人工智能在放射学教育和培训中的应用。

全文（开放获取） · 英文原文

【Abstract】 MOTIVATION: Artificial intelligence (AI) is reshaping radiology through advances in machine learning, deep learning, and generative AI. As these technologies become embedded in diagnostic workflows, radiology education must adapt to prepare learners for AI-enabled clinical practice. This review aimed to synthesize current evidence, identify educational priorities, and highlight challenges and opportunities for implementation. INTRODUCTION: The growing adoption of AI in medical imaging requires radiology curricula to extend beyond image interpretation and encompass AI literacy, ethical reasoning, critical appraisal, and human-AI collaboration. Understanding how AI is currently incorporated into radiology education is essential for developing effective training frameworks. This review examined educational approaches, learner outcomes, and implementation challenges associated with AI in radiology education. METHODOLOGY: A scoping review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR), using the Population-Concept-Context(PCC) framework to define the review question and eligibility criteria. Searches were performed in Scopus, PubMed, and IEEE Xplore for studies published between 2020 and 2025. After screening and eligibility assessment, 29 original studies were included. Educational outcomes were classified into skill development, engagement, and performance domains. RESULTS AND DISCUSSION: Skill development was the most frequently investigated outcome (55.2%), followed by performance (27.6%) and engagement (17.2%). Approximately 86% of studies reported positive or improved educational outcomes. AI-based interventions enhanced learner confidence, AI literacy, diagnostic reasoning, and readiness for clinical implementation. Generative AI tools showed promise for tutoring, assessment, and self-directed learning but raised concerns regarding reliability, hallucinations, bias, and ethical use. Common barriers included limited faculty expertise, insufficient formal training, curricular overcrowding, inadequate infrastructure, and governance challenges. CONCLUSION: AI has significant potential to strengthen radiology education through personalized learning, competency development, and technology-enhanced training. Nevertheless, sustainable implementation requires standardized curricula, faculty development, competency frameworks, and robust ethical oversight. Future research should focus on longitudinal outcomes and evidence-based educational models that prepare radiology professionals for increasingly AI-integrated healthcare systems. 【Introduction】 Radiology education increasingly relies on well-designed curricula that balance interpretive expertise with professional competencies like communication (), safety (), and quality improvement (). While traditional training has long utilized apprenticeship models—incorporating supervised readouts (), didactic lectures (), and case-based learning to build clinical judgment—the modern environment introduces new pressures (). Expanding exam types (), accelerated throughput (), and variable case mixes now demand more scalable (), data-driven educational strategies (). These challenges have shifted the focus toward individualized learning trajectories and measurable competency outcomes that can be implemented without increasing faculty workload or compromising patient care (). The maturation of artificial intelligence from a research interest to a clinical and operational staple is fundamentally reshaping radiology workflows (). As machine learning and deep learning algorithms for image reconstruction (), triage, and reporting support move into routine practice (), the requirements for medical training are shifting. Radiologists must now develop a technical and clinical understanding of how these algorithmic outputs are generated (), their specific failure modes, and the nuances of communicating AI-informed findings to patients and colleagues. This shift creates a dual mandate for modern radiology education: curricula must both prepare trainees to work within AI-enabled environments and utilize AI-driven tools to enhance the teaching, assessment, and remediation of radiologic skills (). AI literacy and AI-enabled pedagogy represent two distinct but interconnected educational mandates (). AI literacy focuses on the technical and ethical foundations of machine learning (), including performance metrics, dataset bias, generalizability, and regulatory issues. In clinical radiology (), these competencies are essential for maintaining safety and quality, as misinterpreting algorithmic outputs or failing to recognize model drift can result in diagnostic errors and inequitable patient care. Consequently, AI education must supplement (), rather than replace, core interpretive training by emphasizing clinical reasoning, accountability, and patient-centered communication. Effective curricula prioritize clinically relevant foundations, such as statistics and bias, before addressing practical implementation, workflow integration, and governance (). AI-enabled pedagogy uses artificial intelligence as a primary instructional tool (), leveraging radiology's data-rich environment of image repositories, structured reports, and performance metrics. This data allows for “precision education,” whe ()re AI tailors content, pacing, and feedback to the specific strengths and weaknesses of individual trainees. Systems can automate case selection based on learner proficiency (), generate targeted assessments, provide adaptive hints, and identify systematic interpretive errors—tasks that traditionally demand significant faculty time. The integration of generative AI and large language models further enables interactive tutoring (), draft report feedback, and simulated clinical reasoning exercises. However, the adoption of these tools requires addressing concerns regarding algorithmic reliability (), potential hallucinations, data privacy, and the critical distinction between educational assistance and the substitution of clinical judgment. Implementing AI in radiology education faces significant barriers that mirror the challenges of clinical deployment (). Primary obstacles include a lack of faculty expertise (), limited time for curriculum development, and inconsistent access to technical infrastructure or curated datasets. Furthermore (), uncertainty persists regarding competency standards, data governance, and the mitigation of bias. While baseline AI knowledge varies among trainees and professionals (), there is a consistent demand for structured training. Educational leaders must now address the practicalities of integrating AI into already dense curricula (), preparing faculty for new roles, and moving beyond superficial familiarity toward critical appraisal skills. Given the high-stakes, team-based nature of the field, effective AI education must extend beyond algorithmic theory to encompass human factors (), interdisciplinary collaboration, and real-world workflow integration. The expanding literature on AI in radiology education is fragmented across diverse study designs (), learner populations, and intervention types. While some research evaluates structured AI curricula, other studies focus on AI-driven educational tools such as automated scoring, case selection, and adaptive feedback. Outcome measures are similarly inconsistent (), ranging from subjective learner satisfaction and perceived readiness to objective data on diagnostic accuracy, efficiency, and knowledge retention. Despite promising results, recurrent methodological limitations including small sample sizes, brief follow-up periods, and heterogeneous interventions hinder the generalizability of these findings. Additionally, the rapid evolution of generative AI complicates formal evaluation, as technological capabilities often outpace traditional research cycles and curricular updates. These gaps make it difficult to determine which specific educational components most effectively drive long-term competency. This structured review examines a specific gap in the literature on artificial intelligence in radiology education. Although previous reviews have discussed AI in radiology or medical education more broadly, less attention has been given to how AI has been examined in radiology training in relation to educational approaches, learner groups, and reported educational outcomes. Accordingly, this review aimed to synthesize recent evidence on the integration of AI in radiology education, with emphasis on AI literacy curricula, AI-enabled instructional interventions, learner populations, reported educational outcomes, and the methodological limitations of the available publications. 【Methodology】 This scoping review was reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) to promote transparent and complete reporting of the literature identification, screening, eligibility assessment, data charting, and evidence synthesis processes (). The review methodology was also informed by established guidance for conducting scoping reviews (). Unlike traditional systematic reviews that often aim to determine the efficacy of specific interventions, this scoping review was designed to map the current state of knowledge, identify research gaps, and clarify concepts regarding the integration of AI into radiology education. The systematic literature search was executed across three major interdisciplinary databases: PubMed, Scopus, and IEEE Xplore. The temporal scope was defined as the five years from 2020 to 2025, a timeframe that encompasses the rapid acceleration of AI integration in medical training following the emergence of advanced machine learning and generative AI architectures. The search strategy employed a combination of Boolean operators and controlled vocabulary centered on three pillars: AI-related technologies, the clinical domain of radiology, and the educational environment. To ensure a comprehensive analysis, this review synthesized evidence across diverse learner demographics ranging from medical students and radiology residents to established professionals and educators and various study designs, including empirical research, curriculum proposals, and position statements. By integrating findings from these disparate sources, the study aimed to provide a structured taxonomy of current pedagogical strategies. The analytical approach focused on categorizing the literature based on reported educational outcomes, such as AI literacy, learner engagement, diagnostic reasoning, and technical competency. This synthesis highlights how technology-supported instructional approaches are being leveraged to supplement conventional clinical training. Furthermore, this review identifies areas where the evidence remains limited, particularly concerning the long-term impact of AI literacy on clinical performance and the lack of standardized global curricula. By mapping these trends, this study provides a foundational resource for future research, informing the development of evidence-based educational frameworks that prepare the next generation of radiologists to navigate an increasingly automated clinical landscape. This methodological approach ensures that the findings are not only representative of the current academic discourse but also actionable for institutional curriculum designers and medical educators. 【Design and application of the PCC framework】 This scoping review was reported in accordance with the PRISMA-ScR guideline and conducted using established methodological guidance for scoping reviews (, ). To delineate the scope of the investigation and establish clear eligibility criteria, the Population–Concept–Context (PCC) framework was applied (). To delineate the scope of the investigation and establish precise inclusion criteria, the PCC framework was utilized (). While the PICO framework is the conventional standard for systematic reviews aiming to quantify the comparative effectiveness of clinical interventions (, ), it is often overly restrictive for exploring the breadth of educational literature. The PCC framework was selected for this study to enable a broader, more descriptive mapping of the research landscape, allowing for the synthesis of diverse educational models, theoretical frameworks, and qualitative observations that might otherwise be excluded (). The specific application of this framework is detailed in , which outlines the core elements used to structure the research question and define inclusion criteria. PCC framework defining the core elements used to structure the research question and inclusion criteria for the systematic review on artificial intelligence in radiology education. To ensure consistency in the synthesis of results, the diverse educational outcomes reported across the selected literature were aggregated into three mutually exclusive thematic categories: skill development, engagement, and performance. Each included study was assigned exclusively to one category according to its primary reported educational focus. The engagement category captures research centered on the affective domain, including learner attitudes, perceptions, professional confidence, interest in AI-related tracks, and overall willingness to adopt novel technological tools in clinical practice. The performance category accounts for studies evaluating applied or objective outcomes, such as gains in diagnostic reasoning, measurable improvements in image interpretation, or the refinement of reporting skills during simulated or real-world tasks. Each publication was assigned to the category that best reflected its primary reported focus through a collaborative, multi-reviewer consensus process. This thematic clustering allows the review to move beyond descriptive summary, providing a structured overview of where the current discourse is positioned. By organizing the findings in this manner, the review identifies whether current research prioritizes foundational AI literacy, the psychological integration of the learner, or the tangible enhancement of clinical proficiency, thereby highlighting gaps where pedagogical innovation may be required. Based on this framework, the review question was defined as follows: “How has artificial intelligence been implemented and evaluated in radiology education to support learning, assessment, and competency development among radiology learners?” A structured Boolean search strategy was then applied to retrieve studies located at the explicit intersection of artificial intelligence, radiology, and education. Search terms were selected to identify studies involving radiology learners, AI-related educational interventions, and teaching or training contexts relevant to the review question. The search strategy was intended to identify studies aligned with the predefined eligibility criteria for the period 2020 to 2025. The search formulation used in the selected databases is summarized in . Search strategy and number of records retrieved by database for publications on artificial intelligence in radiology education (2020–2025). The complete database-specific search strategies, including the exact search strings, field tags, Boolean operators, filters, and syntax adaptations used for Scopus, PubMed, and IEEE Xplore, are provided in to facilitate the reproducibility of the search process. 【Database selection and search strategy】 This review searched Scopus, PubMed, and IEEE Xplore to identify publications explicitly indexed at the intersection of artificial intelligence, radiology, and education. These databases were selected to cover the biomedical, educational, and technical dimensions of the topic. PubMed was used to capture literature in radiology and medical education. Scopus was included for its broad interdisciplinary coverage across health sciences, education, and applied research. IEEE Xplore was included to identify technically oriented publications with potential educational relevance. The search was limited to peer-reviewed publications published between 2020 and 2025. This predefined time window was selected to capture the recent phase of AI adoption in radiology education, including studies addressing AI literacy, educational applications of machine learning, and emerging teaching-oriented uses of generative AI in imaging-related contexts. The final search was conducted on 7 March 2026. The Boolean search strategy was intentionally specific and was designed to prioritize publications in which the educational focus was explicit in the indexed record. The core search structure combined terms related to artificial intelligence, radiology, and education. This approach was selected to reduce retrieval of technically focused studies in medical imaging that did not include a teaching, training, or curriculum-related objective. The search formulation applied in each database is summarized in . The same core search structure was applied across databases, with syntax adapted where required by each platform. This focused strategy was intended to identify publications explicitly indexed at the intersection of artificial intelligence, radiology, and education, rather than all AI-related radiology publications with indirect educational relevance, the selection of specific databases was driven by their unique contributions to the research scope: Scopus was selected for its robust interdisciplinary coverage, encompassing health sciences, educational theory, and broader applied research. PubMed served as a foundational resource for retrieving highly specific biomedical, radiological, and medical education literature. IEEE Xplore was included to ensure the capture of highly technical and AI-oriented educational applications that might be less prominent in clinical-only databases. Because the literature on artificial intelligence in healthcare and medical imaging is broad, this strategy prioritized publications with a clear pedagogical, curricular, or learner-related focus and reduced inclusion of AI radiology studies without an educational objective. This search design should be interpreted as a focused strategy for a structured educational synthesis, not as an exhaustive search of all AI-related radiology literature. The volume of records retrieved demonstrated a consistent upward trend across all three databases throughout the study period, although the specific growth patterns varied by source. Scopus yielded the highest number of records, showing a substantial increase from 91 publications in 2020 to 549 by 2025. During this same timeframe, records from IEEE Xplore grew from 40 to 110, while PubMed citations rose from 4 to 28. These results, which characterize the annual distribution of records retrieved by the search strategy, are visually summarized in . Annual number of publications indexed in Scopus, PubMed, and IEEE Xplore from 2020 to 2025 related to artificial intelligence in radiology education. 【Identification】 The study-selection process was documented in accordance with PRISMA-ScR reporting recommendations. summarizes the stages of record identification, screening, eligibility assessment, and final inclusion. PRISMA 2020 flow diagram illustrating the identification, screening, eligibility assessment, and inclusion of publications in this scoping review. The identification phase began with a comprehensive search across three databases, retrieving a total of 2,095 records: 1,603 from Scopus, 103 from PubMed, and 389 from IEEE Xplore. From this initial articles, 1,221 records were excluded: 567 were duplicates, 395 were identified as books, chapters, or conference papers, and 259 were excluded because their keywords did not specifically address the intersection of AI and radiology education. This resulted in 877 records proceeding to the screening phase. During screening, 834 records were excluded based on abstract assessment, specifically for failing to address core thematic areas: “Radiology—Education” (n = 295), “Radiology—Training” (n = 295), “Radiology—AI” (n = 128), “AI—Education” (n = 39), “AI—Training” (n = 183), and the exclusion of meta-analyses (n = 189). Following this stage, 43 records were assessed for eligibility. Finally, 14 review articles were excluded, resulting in 29 studies being included in the final review analysis. 【Screening】 Initially, the identification phase yielded 2,095 total records. From this stage, 1,221 re