into gross misclassification of otherwise normal studies, with obvious implications for patient safety and clinician trust [ 50 , 51 ]. The most validated defense is adversarial training, in which perturbed samples are injected during learning so the model internalizes more robust decision boundaries [ 52 ]. Hardening should be complemented by runtime integrity checks that flag anomalous inputs, plus secure data pipelines that block model-parameter tampering. While no single safeguard is fool-proof, a layered approach that combines robustness training, continuous validation, and incident logging currently offers the best balance of practicality and protection as attack tactics continue to evolve. AI Synthetic Data AI generative models learn to synthesize data by first studying large collections of real-world examples and identifying the underlying statistical patterns and relationships within them. In the medical domain, this means that AI can create synthetic electronic health records, imaging studies, or physiological waveforms that mirror real patient cases, enabling model development and testing without exposing actual patient identifiers. Once trained, these models can recreate new, artificial samples that mimic the complexity and variability of the original data, without directly copying any individual record. Generative models have the ability to produce authentic yet artificial patient data that is not linked to actual people [ 53 ]. This capability has the potential to facilitate machine learning without the need for long-term utilization of real patient data. This is particularly valuable in healthcare, where strict privacy regulations such as HIPAA and GDPR limit the sharing of identifiable patient information, and where rare disease datasets are often too small to train robust AI systems. Synthetic data can help overcome both privacy barriers and data scarcity by creating balanced, representative datasets for research and model validation. These models may initially need actual data to construct the generative model; however, they provide a potential solution [ 54 , 55 ]. An example of this is the EHR-M-GAN (generative adversarial network), which has the ability to accurately represent the complex and interconnected patterns of patient data across time, including different types of data [ 56 ]. Beyond GANs, other approaches such as variational autoencoders (VAEs) and physics-based simulators have been used to generate medical data. In radiology, VAEs have been applied for tasks such as synthesizing lung nodules in CT scans, generating PET images from MRI inputs, and enhancing low-dose CT quality [ 57 ]. Diffusion models have been also used to generate and denoise medical images. Diffusion models are two-stage models that work by progressively corrupting data with noise in a forward process and training a neural network, usually a denoising U-Net, to reverse this process. In the forward process, Gaussian noise is added to the data over a series of timesteps, eventually turning it into pure isotropic Gaussian noise. The model learns to predict and remove this noise, step by step, starting from a random noise sample. Each timestep encodes the amount of noise present, and the U-Net uses this information to denoise the sample iteratively, gradually reconstructing the original data. This method produces high-quality synthetic data through a precise sequence of learned denoising steps [ 58 ]. In recent times, diffusion models have been known to supersede GANs in generating high-fidelity images in general [ 59 ] and are thus being widely used in medical image generation [ 60 ]. In radiology, diffusion models have been explored not only for creating realistic chest X-rays but also for cross-modality synthesis, such as generating MRI scans from CT inputs to assist in multi-modal AI training without requiring additional patient scans [ 61 ]. These generative approaches may successfully reduce ongoing privacy issues by generating realistic yet fictional patient data that is not associated with real persons. However, it has been observed that generative models are also biased to a specific demographic which is why debiasing such models might be necessary before deployment [ 62 ]. Examples include class-conditional re-weighting of minority groups, embedding an adversarial discriminator that penalizes demographic leakage, and applying rejection sampling to filter biased outputs (e.g., GAMMA-FACE framework) [ 62 ]. There are different types of diffusion models, such as denoising diffusion probabilistic models (DDPM), which focus on pixel-space synthesis, and latent diffusion models (LDM), which operate in a compressed latent space for efficiency. Unconditional diffusion models generate data without relying on any specific input or condition, producing diverse outputs by sampling from the learned data distribution. Conditional and guided diffusion models, on the other hand, incorporate external information, like class labels or text prompts, to steer the generation process toward desired outputs. In Fig. 3 , we show some examples of using an unconditional DDPM to generate chest X-ray Images from the NIH chest X-ray dataset [ 63 , 64 ]. We preprocess the dataset to train it on one-fourth of the original images which is approximately 28,000 images and each sample is resized to 256 × 256 for efficient computation. The top two rows of Fig. 3 showcase exemplar chest X-ray images generated by iterative denoising over 1000 timesteps, while the bottom two rows display examples from the original dataset used for training the model. Fig. 3 . Open in a new tab Top two rows : Exemplar of chest X-ray images generated using an unconditional DDPM by iterative denoising for 1000 timesteps. Bottom two rows : Example images from the original dataset that this DDPM was trained on Radiologists qualitatively reviewed the synthetic chest X-ray images, confirming that they retained key structural and pathological features consistent with real data. These results highlight the potential of unconditional diffusion models to generate high-quality and diagnostically relevant synthetic medical images, further reinforcing their utility in medical image generation tasks. However, the long-term success of synthetic data in radiology will depend on standardized evaluation frameworks, multi-center validation, and ongoing monitoring to ensure that such datasets do not inadvertently introduce hidden biases or compromise patient privacy through data leakage. Data Sharing Data sharing presents another significant challenge due to potential discrepancies in the rules governing personal health information between the location where it is created and the location where it is processed and utilized for deep learning algorithms. In Europe, health-related information is rigorously safeguarded by the GDPR. In contrast, in the USA, specialized legislation such as HIPAA is in place to protect health data [ 65 , 66 ]. These jurisdictional differences can create vulnerabilities in data-sharing laws that might be exploited [ 67 , 68 ]. The disparities between GDPR and HIPAA, for instance, introduce potential vulnerabilities in cross-border data sharing. Additionally, obtaining training data from diverse continents and ethnic backgrounds is crucial for ensuring the global utility of AI models. This emphasizes the urgent need to find solutions to the ethical and legal obstacles associated with such global sharing of healthcare data. The existence of numerous worldwide protocols can add complexity to this procedure, and providing separate training for each ethnic group may not be cost-effective [ 69 ]. Most critically, deep learning algorithms are prone to data biases. To guarantee these networks are useful worldwide, training data from many continents, geographies, and cultures is essential. This necessitates training the network with diverse inputs, highlighting the imperative to resolve ethical and legal issues related to data diversity and sharing [ 32 ]. In cardiovascular imaging, where subtle anatomical variations and modality-specific calibration are common, limited data sharing can lead to underperforming algorithms in underserved or non-Western populations. The lack of global data harmonization undermines the promise of AI to democratize high-quality diagnostics [ 70 , 71 ]. Additional barriers to data sharing include the following [ 72 , 73 ]: technical incompatibilities in data formats (e.g., non-standard DICOM tags) lack of uniform data annotation protocols institutional reluctance to share proprietary data or patient information lack of incentives or credit for data contributors Emerging initiatives, such as federated learning, can help mitigate these limitations by enabling collaborative AI development without centralized data pooling. While promising, these methods require robust coordination, interoperable infrastructure, and institutional trust to succeed. Efforts such as the Medical Imaging and Data Resource Center (MIDRC) in the USA and the Open Access Series of Imaging Studies (OASIS) show how centralized platforms, when ethically governed, can accelerate AI development by pooling anonymized datasets. However, widespread adoption remains constrained by legal and operational hurdles [ 74 – 76 ]. Beyond legal and regulatory aspects, other challenges to data sharing exist. Interoperability is an obstacle to data sharing between entities regardless of how close or far they are from one another geographically. In health informatics, interoperability refers to the capacity of diverse information systems, applications, and devices to exchange, interpret, and use data seamlessly across organizational and geographic boundaries. Achieving this requires not only technical connectivity but also standardized data models and terminologies that preserve meaning during transfer. One of the most widely adopted frameworks supporting this is HL7 (Health Level Seven International), a global standards body that develops specifications for electronic health information exchange. Among its most influential recent developments is FHIR (Fast Healthcare Interoperability Resources), first introduced in 2014 and now implemented in healthcare systems worldwide. FHIR defines modular “resources” (e.g., Patient, Observation, ImagingStudy) and uses modern web technologies such as RESTful APIs and JSON/XML formats, enabling faster, more flexible integration between clinical, imaging, and research systems. Originally piloted in countries such as the USA, Canada, and Australia, FHIR adoption has expanded rapidly into Europe, the UK, and parts of Asia, becoming a cornerstone in efforts to make clinical and imaging data shareable at scale.[a]. Cardiovascular imaging datasets travel through a long chain. Modalities (CT, MR, echo) are transferred to PACS/VNAs then to EHRs/registries then potentially to research platforms. Each transaction exposes an interoperability seam [ 77 ]. Even when DICOM is used consistently, critical context (acquisition parameters, labeling, versioned annotations) is often incomplete, inconsistently encoded, or trapped in proprietary fields; this degrades reusability, hampers de-identification, and fuels site-specific batch effects that complicate multi-center analysis. Recent work shows that mapping imaging to computable clinical data models remains non-trivial at scale, requiring bespoke pipelines to transform DICOM metadata into FHIR and revealing gaps in terminologies and resource maturity [ 78 – 80 ]. Additionally, research tooling itself is in transition. Imaging-aware workflows built on DICOMweb + FHIR are maturing, and position papers in radiology emphasize standardization and tight integration between images and structured clinical data. But at present, deploying these architectures requires substantial local engineering, careful metadata curation, organizational change, and practical frictions that limit privacy-preserving sharing at scale for cardiovascular imaging cohorts [ 81 , 82 ]. To unlock the full potential of AI in radiology, there is a critical need for international standards, cross-border data use agreements, and ethical frameworks that balance innovation with patient privacy. Mitigation Strategies: Federated Learning, Differential Privacy, and Beyond To mitigate these risks, several privacy-preserving technologies have been developed. Figure 4 illustrates potential solutions. Federated learning, for instance, allows multiple institutions to collaborate on training AI models without sharing raw data. This decentralized approach ensures that data remains at its source, significantly reducing the risk of breaches during data transfer. Federated learning has already been piloted in several healthcare applications, such as training AI models for medical imaging across hospitals without centralizing sensitive data [ 83 , 84 ]. Fig. 4. Open in a new tab A multi-layered security framework for protecting patient data in AI applications used in radiology and cardiovascular imaging Differential privacy, another key approach, adds noise to data before it is shared, preserving the privacy of individuals while still allowing for meaningful AI training [ 85 ]. In differential privacy, carefully calibrated noise is added to data points, typically from Gaussian o Laplacian distributions, to protect individual identities while retaining the data’s overall structure [ 86 ]. This process ensures that, although individual entries are slightly altered, aggregate analyses remain largely unaffected, preserving the reliability of insights. For healthcare research, this means that diagnostics and outcomes can be analyzed without compromising patient privacy, as the noise cancels out at the group level, allowing for accurate epidemiological insights [ 87 ]. For example, differential privacy has been used in healthcare research to anonymize patient data in large-scale epidemiological studies, enabling insights without compromising individual privacy [ 88 ]. Moreover, advanced cryptography techniques like secure multi-party computation (SMPC) and homomorphic encryption (HE) allow for the secure computation of AI algorithms on encrypted data, ensuring that sensitive health information remains protected throughout the process. These methods have seen increasing adoption in healthcare AI to secure patient data during training and testing of models [ 89 ]. Blockchain technology offers another promising solution for securing healthcare data. By creating immutable, decentralized records of transactions, blockchain can ensure data integrity and traceability, making it difficult for unauthorized parties to tamper with patient records. One notable example is Estonia’s eHealth Foundation, which has successfully implemented blockchain to secure over one million medical records, allowing patients to control access to their own health data [ 90 ]. Nevertheless, the integration of these technologies into healthcare is now in its early phase. It is essential to collaborate and make efforts to standardize data security procedures in the field of AI. Organizations such as the International Organization for Standardization (ISO) and other entities are now collaborating to develop universally applicable standards that guarantee data security in AI applications [ 91 , 92 ]. In the USA, projects are exploring the use of blockchain for securely sharing medical data between hospitals, patients, and insurance companies. For example, the MedRec system, developed by MIT, leveraged blockchain to create a transparent and secure method for managing medical records [ 93 ]. This ensures that patients maintain ownership of their data while enabling healthcare providers to access up-to-date information, reducing the risk of errors and breaches. These efforts are part of a broader trend toward using blockchain to tackle the complex challenges of healthcare data security, ensuring that sensitive information remains secure even as it is shared across different entities. To sum up, as AI continues to transform healthcare, robust data protection strategies are crucial to safeguard patient privacy and maintain trust. Breaches, such as the ransomware attacks on UHS and Blackbaud, highlight the need for advanced security measures like federated learning, differential privacy, and blockchain technology. The integration of these technologies into healthcare remains in its early stages, but ongoing pilot projects and collaborations indicate their potential to address current vulnerabilities. Standardizing these practices across the industry, guided by organizations such as the International Organization for Standardization (ISO), will be essential in creating a secure, ethical future for AI-driven healthcare. Figure 5 represents data flow in the cloud. Fig. 5. Open in a new tab Data flow in the cloud Data Ownership and Accountability Health systems using AI and ML, particularly in fields like radiology and cardiology, face significant challenges in defining responsibility and accountability. As these technologies become more integrated into clinical practice, the question of who bears responsibility for their implementation, use, and outcomes becomes increasingly complex. This shared responsibility among physicians, developers, healthcare institutions, industry, and regulatory bodies can lead to ethical dilemmas, particularly in high-stake areas like cardiovascular imaging. One of the primary ethical dilemmas arises when AI-derived diagnostic recommendations differ from a physician’s judgment. For example, if an AI system in cardiovascular imaging recommends a particular intervention that a radiologist/cardiologist disagrees with, but the treating physician/primary provider decides to follow the AI recommendation, who is responsible if the patient experiences a negative outcome? Conversely, if the physician disregards the AI recommendation and an adverse event occurs, there is potential for ethical and legal conflicts regarding the accountability of both the physician and the AI system’s developers [ 8 ]. 1- Physicians and Radiologists The integration of AI into clinical imaging workflows will inevitably reshape the roles and responsibilities of radiologists and physicians. While AI holds the promise of enhancing diagnostic accuracy, reducing workload, and accelerating reporting times, its successful implementation depends on clinician trust, proper training, and clearly defined accountability structures [ 94 , 95 ]. This can be particularly challenging in cardiovascular imaging, where AI models are being used to predict complex conditions. Further complicating factors include the primary explainability of the models and associated transparent vs. black box nature of the model in question as well as how well the radiologist or other physician understands the model and its capabilities [ 96 , 97 ]. Physicians must balance the AI’s recommendations with their clinical judgment, patient-specific factors including demographics pertinent to the patient as compared to that of the training dataset of the model in question, and ethical considerations such as patient autonomy [ 94 , 95 ]. Surveys have shown that clinicians remain cautiously optimistic about AI. A 2021 European Society of Radiology (ESR) white paper revealed that most radiologists view AI as a supportive tool rather than a replacement, with 64% believing AI will enhance their diagnostic capabilities but should not operate autonomously. Similarly, a 2020 US survey among cardiologists and radiologists reported that while many see AI as potentially valuable, concerns remain about explainability, legal liability, and workflow disruption [ 98 – 101 ]. To facilitate clinical adoption, institutions should: provide targeted training in AI literacy, including interpretation of algorithm outputs integrate AI seamlessly into existing imaging platforms (e.g., PACS) ensure explainable AI (XAI) elements are available, allowing clinicians to verify decisions establish clear protocols on who holds responsibility when AI-informed decisions lead to adverse outcomes Importantly, AI should be positioned as an augmentative tool, not a replacement. In cardiovascular imaging, AI can assist with tasks such as automated left ventricular function quantification or coronary artery calcium scoring, but final interpretation and decision-making must remain under the physician’s control [ 102 , 103 ]. Maintaining human oversight is vital for trust and ethical accountability, especially when AI models operate as black boxes. There is also growing recognition of the need for interdisciplinary collaboration [ 104 ]. Radiologists, data scientists, and engineers must work together from the earliest stages of AI development to ensure clinical relevance, reduce bias, and improve usability. Institutions should support multidisciplinary AI committees to oversee evaluation, implementation, and post-deployment monitoring. Additionally, physicians have a moral obligation to use AI technologies to improve patient outcomes while avoiding harm. This includes recognizing biases in AI algorithms that may exacerbate health disparities, ensuring that AI tools are applied equitably across diverse patient populations, and addressing issues of fairness in AI decision-making. Physicians should also ensure that AI systems are continuously monitored for performance and reliability, maintain patient autonomy by informing them of AI involvement in care, and make sure that decisions taken by the algorithm are accurate, understandable, and explainable using their domain expertise. Ethical dilemmas may arise when AI models perform better for certain demographic groups, creating a disparity in healthcare outcomes [ 105 , 106 ]. Ultimately, empowering radiologists and physicians to lead AI integration will enhance acceptance, safeguard patient safety, and ensure that technological innovation aligns with clinical priorities. 2- Developers and Manufacturers of AI Systems AI developers, both academic researchers and commercial manufacturers, play a critical role in ensuring that medical AI systems are clinically useful, ethically sound, and legally compliant. In the context of radiology and cardiovascular imaging, developers must address unique challenges including variable imaging protocols, patient heterogeneity, and strict data protection standards [ 107 – 109 ]. However, ethical dilemmas arise when there is a lack of transparency in how AI models function. The “black box” nature of many AI systems, where the decision-making process is not fully understandable even to the developers, raises concerns about accountability. Developers must ensure that their AI models are explainable, particularly in sensitive areas like cardiovascular diagnostics, where patients and physicians alike need to understand the rationale behind AI-driven decisions. Recent governance tools provide pragmatic safety nets when full mechanistic explainability is unattainable: the PRIME checklist for cardiovascular-imaging AI requires prospectively defined update triggers, public model-card metrics, and ongoing audit logs, while the IMDRF 2025 Good Machine Learning Practice/SaMD guidance mandates continuous post-market surveillance to detect drift, bias, and emerging risks. Together these frameworks turn routine performance monitoring—rather than saliency maps—into the primary assurance mechanism for clinical safety and ethical integrity [ 110 , 111 ]. One of the most impactful examples of collaborative standardization is the development of AI tools for coronary artery calcium (CAC) scoring. Recognizing the need for consistency across vendors and imaging modalities, manufacturers collaborated to establish standardized CAC thresholds and validated reference datasets. These efforts have enabled more reliable and comparable risk assessments in clinical cardiology and have facilitated regulatory approval of CAC-based AI software [ 112 , 113 ]. Another area where developers have contributed meaningfully is the adoption and expansion of the DICOM (Digital Imaging and Communications in Medicine) standard. AI models often require high-fidelity, annotated input data with uniform metadata. Enhancing DICOM standards to include structured reports, segmentation labels, and metadata harmonization has been essential for training robust algorithms. Initiatives such as IHE (Integrating the Healthcare Enterprise) further support cross-platform integration and interoperability between AI software and clinical imaging systems [ 82 , 114 , 115 ]. In addition, it is imperative to provide healthcare personnel with comprehensive documentation and training, clarifying the operational mechanisms of AI systems, the data they utilize, and the decision-making processes they employ. This level of transparency builds assurance and enables the accurate utilization of AI techniques in clinical practice [ 116 , 117 ]. Moreover, the ethical responsibility of developers extends beyond the initial deployment of AI systems. Finally, the emergence of model monitoring tools such as performance drift detectors and bias auditing systems will be increasingly critical in the post-deployment phase. Developers must ensure that deployed models are not only effective at the time of release but also remain safe and equitable as clinical environments evolve. Healthcare Institutions Healthcare institutions, including hospitals and clinics, are responsible for integrating AI technologies into their workflows. This may include and not limited to selecting AI tools that meet clinical needs, are ethically developed when it comes to data privacy/handling and patient autonomy, are preferably explainable, have optimal diagnostic yield with scientifically proven improved patient outcomes and cost effective besides providing adequate training for healthcare personnel, and establishing monitoring systems to ensure AI systems function as intended. Ethical dilemmas arise when institutions prioritize cost-saving (which is different than cost-effective) AI solutions over patient safety, potentially compromising care quality [ 118 , 119 ]. For example, in cardiovascular imaging, institutions might adopt AI algorithms that streamline the interpretation of cardiovascular imaging studies, e.g., cardiac CT and cardiac MRI, but if these algorithms are not adequately tested in diverse patient populations, they could introduce biases that impact patient outcomes. Institutions must also address the challenge of cybersecurity, ensuring that patient data used in AI applications is protected according to regulatory standards like GDPR and HIPAA. In the event of a data breach, ethical and legal questions emerge about the institution’s responsibility to protect patient privacy [ 107 , 120 ]. LLMs and Data Ownership The rise of large language models (LLMs) such as GPT-4 has transformed how data is collected, shared, and reused in medicine and radiology. Unlike traditional AI systems that are trained on narrowly curated datasets, LLMs are often built from vast collections of internet, academic, and clinical text, which may include sensitive medical information obtained from public datasets, inadvertently disclosed case reports, or scraped institutional materials [ 121 ]. This shift raises new concerns about patient privacy, because once information enters an LLM’s training set, it cannot easily be removed without retraining the entire model [ 122 ]. There is also evidence that LLMs can “memorize” and reproduce rare or unique data points verbatim, potentially re-exposing identifiable health information. Meaning LLMs can store exact sequences of words, sentences, or even full paragraphs from its training data, and later reproduce them exactly as they appeared, without altering a single word or punctuation mark [ 123 ]. In radiology, this risk extends to model outputs that might recall details from anonymized imaging reports or structured datasets used during training, thereby undermining de-identification safeguards. Furthermore, LLM APIs often process user queries in the cloud, where data is sent for processing, meaning that clinical prompts or pasted patient information could be stored or logged by third-party providers, creating additional points of vulnerability [ 124 ]. These concerns have intensified with high-profile legal and ethical challenges. For example, in December 2023 the New York Times sued OpenAI and Microsoft, alleging that their LLMs had been trained on millions of copyrighted articles without permission and in ways that allowed the models to reproduce substantial portions of those works [ 125 ]. While this case is about journalism, it has clear parallels to medicine: if proprietary imaging datasets, radiology reports, or patient education materials were used in LLM training without consent, similar disputes over intellectual property and privacy could emerge. The opacity of LLM training pipelines makes it difficult for hospitals, regulators, or patients to know whether their data has been included, complicating consent, ownership, and accountability frameworks. As LLM-driven tools become embedded in radiology workflows, such as automated report drafting or patient communication, institutions must establish clear governance for data sourcing, implement privacy-preserving fine-tuning methods, and require transparent documentation of training data provenance to prevent both ethical breaches and legal exposure. Transparency Building trust among patients, healthcare providers, and other stakeholders depends on transparency in AI and ML in healthcare. This includes clarity regarding how AI systems operate, the data they utilize, and the choices they generate. Transparency in AI decision-making is essential for building confidence among both patients and healthcare providers. When AI recommendations are clear and understandable, patients are more likely to trust these technologies, particularly in high-stake fields such as cardiovascular imaging [ 126 ]. However, transparency must be balanced with security and privacy concerns to avoid unintended consequences, such as exposing sensitive information or creating vulnerabilities to cyberattacks [ 95 ]. One key aspect of transparency is patient-centered communication. AI systems should be designed with user-friendly interfaces that allow patients to understand the reasoning behind AI-generated recommendations [ 127 ]. For instance, in cardiovascular imaging, AI systems could present diagnostic findings in a way that highlights both the certainty and uncertainty of the analysis. This can involve using visual aids, such as probability scores or confidence intervals, to convey the likelihood of different outcomes. Providing this information enables patients to make more informed decisions about their care and fosters a sense of collaboration between patients, physicians, and AI technologies [ 128 ]. A key aspect of transparency is communicating uncertainty, especially when AI systems produce probabilistic outputs or operate under variable data quality [ 129 ]. These uncertainties must be clearly conveyed to both physicians and patients to ensure safe and informed decision-making. Clinicians, in particular, require insight into model confidence and limitations in order to appropriately integrate AI outputs into diagnostic workflows. Another core element is the availability of explanation mechanisms. While earlier sections have discussed the importance of explainable AI (XAI), here it is worth emphasizing that transparency must extend to system design, training data sources, and validation procedures. AI systems deployed in clinical practice should clearly document [ 130 , 131 ]: the datasets used for training, including population diversity and imaging modalities performance metrics, stratified by subgroups (e.g., age, sex, ethnicity) known limitations and potential failure modes Moreover, AI systems should be transparent about their limitations. For example, if an AI model is known to be less accurate for certain demographics or conditions, this information should be communicated clearly to both the patient and the healthcare provider [ 132 ]. This helps set realistic expectations and allows for a more nuanced interpretation of AI-driven recommendations. Addressing uncertainty in AI predictions, especially in complex areas like cardiovascular diagnostics, can prevent over-reliance on these systems and encourage collaborative decision-making between AI and human expertise. To enhance patient-centered transparency, AI developers must prioritize user-friendly interfaces that allow both patients and radiologists to interact with AI systems effectively [ 127 ]. On the other hand, radiologists must educate AI developers and offer assistance in the development process. Since most AI interpretation and implementation occur at the radiologist level, interfaces should be designed to support radiologists in conveying complex findings to patients in an accessible manner. This involves simplifying the presentation of complex data for patient comprehension, while providing radiologists with detailed visualizations and insights that support their clinical expertise. For instance, in cardiovascular imaging, AI could use visual representations of heart structures to aid radiologists in explaining diagnostic findings, offering layered views of information. These could include both comprehensive clinical data for radiologists and simplified visuals for patient education, ensuring patients grasp the relevance of findings to their health [ 128 ]. Beyond previous regulatory bodies we discussed, certain countries are pioneering AI transparency efforts. For example, the UK’s National Health Service (NHS) has launched the NHS AI Lab, which is focused on developing and regulating AI in healthcare. The NHS AI Lab emphasizes transparency, explainability, and patient involvement in AI development, aiming to build public trust in AI-driven healthcare solutions [ 133 ]. In cardiovascular imaging, the NHS has been piloting AI technologies that help detect heart disease earlier and more accurately, with an emphasis on ensuring that these technologies are transparent to both clinicians and patients. Additionally, Singapore has emerged as a leader in AI governance, with its Ministry of Health issuing the AI in Healthcare Guidelines, which set out principles for AI deployment, including transparency and explainability [ 134 ]. These guidelines emphasize the importance of making AI models interpretable by healthcare providers and ensuring that patients understand how AI is used in their care. Singapore’s approach highlights the importance of a patient-centered regulatory framework that balances innovation with safety and transparency. Moreover, collaboration between radiologists and data scientists or AI engineers should be actively encouraged, enabling both groups to combine their expertise and translate it into the greatest possible benefit for patients. Radiologists must be involved in developing not only the product but also in the deployment process, exchanging knowledge with developers. Ultimately, transparent AI systems empower clinicians to exercise critical judgment, enable patients to make informed choices, and support regulators in monitoring safe implementation. Not only that, but it will also empower engineers and developers with enough guidance, domain expertise, and understanding of the problems they tackle. It is not just a technical feature, nor is it just a clinical and ethical imperative but a blend of both. Reliability and Validity Reliability and validity are not only statistical benchmarks for any AI system, but they are operational metrics where AI systems are tested in real busy clinical environments. Also, reliability refers to consistent reproducible performance across different settings, while validity denotes the model’s ability to produce accurate, clinically meaningful results. In medical imaging, particularly in cardiovascular contexts, these qualities are foundational to safe and ethical AI integration [ 135 ]. Although AI models have shown significant potential in accurately analyzing radiology images, development and implementation face substantial challenges [ 136 ]. These challenges include the need for extensive and well-labeled datasets, the potential for automation bias, and variability in image quality [ 137 , 138 ]. Additionally, it is critical to distinguish between regulatory approval and clinical effectiveness. AI tools that meet regulatory standards such as FDA clearance or CE marking are often evaluated using retrospective datasets under controlled conditions [ 139 ]. Real-world implementation, by contrast, introduces challenges such as imaging protocol variability, scanner heterogeneity, and patient demographic shifts. As highlighted by van Leeuwen et al., performance often declines in prospective or external validation settings, raising concerns about generalizability and robustness [ 6 , 140 , 141 ]. These discrepancies are partly due to dataset shift and population drift, where changes in patient demographics, disease prevalence, or imaging technology reduce model performance over time. For example, an AI model trained on coronary CT scans from a single tertiary center may underperform when applied to a rural population with different baseline risks or imaging equipment [ 139 ]. A successful deployment of an AI system requires significant resources. These resources go beyond procurement and integration alone. Deploying clinical AI at scale requires specialized multidisciplinary teams. Such team should include radiologists, AI engineers, data scientists, IT professionals, workflow integrators, and others [ 141 ]. This is crucial to collaboratively oversee performance, troubleshoot failures, and coordinate updates. A recent example comes from the University of Miami Miller School of Medicine, where an IRB-approved pilot leveraged such a team to evaluate staffing, infrastructure, workflows, and financial feasibility before scaling AI tools into radiology practice. It also demands dedicated infrastructure, like GPU-enabled servers or hybrid cloud-edge systems, secure DICOM pipelines, and data processing capabilities tailored for high-throughput imaging. The implementation of a vendor-neutral AI deployment platform, combining radiologists, data scientists, software developers, and hospital IT, is a proven solution to manage the complexity of maintaining multiple narrow AI algorithms in clinical practice [ 142 ]. Moreover, workflow adaptation is essential: training radiologists and technologists, embedding fallback protocols for AI downtime, and ensuring seamless integration into RIS/PACS systems are critical [ 143 ]. Continuous governance including drift detection, recalibration cycles, version control, and empowered decision rights to suspend models is vital. Most of such governance would be carried out by data scientists and AI engineers but radiologists must be well-informed when and why models could be suspended [ 144 ]. Finally, budgetary planning must encompass not just licensing but also costs for staff time, IT support, cybersecurity, infrastructure maintenance, and compliance audits, as highlighted in the Miller School study which modeled substantial downstream clinical and financial benefits from investment in these areas [ 144 ]. To address this AI tools should disclose confidence scores and be integrated into multimodal decision pathways, where human oversight is maintained. This is particularly important for safety–critical applications such as coronary artery stenosis detection or automated ejection fraction measurement. Institutions should also support post-market surveillance programs, similar to those used for pharmaceuticals, to track AI performance longitudinally. Independent registries for AI outcomes stratified by use-case, population, and vendor would help build transparency and accelerate continuous improvement [ 145 ]. Ultimately, an AI tool’s value lies not in its regulatory status, but in its sustained, equitable benefit to real-world patient care. An increasingly recognized pillar of reliability is lifecycle governance—the set of processes that ensure an algorithm remains safe and performant long after the carefully controlled conditions of its pivotal study. Best-practice frameworks now recommend that every high-risk model be overseen by an institutional AI oversight board made up of clinicians, data scientists, ethicists, and patient representatives. This board authorizes an initial silent-trial phase, defines minimum acceptance thresholds (e.g., AUC, calibration error) for local deployment, and mandates routine drift audits at pre-specified intervals (monthly for rapid-learning models, quarterly for static ones). If the board’s dashboard signals a clinically relevant decline—commonly set at a ≥ 5% drop in primary accuracy or a statistically significant shift in input-data distribution—the model must be recalibrated or withdrawn pending investigation. International regulators are converging on similar expectations. The FDA’s 2025 draft guidance on AI-enabled device-software functions requires manufacturers to file a predetermined change control plan that spells out post-market monitoring methods, the criteria that will trigger a model update, and the evidence package needed before the updated model can be re-enabled in routine care. Across the Atlantic, the EU Artificial Intelligence Act classifies most imaging algorithms as “high-risk” systems and imposes a continuous post-market monitoring system with obligatory performance logging, incident reporting, and periodic summary reports to national competent authorities; these provisions start to apply from August 2025 for general-purpose models and from August 2027 for embedded medical-device AI. Together, these regulatory instruments reinforce the ethical imperative for healthcare institutions to embed robust drift detection and recalibration workflows—governed by transparent, multidisciplinary decision-making—into the very fabric of AI deployment. Bias One of the key ethical challenges in AI for cardiovascular imaging is managing bias, which can stem from the datasets used to train these models, technical imaging factors, and variations in data acquisition protocols across different centers. Bias in cardiovascular imaging AI refers to systematic deviations in model performance that lead to unequal accuracy across patient subgroups, clinical contexts, or imaging environments [ 69 ]. From a technical perspective, these deviations can emerge from several sources. Representation bias arises when training datasets fail to proportionally capture the diversity of the intended clinical population in terms of demographics, comorbidities, and disease phenotypes Cardiovascular imaging datasets often over-represent certain populations, typically patients from specific geographic regions, ethnic backgrounds, or socioeconomic groups. Consequently, models trained on such data may not perform as reliably for diverse patient populations. For example, a system trained primarily on European-descent cohorts may not account for phenotypic differences in African or Asian patients, resulting in measurable reductions in sensitivity and specificity when deployed in these settings. Furthermore, biases related to comorbidities, gender, and age can significantly impact the performance of AI models in cardiovascular imaging. Spectrum and measurement biases are the ones to look out for in such cases. Spectrum bias occurs when the distribution of disease severity, comorbidity profiles, or imaging findings in the training set does not match real-world case mixes. Measurement bias can be introduced by heterogeneity in image acquisition protocols, reconstruction algorithms, and equipment calibration across sites [ 146 ]. For example, cardiovascular diseases often present differently in women compared to men, and an AI model trained predominantly on male patients might miss important signs of disease in female patients [ 147 ]. Similarly, age-related changes in the cardiovascular system can affect imaging results, leading to potential bias if AI models are not adequately trained to account for these differences. Technical factors such as variations in image acquisition protocols, scanning techniques, and equipment can also introduce bias, further complicating the generalizability of AI models [ 95 ]. Automation bias occurs when humans favor machine-generated conclusions over contradictory facts or human decisions [ 148 ]. Automation bias causes omissions and commissions [ 149 ]. Omission mistakes arise when radiologists ignore AI tool failures. Omission mistakes are more likely in high decision flow rates, where radiologist reads radiological exams quickly. This is aggravated by AI conclusions based on small features humans cannot notice. Commission mistakes result from accepting or implementing a machine’s conclusion despite opposing evidence or from misinterpretation of machine-generated results. Automation bias concerns may be higher in resource-poor populations especially in a setting where there may not be a local radiologist to disapprove outcomes [ 95 ]. To mitigate the impact of bias and improve the generalizability of AI in cardiovascular imaging, several strategies can be employed. First, dataset diversification should be pursued by incorporating underrepresented populations and a broader range of disease presentations [ 81 ]. This includes incorporating data from different populations, including underrepresented groups, and ensuring that the data reflects a wide range of clinical scenarios. One approach to achieving this is through the use of conditional generative models, which can help generate synthetic data that captures diverse characteristics and capturing rare groups [ 150 ], hereby enhancing dataset diversity and improving the generalizability of AI models in medical imaging applications [ 96 , 151 ]. Second, multicenter collaborations are essential for introducing acquisition and population heterogeneity into training data, thereby improving robustness [ 6 ]. Additionally, standardization of acquisition protocols can reduce measurement bias, and site-stratified validation can detect population-specific performance degradation prior to clinical rollout. Once deployed, post-market bias audits and drift surveillance can identify emerging disparities in output distributions, while structured human-in-the-loop review processes can help counter automation bias by ensuring that discordant findings trigger expert review rather than passive acceptance. It is also important to note that mitigating bias has become an increasingly active research focus in recent years, driven by the rapid expansion of the medical AI market. Local Evaluation and Drift Governance Local silent trials. Before a high-risk imaging algorithm is exposed to clinicians, it should complete a silent trial —a time-boxed phase in which the model runs in the background on consecutive patients, its outputs hidden from decision-making, while performance is benchmarked against verified ground-truth labels. Silent trials routinely surface workflow frictions and failure modes that are invisible in retrospective testing, providing a low-risk bridge from bench to bedside [ 152 ]. Drift surveillance with SPC charts. Once live, reliability is best tracked with statistical process-control (SPC) charts , an established quality-engineering tool now applied to AI. By plotting core metrics—area under the curve, calibration error, or a Kolmogorov–Smirnov statistic for input distributions—against three-sigma control limits, SPC dashboards flag distribution shifts or accuracy decay early enough for remediation [ 153 ]. Trigger thresholds can be tuned to model criticality (e.g., monthly reviews for rapid-learning systems, quarterly for static ones). Decision rights and oversight. Threshold breaches should be adjudicated by a multidisciplinary AI oversight board chaired by the chief medical officer but including data scientists, frontline clinicians, informaticians, legal/compliance officers, and patient advocates. The board ratifies acceptance criteria, approves recalibration or de-commissioning, and publishes audit summaries—ensuring that no single stakeholder can override patient-safety considerations for operational convenience. Conclusion In conclusion, the integration of AI in cardiovascular imaging presents complex ethical challenges that must be balanced with patient rights and welfare. Ethical AI in healthcare demands a focus on transparency, fairness, informed consent, patient privacy, and data ownership. Researchers, developers, and regulators must work together to create robust standards that address these concerns while ensuring that AI models are explainable, generalizable, and transparent about their biases. The ethical challenges of AI in cardiovascular imaging require a sociotechnical approach that views technical affordances and social practices as interconnected. This perspective emphasizes the need to design distributed responsibilities [ 154 ], recognize how technologies evolve through use beyond their initial design [ 155 ], and foreground the uncertainties inherent in technological deployment [ 156 , 157 ]. Socio-technical frameworks such as Responsible Research and Innovation (RRI) [ 158 , 159 ] can inform strategies to incorporate these considerations in clinical technology development and application. In brief, the RRI framework aims to orient technoscientific knowledge development and application around the core values of anticipation, reflexivity, inclusion, and responsiveness [ 160 ], fostering dialogue with the social sciences and humanities to enhance the societal benefits of innovation [ 161 ]. Further work could explore how sociotechnical frameworks such as RRI could improve our capacity to address the ethical challenges of AI in cardiovascular imaging. As AI continues to evolve, ethical considerations should guide its development, with a focus on maximizing patient benefits while minimizing risks and constantly reflecting on our current practice. This requires continuous collaboration among all stakeholders to ensure that AI not only advances medical practice but also protects the dignity, privacy, and trust of patients. Technological change brings about new ethical challenges and often prompts us to reexamine existing ones. Embedding ethical principles and socio-technical perspectives into clinical settings means constantly reflecting on how technological changes in our practice improve patient outcomes while also anticipating and being responsive to the new vulnerabilities they help shape. This approach can provide cardiovascular practitioners with the confidence that they are responsibly harnessing the potential of new technologies. Acknowledgements We gratefully acknowledge the paper by Rahman et al., “Exploring the Effect of Image Enhancement Techniques on COVID-19 Detection using Chest X-ray Images” (Rahman et al. 2020, Computers in Biology and Medicine, 10.1016/j.compbiomed.2021.104607), for inspiring aspects of our methodology in generating synthetic chest X-ray images. Additionally, we thank B.P. for providing the detailed methods employed in creating the generative images used in this study. Data Availability Not applicable. Declarations Ethics Approval and Consent to Participate Not applicable. Consent for Publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Kaul V, Enslin S, Gross SA. History of artificial intelligence in medicine. Gastrointest Endosc. 2020;92(4):807-12. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Coiera EW. Artificial Intelligence in Medicine: The Challenges Ahead. Journal of the American Medical Informatics Association. 1996;3(6):363-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Jiang B, Guo N, Ge Y, Zhang L, Oudkerk M, Xie X. Development and application of artificial intelligence in cardiac imaging. British Journal of Radiology. 2020;93(1113):20190812. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Bernard O, Lalande A, Zotti C, Cervenansky F, Yang X, Heng PA, et al. Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the Problem Solved? IEEE Transactions on Medical Imaging. 2018;37(11):2514-25. [ DOI ] [ PubMed ] [ Google Scholar ] 5. Sengupta PP, Dey D, Davies RH, Duchateau N, Yanamala N. Challenges for augmenting intelligence in cardiac imaging. The Lancet Digital Health. 2024;6(10):e739-e48. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Yang Y, Zhang H, Gichoya JW, Katabi D, Ghassemi M. The limits of fair medical imaging AI in real-world generalization. Nature Medicine. 2024;30(10):2838-48. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Eche T, Schwartz LH, Mokrane F-Z, Dercle L. Toward Generalizability in the Deployment of Artificial Intelligence in Radiology: Role of Computation Stress Testing to Overcome Underspecification. Radiology: Artificial Intelligence. 2021;3(6):e210097. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Kohli M, Prevedello LM, Filice RW, Geis JR. Implementing Machine Learning in Radiology Practice and Research. AJR Am J Roentgenol. 2017;208(4):754-60. [ DOI ] [ PubMed ] [ Google Scholar ] 9. Erickson BJ, Korfiatis P, Akkus Z, Kline TL. Machine Learning for Medical Imaging. Radiographics. 2017;37(2):505-15. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Mittelstadt BD, Floridi L. The Ethics of Big Data: Current and Foreseeable Issues in Biomedical Contexts. Sci Eng Ethics. 2016;22(2):303-41. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Geis JR, Brady A, Wu CC, Spencer J, Ranschaert E, Jaremko JL, et al. Ethics of artificial intelligence in radiology: summary of the joint European and North American multisociety statement. Insights into Imaging. 2019;10(1):101. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Murdoch B. Privacy and artificial intelligence: challenges for protecting health information in a new era. BMC Medical Ethics. 2021;22(1):122. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Brady AP, Neri E. Artificial Intelligence in Radiology-Ethical Considerations. Diagnostics (Basel). 2020;10(4). [ DOI ] [ PMC free article ] [ PubMed ] 14. Kadam RA. Informed consent process: A step further towards making it meaningful! Perspectives in Clinical Research. 2017;8(3). [ DOI ] [ PMC free article ] [ PubMed ] 15. Geis J.R. BAP, Wu C.C., Spencer J., Ranschaert E., Jaremko J.L., Langer S.G., Borondy Kitts A., Birch J., Shields W.F., et al. Ethics of AI in Radiology: Joint European and North American Multisociety Statement. [(accessed on 16 April 2020)] [Available from: https://www.acr.org/-/media/ACR/Files/Informatics/Ethics-of-AI-in-Radiology-European-and-North-American-Multisociety-Statement%2D%2D6-13-2019.pdf . 16. Fotaki A, Puyol-Antón E, Chiribiri A, Botnar R, Pushparajah K, Prieto C. Artificial Intelligence in Cardiac MRI: Is Clinical Adoption Forthcoming? Front Cardiovasc Med. 2021;8:818765. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. De-identification of Protected Health Information: How to Anonymize PHI [Available from: https://www.hipaajournal.com/de-identification-protected-health-information/ . 18. Kalkman S, Mostert M, Gerlinger C, van Delden JJM, van Thiel G. Responsible data sharing in international health research: a systematic review of principles and norms. BMC Med Ethics. 2019;20(1):21. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. GDPR Versus PIPL – Key Differences and Implications for Compliance in China [Available from: https://www.china-briefing.com/news/pipl-vs-gdpr-key-differences-and-implications-for-compliance-in-china/ . 20. Health Information Privacy [Available from: https://www.hhs.gov/hipaa/for-professionals/privacy/laws-regulations/index.html . 21. Tovino SA. Artificial Intelligence and the HIPAA Privacy Rule: A Primer. Houston Journal of Health Law & Policy. 2025;24(1):77-126. [ Google Scholar ] 22. Dankar FK, El Emam K, Neisa A, Roffey T. Estimating the re-identification risk of clinical data sets. BMC Medical Informatics and Decision Making. 2012;12(1):66. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Garfinkel S. De-Identification of Personal Information. NIST Interagency/Internal Report (NISTIR), National Institute of Standards and Technology, Gaithersburg, MD; 2015. [ Google Scholar ] 24. Sweeney L, Yoo JS, Perovich L, Boronow KE, Brown P, Brody JG. Re-identification Risks in HIPAA Safe Harbor Data: A study of data from one environmental health study. Technol Sci. 2017;2017. [ PMC free article ] [ PubMed ] 25. Schwarz CG, Kremers WK, Therneau TM, Sharp RR, Gunter JL, Vemuri P, et al. Identification of Anonymous MRI Research Participants with Face-Recognition Software. N Engl J Med. 2019;381(17):1684-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Schwarz CG, Kremers WK, Lowe VJ, Savvides M, Gunter JL, Senjem ML, et al. Face recognition from research brain PET: An unexpected PET problem. NeuroImage. 2022;258:119357. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Jwa AS, Koyejo O, Poldrack RA. Demystifying the likelihood of reidentification in neuroimaging data: A technical and regulatory analysis. Imaging Neuroscience. 2024;2:1-18. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Aryanto KYE, Oudkerk M, van Ooijen PMA. Free DICOM de-identification tools in clinical research: functioning and safety of patient privacy. European Radiology. 2015;25(12):3685-95. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Williamson SM, Prybutok V. Balancing Privacy and Progress: A Review of Privacy Challenges, Systemic Oversight, and Patient Perceptions in AI-Driven Healthcare. Applied Sciences. 2024;14(2):675. [ Google Scholar ] 30. Li T, Sahu AK, Talwalkar A, Smith V. Federated Learning: Challenges, Methods, and Future Directions. IEEE Signal Processing Magazine. 2020;37(3):50-60. [ Google Scholar ] 31. Petersen SE, Abdulkareem M, Leiner T. Artificial Intelligence Will Transform Cardiac Imaging-Opportunities and Challenges. Front Cardiovasc Med. 2019;6:133. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Yadav N, Pandey S, Gupta A, Dudani P, Gupta S, Rangarajan K. Data Privacy in Healthcare: In the Era of Artificial Intelligence. Indian Dermatol Online J. 2023;14(6):788-92. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Lee RS, Gimenez F, Hoogi A, Miyake KK, Gorovoy M, Rubin DL. A curated mammography data set for use in computer-aided detection and diagnosis research. Sci Data. 2017;4:170177. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Halling-Brown MD, Warren LM, Ward D, Lewis E, Mackenzie A, Wallis MG, et al. OPTIMAM Mammography Image Database: A Large-Scale Resource of Mammography Images and Clinical Data. Radiol Artif Intell. 2021;3(1):e200103. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Shandhi MMH, Singh K, Janson N, Ashar P, Singh G, Lu B, et al. Assessment of ownership of smart devices and the acceptability of digital health data sharing. NPJ Digit Med. 2024;7(1):44. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Li M, Xu P, Hu J, Tang Z, Yang G. From challenges and pitfalls to recommendations and opportunities: Implementing federated learning in healthcare. Medical Image Analysis. 2025;101:103497. [ DOI ] [ PubMed ] [ Google Scholar ] 37. McGraw D, Mandl KD. Privacy protections to encourage use of health-relevant digital data in a learning health system. NPJ Digital Medicine. 2021;4(1):2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Hu Y, Li Z, Liu Z, Zhang Y, Qin Z, Ren K, et al. Membership Inference Attacks Against Vision-Language Models2025. 39. El Emam K, Mosquera L, Bass J. Evaluating Identity Disclosure Risk in Fully Synthetic Health Data: Model Development and Validation. J Med Internet Res. 2020;22(11):e23139. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Dar SUH, Seyfarth M, Ayx I, Papavassiliu T, Schoenberg SO, Siepmann RM, et al. Unconditional Latent Diffusion Models Memorize Patient Imaging Data: Implications for Openly Sharing Synthetic Data. 2025. [ DOI ] [ PubMed ] 41. Paterick ZR, Paterick TE. Preparticipation Cardiovascular Screening of Student-Athletes with Echocardiography: Ethical, Clinical, Economic, and Legal Considerations. Current Cardiology Reports. 2019;21(3):16. [ DOI ] [ PubMed ] [ Google Scholar ] 42. How do the European Union’s GDPR and China’s PIPL regulate cross-border data flows? [Available from: https://ipr.blogs.ie.edu/2025/01/27/how-do-the-european-unions-gdpr-and-chinas-pipl-regulate-cross-border-data-flows/ . 43. Wang C, Zhang J, Lassi N, Zhang X. Privacy Protection in Using Artificial Intelligence for Healthcare: Chinese Regulation in Comparative Perspective. Healthcare [Internet]. 2022; 10(10). [ DOI ] [ PMC free article ] [ PubMed ] 44. Strengthening cybersecurity for patient care and data protection. Available from: https://www.ruralhealth.us/blogs/2025/04/strengthening-cybersecurity-for-patient-care-and-data-protection . [ DOI ] [ PMC free article ] [ PubMed ] 45. Information security, cybersecurity and privacy protection. Available from: https://www.iso.org/standard/27001 . 46. Seh AH, Zarour M, Alenezi M, Sarkar AK, Agrawal A, Kumar R, et al. Healthcare Data Breaches: Insights and Implications. Healthcare (Basel). 2020;8(2). [ DOI ] [ PMC free article ] [ PubMed ] 47. Bowers GM, Kleinpeter ML, Rials WT. Securing Your Radiology Practice: Evidence-Based Strategies for Radiologists Compiled From 10 Years of Cyberattacks and HIPAA Breaches Involving Medical Imaging. Perspect Health Inf Manag. 2022;19(3):1c. [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. Smart Ethics in the Digital World Proceedings of the ETHICOMP 2024. 49. Reddy J. Data breaches in healthcare security systems: University of Cincinnati; 2021. [ Google Scholar ] 50. Finlayson SG, Bowers JD, Ito J, Zittrain JL, Beam AL, Kohane IS. Adversarial attacks on medical machine learning. Science. 2019;363(6433):1287-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Ma X, Niu Y, Gu L, Wang Y, Zhao Y, Bailey J, et al. Understanding adversarial attacks on deep learning based medical image analysis systems. Pattern Recognition. 2021;110:107332. [ Google Scholar ] 52. Goodfellow IJ, Shlens J, Szegedy C. Explaining and Harnessing Adversarial Examples. 2015. [ Google Scholar ] 53. Choi E, Biswal S, Malin B, Duke J, Stewart WF, Sun J. Generating Multi-label Discrete Patient Records using Generative Adversarial Networks. In: Finale D-V, Jim F, David K, Rajesh R, Byron W, Jenna W, editors. Proceedings of the 2nd Machine Learning for Healthcare Conference; Proceedings of Machine Learning Research: PMLR; 2017. p. 286--305. [ Google Scholar ] 54. Yoon J, Drumright LN, Van Der Schaar M. Anonymization through data synthesis using generative adversarial networks (ads-gan). IEEE journal of biomedical and health informatics. 2020;24(8):2378-88. [ DOI ] [ PubMed ] [ Google Scholar ] 55. Baowaly MK, Lin C-C, Liu C-L, Chen K-T. Synthesizing electronic health records using improved generative adversarial networks. Journal of the American Medical Informatics Association. 2019;26(3):228-41. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Li J, Cairns BJ, Li J, Zhu T. Generating synthetic mixed-type longitudinal electronic health records for artificial intelligent applications. NPJ Digit Med. 2023;6(1):98. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 57. Teli A. Uncover This Tech Term: Variational Autoencoders. Korean J Radiol. 2025;26(6):616-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 58. Chen H, Xiang Q, Hu J, Ye M, Yu C, Cheng H, et al. Comprehensive exploration of diffusion models in image generation: a survey. Artificial Intelligence Review. 2025;58(4):99. [ Google Scholar ] 59. Dhariwal P, Nichol A. Diffusion models beat GANs on image synthesis. Proceedings of the 35th International Conference on Neural Information Processing Systems. Curran Associates Inc; 2024. p. Article 672. 60. Khader F, Müller-Franzes G, Tayebi Arasteh S, Han T, Haarburger C, Schulze-Hagen M, et al. Denoising diffusion probabilistic models for 3D medical image generation. Scientific Reports. 2023;13(1):7303. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 61. Zhu L, Xue Z, Jin Z, Liu X, He J, Liu Z, et al. Make-A-Volume: Leveraging Latent Diffusion Models for Cross-Modality 3D Brain MRI Synthesis. 2023. [ Google Scholar ] 62. Pal B, Kannan A, Kathirvel RP, O’Toole AJ, Chellappa R, editors. GAMMA-FACE: GAussian Mixture Models Amend Diffusion Models for Bias Mitigation in Face Images. Computer Vision – ECCV 2024; 2025 2025//; Cham: Springer Nature Switzerland. 63. NIH Chest X-rays. 64. Rahman T, Khandakar A, Qiblawey Y, Tahir A, Kiranyaz S, Abul Kashem SB, et al. Exploring the effect of image enhancement techniques on COVID-19 detection using chest X-ray images. Computers in Biology and Medicine. 2021;132:104319. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 65. Price WN, 2nd, Cohen IG. Privacy in the age of medical big data. Nat Med. 2019;25(1):37-43. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 66. Panch T, Mattie H, Celi LA. The "inconvenient truth" about AI in healthcare. NPJ Digit Med. 2019;2:77. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 67. Iacobucci G. Patient data were shared with Google on an "inappropriate legal basis," says NHS data guardian. Bmj. 2017;357:j2439. [ DOI ] [ PubMed ] [ Google Scholar ] 68. Vincent J. Privacy advocates sound the alarm after Google grabs DeepMind UK health app. The Verge. 2018;14. 69. Ferrara E. Fairness and Bias in Artificial Intelligence: A Brief Survey of Sources, Impacts, and Mitigation Strategies. Sci [Internet]. 2024; 6(1). 70. Rahnasto J. Genetic data are not always personal-disaggregating the identifiability and sensitivity of genetic data. J Law Biosci. 2023;10(2):lsad029. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 71. Kumar K, Samanth M, Bharathi M, Rane S, Kumar A, Mufeed Ahmad S, et al. Advancements in Cardiovascular Imaging Modalities: Integrating Artificial Intelligence and Multi-modal Imaging for Enhanced Diagnosis, Risk Stratification, and Treatment Monitoring. Journal of Cardiovascular Disease Research. 2025;16:355-94. [ Google Scholar ] 72. Piergiovanni M, Mennecozzi M, Barale-Thomas E, Danovi D, Dunst S, Egan D, et al. Bridging imaging-based in vitro methods from biomedical research to regulatory toxicology. Arch Toxicol. 2025;99(4):1271-85. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 73. Fassnacht M, Benz C, Heinz D, Leimstoll J, Satzger G. Barriers to Data Sharing among Private Sector Organizations2023. 74. Rieke N, Hancox J, Li W, Milletarì F, Roth HR, Albarqouni S, et al. The future of digital health with federated learning. NPJ Digit Med. 2020;3:119. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 75. Li N, Lewin A, Ning S, Waito M, Zeller MP, Tinmouth A, et al. Privacy-preserving federated data access and federated learning: Improved data sharing and AI model development in transfusion medicine. Transfusion. 2025;65(1):22-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 76. Farahani B, Monsefi AK. Smart and collaborative industrial IoT: A federated learning and data space approach. Digital Communications and Networks. 2023;9(2):436-47. [ Google Scholar ] 77. Osamika D, Adelusi B, Theresa M, Kelvin-Agwu C, Mustapha A, Forkuo A, et al. A Critical Review of Health Data Interoperability Standards: FHIR, HL7, and Beyond. 2025. [ Google Scholar ] 78. Mackenzie A, Lewis E, Loveland J. Successes and challenges in extracting information from DICOM image databases for audit and research. Br J Radiol. 2023;96(1151):20230104. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 79. Iancu A, Bauer J, May MS, Prokosch HU, Dörfler A, Uder M, et al. Large-Scale Integration of DICOM Metadata into HL7-FHIR for Medical Research. Methods Inf Med. 2024;63(3-04):77-84. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 80. Tabari P, Costagliola G, De Rosa M, Boeker M. State-of-the-Art Fast Healthcare Interoperability Resources (FHIR)–Based Data Model and Structure Implementations: Systematic Scoping Review. JMIR Med Inform. 2024;12:e58445. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 81. Tang S-T, Tjia V, Noga T, Febri J, Lien C-Y, Chu W-C, et al. Creating a Medical Imaging Workflow Based on FHIR, DICOMweb, and SVG. Journal of Digital Imaging. 2023;36(3):794-803. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 82. Jeon K, Park W, Kahn C, Nagy P, You S, Yoon SH. Advancing Medical Imaging Research Through Standardization: The Path to Rapid Development, Rigorous Validation, and Robust Reproducibility. Investigative radiology. 2024;60. [ DOI ] [ PMC free article ] [ PubMed ] 83. Yang Q, Liu Y, Chen T, Tong Y. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST). 2019;10(2):1-19. [ Google Scholar ] 84. Moshawrab M, Adda M, Bouzouane A, Ibrahim H, Raad A. Reviewing Federated Machine Learning and Its Use in Diseases Prediction. Sensors (Basel). 2023;23(4). [ DOI ] [ PMC free article ] [ PubMed ] 85. Liu J, Hu Y, Guo X, Liang T, Jin W. Differential privacy performance evaluation under the condition of non-uniform noise distribution. Journal of Information Security and Applications. 2022;71:103366. [ Google Scholar ] 86. Bi X, Shen X. Distribution-Invariant Differential Privacy. (0304-4076 (Print)). [ DOI ] [ PMC free article ] [ PubMed ] 87. Dyda A, Purcell M, Curtis S, Field E, Pillai P, Ricardo K, et al. Differential privacy for public health data: An innovative tool to optimize information sharing while protecting data confidentiality. Patterns. 2021;2(12):100366. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 88. Jain P, Gyanchandani M, Khare N. Differential privacy: its technological prescriptive using big data. Journal of Big Data. 2018;5(1):1-24. [ Google Scholar ] 89. Torkzadehmahani R, Nasirigerdeh R, Blumenthal DB, Kacprowski T, List M, Matschinske J, et al. Privacy-Preserving Artificial Intelligence Techniques in Biomedicine. Methods Inf Med. 2022;61(S 01):e12-e27. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 90. Tagde P, Tagde S, Bhattacharya T, Tagde P, Chopra H, Akter R, et al. Blockchain and artificial intelligence technology in e-Health. Environmental Science and Pollution Research. 2021;28(38):52810-31. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 91. Kasyapa MSB, Vanmathi C. Blockchain integration in healthcare: a comprehensive investigation of use cases, performance issues, and mitigation strategies. Front Digit Health. 2024;6:1359858. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 92. Saeed H, Malik H, Bashir U, Ahmad A, Riaz S, Ilyas M, et al. Blockchain technology in healthcare: A systematic review. PLoS One. 2022;17(4):e0266462. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 93. Azaria A. MedRec: Using Blockchain for Medical Data Access and Permission Management. MIT Media Lab. 94. O'neil C. Weapons of math destruction: How big data increases inequality and threatens democracy: Crown; 2017. [ Google Scholar ] 95. Geis JR, Brady A, Wu CC, Spencer J, Ranschaert E, Jaremko JL, et al. Ethics of artificial intelligence in radiology: summary of the joint European and North American multisociety statement. Insights Imaging. 2019;10(1):101. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 96. Marey A, Serdysnki KC, Killeen BD, Unberath M, Umair M. Applications and implementation of generative artificial intelligence in cardiovascular imaging with a focus on ethical and legal considerations: what cardiovascular imagers need to know! BJR|Artificial Intelligence. 2024;1(1):ubae008. [ DOI ] [ PMC free article ] [ PubMed ] 97. Marey A, Arjmand P, Alerab ADS, Eslami MJ, Saad AM, Sanchez N, et al. Explainability, transparency and black box challenges of AI in radiology: impact on patient care in cardiovascular radiology. Egyptian Journal of Radiology and Nuclear Medicine. 2024;55(1):183. [ Google Scholar ] 98. Current practical experience with artificial intelligence in clinical radiology: a survey of the European Society of Radiology. Insights Imaging. 2022;13(1):107. [ DOI ] [ PMC free article ] [ PubMed ] 99. Brady AP, Allen B, Chong J, Kotter E, Kottler N, Mongan J, et al. Developing, Purchasing, Implementing and Monitoring AI Tools in Radiology: Practical Considerations. A Multi-Society Statement from the ACR, CAR, ESR, RANZCR and RSNA. Radiol Artif Intell. 2024;6(1):e230513. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 100. Becker CD, Kotter E, Fournier L, Martí-Bonmatí L, European Society of R. Current practical experience with artificial intelligence in clinical radiology: a survey of the European Society of Radiology. Insights into Imaging. 2022;13(1):107. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 101. Najjar R. Redefining Radiology: A Review of Artificial Intelligence Integration in Medical Imaging. Diagnostics (Basel). 2023;13(17). [ DOI ] [ PMC free article ] [ PubMed ] 102. Grant JK, Javaid A, Carrick RT, Koester M, Kassamali AA, Kim CH, et al. Digital health innovation and artificial intelligence in cardiovascular care: a case-based review. npj Cardiovascular Health. 2024;1(1):26. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 103. Alsharqi M, Edelman ER. Artificial Intelligence in Cardiovascular Imaging and Interventional Cardiology: Emerging Trends and Clinical Implications. Journal of the Society for Cardiovascular Angiography & Interventions. 2025;4(3). [ DOI ] [ PMC free article ] [ PubMed ] 104. Abgrall G, Holder AL, Chelly Dagdia Z, Zeitouni K, Monnet X. Should AI models be explainable to clinicians? Crit Care. 2024;28(1):301. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 105. Fenech ME, Buston O. AI in Cardiac Imaging: A UK-Based Perspective on Addressing the Ethical, Social, and Political Challenges. Front Cardiovasc Med. 2020;7:54. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 106. Ipsos M. Public views of machine learning. Findings From Public Research Engagement Conducted on Behalf of the Royal Society London: Ipsos MORI, The Royal Society. 2017. [ Google Scholar ] 107. Mennella C, Maniscalco U, De Pietro G, Esposito M. Ethical and regulatory challenges of AI technologies in healthcare: A narrative review. Heliyon. 2024;10(4):e26297. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 108. Siala H, Wang Y. SHIFTing artificial intelligence to be responsible in healthcare: A systematic review. Social Science & Medicine. 2022;296:114782. [ DOI ] [ PubMed ] [ Google Scholar ] 109. Larson DB, Harvey H, Rubin DL, Irani N, Tse JR, Langlotz CP. Regulatory Frameworks for Development and Evaluation of Artificial Intelligence-Based Diagnostic Imaging Algorithms: Summary and Recommendations. J Am Coll Radiol. 2021;18(3 Pt A):413-24. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 110. Sengupta PP, Shrestha S, Berthon B, Messas E, Donal E, Tison GH, et al. Proposed Requirements for Cardiovascular Imaging-Related Machine Learning Evaluation (PRIME): A Checklist: Reviewed by the American College of Cardiology Healthcare Innovation Council. JACC: Cardiovascular Imaging. 2020;13(9):2017-35. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 111. Dubowik M. MDR Regulator [Internet]2025 2025/01/30/T14:22:56+00:00. Available from: https://mdrregulator.com/news/mhra-imdrfs-latest-guidance-on-ai-and-medical-device-software . 112. Verma S, Maerkisch L, Paderno A, Gilberg L, Teodorescu B, Meyer M. One scan, multiple insights: A review of AI-Driven biomarker imaging and composite measure detection in lung cancer screening. Meta-Radiology. 2025;3(1):100124. [ Google Scholar ] 113. Wang TW, Tzeng Y-H, Wu K-T, Liu H-R, Hong J-S, Hsu H-Y, et al. Meta-analysis of deep learning approaches for automated coronary artery calcium scoring: Performance and clinical utility AI in CAC scoring: A meta-analysis: AI in CAC scoring: A meta-analysis. Computers in Biology and Medicine. 2024;183:109295. [ DOI ] [ PubMed ] [ Google Scholar ] 114. Larobina M. Thirty Years of the DICOM Standard. Tomography. 2023;9(5):1829-38. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 115. Cohen RY, Sodickson AD. An Orchestration Platform that Puts Radiologists in the Driver's Seat of AI Innovation: a Methodological Approach. J Digit Imaging. 2023;36(2):700-14. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 116. Balasubramaniam N, Kauppinen M, Rannisto A, Hiekkanen K, Kujala S. Transparency and explainability of AI systems: From ethical guidelines to requirements. Information and Software Technology. 2023;159:107197. [ Google Scholar ] 117. Markus AF, Kors JA, Rijnbeek PR. The role of explainability in creating trustworthy artificial intelligence for health care: A comprehensive survey of the terminology, design choices, and evaluation strategies. Journal of Biomedical Informatics. 2021;113:103655. [ DOI ] [ PubMed ] [ Google Scholar ] 118. Maleki Varnosfaderani S, Forouzanfar M. The Role of AI in Hospitals and Clinics: Transforming Healthcare in the 21st Century. Bioengineering (Basel). 2024;11(4). [ DOI ] [ PMC free article ] [ PubMed ] 119. Lämmermann L, Hofmann P, Urbach N. Managing artificial intelligence applications in healthcare: Promoting information processing among stakeholders. International Journal of Information Management. 2024;75:102728. [ Google Scholar ] 120. The impact of the General Data Protection Regulation(GDPR) on artificial intelligence. 121. Bommasani R, Hudson DA, Adeli E, Altman R, Arora S, Arx Sv, et al. On the Opportunities and Risks of Foundation Models. 2022. 122. Kadavath S, Conerly T, Askell A, Henighan T, Drain D, Perez E, et al. Language Models (Mostly) Know What They Know. 2022. [ Google Scholar ] 123. Carlini N, Tramer F, Wallace E, Jagielski M, Herbert-Voss A, Lee K, et al. Extracting Training Data from Large Language Models. 2021. [ Google Scholar ] 124. Hartmann V, Suri A, Bindschaedler V, Evans D, Tople S, West R. SoK: Memorization in General-Purpose Large Language Models. 2023. [ Google Scholar ] 125. New York Times Sues OpenAI and Microsoft Over Use of Copyrighted Work - The New York Times. 126. Martin C, DeStefano K, Haran H, Zink S, Dai J, Ahmed D, et al. The ethical considerations including inclusion and biases, data protection, and proper implementation among AI in radiology and potential implications. Intelligence-Based Medicine. 2022;6:100073. [ Google Scholar ] 127. Amershi S, Weld D, Vorvoreanu M, Fourney A, Nushi B, Collisson P, et al. Guidelines for Human-AI Interaction. Chi '19. 2019:1–13. [ Google Scholar ] 128. Grigera J, Espada JP, Rossi G. AI in User Interface Design and Evaluation. IT Professional. 2023;25(2):20-2. [ Google Scholar ] 129. Saeed W, Omlin C. Explainable AI (XAI): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems. 2023;263:110273. [ Google Scholar ] 130. Arora A, Alderman JE, Palmer J, Ganapathi S, Laws E, McCradden MD, et al. The value of standards for health datasets in artificial intelligence-based applications. Nat Med. 2023;29(11):2929-38. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 131. Bernal J, Mazo C. Transparency of Artificial Intelligence in Healthcare: Insights from Professionals in Computing and Healthcare Worldwide. Applied Sciences. 2022;12(20):10228. [ Google Scholar ] 132. Ueda D, Kakinuma T, Fujita S, Kamagata K, Fushimi Y, Ito R, et al. Fairness of artificial intelligence in healthcare: review and recommendations. Jpn J Radiol. 2024;42(1):3-15. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 133. Chinu, Bansal U. Explainable AI: To Reveal the Logic of Black-Box Models. New Generation Computing. 2024;42(1):53-87. [ Google Scholar ] 134. MOH | Artificial Intelligence in Healthcare. 135. Balagurunathan Y, Mitchell R, El Naqa I. Requirements and reliability of AI in the medical context. Phys Med. 2021;83:72-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 136. Waller J, O’Connor A, Raafat E, Amireh A, Dempsey J, Martin C, et al. Applications and challenges of artificial intelligence in diagnostic and interventional radiology. Pol J Radiol. 2022;87:113-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 137. Aldoseri A, Al-Khalifa KN, Hamouda AM. Re-Thinking Data Strategy and Integration for Artificial Intelligence: Concepts, Opportunities, and Challenges. Applied Sciences [Internet]. 2023; 13(12). 138. Tejani AS, Ng YS, Xi Y, Rayan JC. Understanding and Mitigating Bias in Imaging Artificial Intelligence. RadioGraphics. 2024;44(5):e230067. [ DOI ] [ PubMed ] [ Google Scholar ] 139. Fassi S, Abdullah A, Fang Y, Natarajan S, Masroor A, Kayali N, et al. Not all AI health tools with regulatory authorization are clinically validated. Nature Medicine. 2024;30. [ DOI ] [ PubMed ] 140. Marka AW, Luitjens J, Gassert FT, Steinhelfer L, Burian E, Rübenthaler J, et al. Artificial intelligence support in MR imaging of incidental renal masses: an early health technology assessment. European Radiology. 2024;34(9):5856-65. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 141. Stogiannos N, Gillan C, Precht H, Reis Csd, Kumar A, O'Regan T, et al. A multidisciplinary team and multiagency approach for AI implementation: A commentary for medical imaging and radiotherapy key stakeholders. Journal of Medical Imaging and Radiation Sciences. 2024;55(4):101717. [ DOI ] [ PubMed ] [ Google Scholar ] 142. Leiner T, Bennink E, Mol CP, Kuijf HJ, Veldhuis WB. Bringing AI to the clinic: blueprint for a vendor-neutral AI deployment infrastructure. Insights into Imaging. 2021;12(1):11. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 143. Pollen A. Healthcare AI | Aidoc Always-on AI [Internet]2019 2019/05/30/T09:47:44+00:00. Available from: https://www.aidoc.com/learn/blog/ris-pacs-ai-radiology-workflow/ . 144. Kim B, Romeijn S, van Buchem M, Mehrizi MHR, Grootjans W. A holistic approach to implementing artificial intelligence in radiology. Insights into Imaging. 2024;15(1):22. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 145. Muralidharan V, Adewale BA, Huang CJ, Nta MT, Ademiju PO, Pathmarajah P, et al. A scoping review of reporting gaps in FDA-approved AI medical devices. NPJ Digit Med. 2024;7(1):273. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 146. Seker E, Greer M. Detecting the Potential for Bias in Healthcare Data. Stud Health Technol Inform. 2025;327:1210-4. [ DOI ] [ PubMed ] [ Google Scholar ] 147. Miller RJH, Singh A, Otaki Y, Tamarappoo BK, Kavanagh P, Parekh T, et al. Mitigating bias in deep learning for diagnosis of coronary artery disease from myocardial perfusion SPECT images. Eur J Nucl Med Mol Imaging. 2023;50(2):387-97. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 148. Abdelwanis M, Alarafati HK, Tammam MMS, Simsekler MCE. Exploring the risks of automation bias in healthcare artificial intelligence applications: A Bowtie analysis. Journal of Safety Science and Resilience. 2024;5(4):460-9. [ Google Scholar ] 149. Skitka LJ, Mosier K, Burdick MD. Accountability and automation bias. International Journal of Human-Computer Studies. 2000;52(4):701-17. [ Google Scholar ] 150. Pal B, Roy A, Kathirvel RP, O'Toole AJ, Chellappa R. DiversiNet: Mitigating Bias in Deep Classification Networks across Sensitive Attributes through Diffusion-Generated Data. 2024 IEEE International Joint Conference on Biometrics (IJCB). 2024:1-10. [ Google Scholar ] 151. Paproki A, Salvado O, Fookes C. Synthetic Data for Deep Learning in Computer Vision & Medical Imaging: A Means to Reduce Data Bias. ACM Comput Surv. 2024;56(11):Article 271. [ Google Scholar ] 152. Kwong JCC, Erdman L, Khondker A, Skreta M, Goldenberg A, McCradden MD, et al. The silent trial - the bridge between bench-to-bedside clinical AI applications. Front Digit Health. 2022;4:929508. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 153. Andersen ES, Birk-Korch JB, Hansen RS, Fly LH, Röttger R, Arcani DMC, et al. Monitoring performance of clinical artificial intelligence in health care: a scoping review. JBI Evid Synth. 2024;22(12):2423-46. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 154. Hernandez RJ, Goñi J. Responsible Design for Sustainable Innovation: Towards an Extended Design Process. Processes. 2020;8(12):1574-. [ Google Scholar ] 155. Williams R. The Social Shaping of Technology (SST). Science, Technology, and Society: Cambridge University Press; 2019. p. 138-62. [ Google Scholar ] 156. Jasanoff S. The Ethics of Invention: Technology and the Human Future. New York: W.W. Norton & Company; 2016. [ Google Scholar ] 157. Jasanoff S. Technologies of humility. Nature. 2007;450(7166):33-. [ DOI ] [ PubMed ] [ Google Scholar ] 158. Macnaghten P, Owen R, Stilgoe J, Wynne B, Azevedo A, de Campos A, et al. Responsible innovation across borders: tensions, paradoxes and possibilities. Journal of Responsible Innovation. 2014;1(2):191-9. [ Google Scholar ] 159. Owen R, Macnaghten P, Stilgoe J. Responsible research and innovation: From science in society to science for society, with society. Science and Public Policy. 2012;39(6):751-60. [ Google Scholar ] 160. Stilgoe J, Owen R, Macnaghten P. Developing a framework for responsible innovation. Research Policy. 2013;42(9):1568-80. [ Google Scholar ] 161. Stahl BC, Akintoye S, Bitsch L, Bringedal B, Eke D, Farisco M, et al. From Responsible Research and Innovation to responsibility by design. Journal of Responsible Innovation. 2021;8(2):175-98. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement Not applicable. Articles from Journal of Imaging Informatics in Medicine are provided here courtesy of Springer ACTIONS View on publisher site PDF (1.7 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Ethical Considerations in Patient Privacy and Data Handling for AI in Cardiovascular Imaging and Radiology.#263325
• Algorithmic Governance of Healthcare Data Privacy: Operationalizing HIPAA and GDPR Compliance Through Cloud-Native Auditability and Cryptographic Enforcement#1177629
• HIPAA-Compliant AI Deployment Patterns in Clinical Settings: Privacy-Preserving Techniques and Governance Controls#1115582
• Secure, privacy-preserving and federated machine learning in medical imaging#964635
• Protecting patients’ medical records under the GDPR#1177644