Evaluating AI-powered instruction on Computed Tomography Angiography (CTA) teaching outcomes: a systematic review and meta-analysis Authors: Liang Yuan, Xingyi Xie, Yiqian Wen Published: 2026 ABSTRACT AbstractObjectiveTo systematically evaluate the effectiveness of AI-assisted teaching in computed tomography angiography (CTA) training, focusing on theoretical knowledge, practical skills, image judgment time, and teaching satisfaction.MethodsEight databases were searched for randomized controlled trials (RCTs) from 1990 to 2025. Two reviewers independently screened studies, extracted data, and assessed risk of bias using the Cochrane RoB 2.0 tool. Frequentist and Bayesian meta-analyses were performed, along with subgroup analyses by CTA anatomical type, learner group, and sample size.ResultsSeven RCTs with 515 participants were included. All studies were conducted in China with Chinese-language AI interaction. Frequentist analyses showed that AI significantly improved practical skills and reduced image judgment time. Benefits were also observed for theoretical scores and teaching satisfaction. However, Bayesian analyses revealed statistical uncertainty for theoretical scores and teaching satisfaction due to high heterogeneity. Subgroup analyses identified CTA type, learner background, and sample size as key sources of heterogeneity. All included studies had high or some concerns for risk of bias.ConclusionAI-assisted teaching effectively improves practical skills and shortens image judgment time in CTA training, offering clear educational benefits for radiology trainees. However, evidence for theoretical scores and teaching satisfaction remains uncertain due to high heterogeneity and methodological limitations. The findings support targeted integration of AI into CTA curricula while highlighting the need for large-scale, standardized RCTs to confirm long-term efficacy and generalizability.Keywords: Teaching, Medical teaching, Artificial intelligence, Computed Tomography Angiography (CTA), Radiology Education, Randomized Controlled Trial (RCT) Introduction Radiology education is a core component of undergraduate and postgraduate medical training, with computed tomography angiography (CTA) emerging as a critical skill for clinical decision-making across specialties. However, traditional radiology teaching faces persistent challenges, including limited teaching time from radiologists, inadequate pre-clinical training, and uneven quality of instruction across institutions. A Dutch qualitative study (44 clinicians) confirmed anatomy is foundational to radiology, but current curricula separate them; students also lack knowledge of imaging indications [1]. A national survey of Greek radiology residents further revealed that only 14% were satisfied with their training program’s overall quality, with 45% lacking access to structured curricula and 95% deprived of protected academic time [2]. These gaps often lead to inconsistent learning outcomes, particularly in CTA interpretation which requires mastery of complex anatomical relationships and imaging nuances. Against this backdrop, artificial intelligence (AI) has emerged as a transformative force in medical education, with growing evidence supporting its ability to deliver personalized, standardized, and efficient training [3]. Radiology postgraduates globally have shown familiarity with large language models (LLMs) such as ChatGPT and BingAI, with 71% open to integrating them into learning and 59% using them when traditional resources fall short [4]. Prior studies have shown AI’s efficacy in radiology education, such as enhancing diagnostic accuracy through simulated case training and improving assessment consistency via automated scoring tools [5, 6]. Specifically, AI-driven tools have addressed key pain points in radiology training, including limited exposure to rare cases and variable feedback quality [3, 7], In addition, generative AI models can convert radiology report revisions into actionable educational feedback, with 84.5% of revisions rated as relevant by expert radiologists [8]. Specifically, CTA training targeting complex vascular anatomy and subtle imaging features requires professional imaging visual AI and multimodal diagnostic AI, rather than universal general large language models, to fully adapt to the unique professional characteristics of CTA interpretation. Despite these advancements, the effectiveness of AI-powered instruction in CTA teaching remains incompletely synthesized. Moreover, the potential risks of AI education, including overreliance, reduced critical thinking, algorithmic bias, and ethical issues, have rarely been addressed in prior reviews [9–11]. CTA requires unique training components such as vascular anatomy visualization, contrast enhancement analysis, and clinical–imaging correlation, all of which demand targeted educational strategies. However, no systematic review or meta-analysis has focused specifically on AI-assisted CTA teaching. A recent scoping review of AI in radiology education documented diverse applications but emphasized the lack of standardized curricula and long-term outcome data [12], Meanwhile, studies evaluating LLMs for CTA reporting in preoperative planning have shown promising but highly variable performance across models [13]. Recent evidence has further demonstrated that large language models achieve excellent performance in coronary CTA reporting based on CAD-RADS 2.0 guidelines and can effectively enhance the diagnostic performance of general radiologists in cardiac imaging cases, supporting the rational integration of AI tools into both clinical CTA interpretation and targeted educational training [14, 15]. Distinctively, this study differs uniquely from prior general radiology AI education research, as it innovatively covers four comprehensive outcome dimensions and focuses on radiology or clinical medicine residents and medical students who received CTA courses, forming exclusive evidence adapted to CTA-specific teaching characteristics. Therefore, this systematic review and meta-analysis was conducted to fill these gaps. The primary objectives were to comprehensively evaluate the benefits and limitations of AI-assisted CTA teaching, explore sources of between-study heterogeneity, assess the certainty of evidence using the GRADE approach, and provide balanced, evidence-based recommendations for clinical radiology teaching. Literature review The role and value of artificial intelligence in medical education Artificial intelligence (AI) has reshaped modern education by enabling personalized learning, adaptive feedback, and scalable access to high-quality resources [9]. In healthcare, AI-powered instruction addresses critical limitations of traditional training, including uneven teaching quality, heavy instructor workload, insufficient hands-on practice, and limited exposure to complex clinical cases [16, 17]. In radiology and imaging education, AI tools enhance diagnostic training through interactive visualization, automated image annotation, simulated clinical reasoning, and standardized feedback—functions especially critical for complex modalities such as CTA [18, 19]. Large language models (LLMs) further strengthen educational value by supporting case discussion, explanation generation, and structured teaching feedback [5, 17], aligning with the high willingness of radiology trainees to adopt AI into regular learning [20]. Risks and challenges of AI in education Despite well-documented benefits, excessive or unregulated use of AI in education carries meaningful risks that remain underrecognized in radiology and CTA-specific training. Recent cross-disciplinary educational research has confirmed robust relationships among AI literacy, AI trust, and AI dependency [10, 11]. Higher AI literacy and trust are consistently associated with increased AI dependency. This may reduce learners’ critical thinking, independent problem-solving abilities, and autonomous clinical reasoning [10, 11]. For CTA training, where independent vascular interpretation, artifact recognition, and clinical reasoning are core competencies, overreliance on AI could gradually erode foundational diagnostic skills rather than enhance them. Additional concerns include algorithmic bias, limited model transparency, data privacy risks, and ethical challenges in automated clinical decision-making [9, 17]. Research gaps in AI-assisted CTA teaching Although existing literature supports the potential of AI in radiology education, critical knowledge gaps persist specifically for CTA training: ⑴ No systematic review or meta-analysis has rigorously quantified the effectiveness of AI‑powered instruction for CTA teaching outcomes. ⑵ Potential harms, including AI dependency and reduced critical thinking, are rarely addressed or empirically evaluated in prior radiology education reviews. ⑶ Heterogeneity sources related to anatomical CTA regions, learner proficiency, AI tool categories, and intervention durations remain inadequately explored. ⑷ Formal assessment of evidence quality using structured tools such as GRADE remains rare, limiting the clinical and educational interpretability of the findings. This systematic review and meta-analysis directly address these gaps by providing a balanced synthesis of benefits and risks, quantitatively evaluating teaching efficacy, exploring heterogeneity, and rating evidence certainty. The findings support safe, responsible, and evidence‑based integration of AI into CTA training curricula. Methods The protocol for this study has been formally registered in the International Prospective Register of Systematic Reviews (PROSPERO), with the registration ID: CRD42024497935. Search strategy A comprehensive search was conducted across eight databases, including four English databases (PubMed, Embase, Cochrane Library, Web of Science Core Collection) and four Chinese databases (China National Knowledge Infrastructure [CNKI], VIP Chinese Scientific Journals Database [VIP], Wanfang Medical Database, SinoMed). The search period was restricted to articles published from January 1, 1990, to October 22, 2025, to ensure both comprehensiveness and currency of the data. Core search terms covered three key dimensions: ⑴ AI-powered educational interventions, including “Artificial Intelligence”, “AI”, “machine learning”, “deep learning”, “large language model”, “LLM”, “intelligent assisted teaching”; ⑵ discipline-specific content, including “CTA”, “CT angiography”, “coronary CTA”, “head and neck CTA”, “computed tomography angiography”, “radiology education”; ⑶ study design elements, including “Randomized Controlled Trial”, “RCT”, “control group”, “intervention study”, “comparative study”. For English databases, MeSH terms were combined with free-text words, including “Artificial Intelligence”[Mesh], “Tomography, X-Ray Computed”[Mesh], “Angiography”[Mesh], and “Teaching”[Mesh]. Detailed database-specific search strategies (combining subject terms, free text, and field-specific restrictions) are provided in Appendix 1. Study screening This systematic review strictly follows the standardized PICOS framework to standardize literature screening specifications. ⑴ Population: Radiology or clinical medicine residents and medical students who completed formal CTA specialized teaching courses. ⑵ Intervention: Practical AI-assisted CTA teaching applications, including AI intelligent CTA vascular image interpretation, simulated case scenario training, automatic standardized skill assessment and real-time intelligent teaching feedback. ⑶ Comparator: Traditional manual CTA teaching mode, which adopts teacher offline lectures, conventional film reading guidance and routine clinical case teaching, without any application of AI auxiliary teaching tools. ⑷ Outcomes: Theoretical knowledge assessment scores, professional operational skill scores, CTA image diagnosis judgment time, and multi-dimensional teaching satisfaction evaluation. ⑸ Study design: Prospective randomized controlled trials (RCTs). Inclusion criteria ⑴ Only randomized controlled trials (RCTs) were included. ⑵ The study population consisted of radiology or clinical medicine residents or medical students who received CTA courses. ⑶ The evaluation system outcomes included theoretical scores, skills scores, image judgment time, and teaching satisfaction. Importantly, Studies that met at least one of the above outcome indicators were eligible for inclusion, and simultaneous reporting of all four dimensions was not mandatory for trial enrollment. ⑷ Only peer-reviewed journal articles in Chinese or English were included. Grey literature such as postgraduate theses and dissertations was excluded, owing to non-uniform data reporting, absent rigorous peer review, and variable methodological quality, which fails to comply with rigorous quantitative evidence synthesis standards for meta-analysis. Exclusion criteria ⑴ Non-RCTs or studies lacking key outcome measures were excluded. ⑵ Studies not involving AI-assisted CTA teaching interventions were excluded. ⑶ Studies with major methodological deficiencies in RCT teaching trials were excluded, including improper randomization, absence of blinding, incomplete outcome reporting, and unbalanced baseline characteristics, as these issues substantially compromise the internal validity and statistical reliability of study findings. ⑷ Articles with non-continuous theoretical scores, skills scores, image judgment time (continuous variables), and simple yes/no dichotomous teaching satisfaction outcomes were excluded, as they cannot meet the standardized statistical analysis requirements of meta-analysis. Study selection Two independent researchers (Xingyi Xie, Yiqian Wen) conducted study selection in accordance with PRISMA guidelines. First, duplicated records were removed. Second, titles and abstracts were screened to exclude irrelevant studies; potentially eligible studies were retrieved for full-text review against the predefined inclusion/exclusion criteria. Discrepancies were resolved via discussion, with consultation of a third senior researcher (Liang Yuan) if consensus failed. Published and unpublished studies were sought, and additional studies were identified through reference list checking and forward citation searching. Data extraction The same two researchers independently extracted data using a pre-designed structured form. Extracted information included: first author, publication year, sample size allocated to the AI-powered CTA teaching group versus the traditional teaching group, type of AI-powered teaching intervention, specific details of traditional teaching methods, key outcome indicators (theoretical scores, skill scores, image judgment time, and teaching satisfaction), as well as raw or extractable summary data required for meta-analysis. The risk of bias in included studies was independently assessed by the two researchers using the Cochrane Risk of Bias 2.0 (RoB 2.0) tool, with discrepancies resolved following the same procedure as that adopted for study selection. Statistical analysis All meta-analyses were performed in R (Version 4.5.1) using R Studio. For frequentist meta-analyses, the meta and metafor packages were utilized: standardized mean differences (SMDs) with 95% confidence intervals (CIs) were estimated for continuous outcomes (theoretical scores, skill scores, image judgment time), while risk ratios (RRs) with 95% CIs were calculated for the binary outcome of teaching satisfaction to quantify group differences. For continuous outcomes, Hedges’ g (bias-corrected standardized mean difference, SMD) was used as the effect size measure to account for potential small-sample bias in the included studies. The meta package was additionally used for heterogeneity visualization and preliminary effect synthesis. Heterogeneity across studies was assessed using the chi-square test and I² statistic: a fixed-effects model was applied if I² ≤ 50% (indicating low-to-moderate heterogeneity), whereas a random-effects model was used if I² > 50% (indicating high heterogeneity). For the outcome with high heterogeneity and the largest number of included studies, predefined subgroup analyses were conducted to explore potential sources of variability, focusing on three dimensions: CTA anatomical type (Head-Neck CTA, Comprehensive CTA, Coronary CTA), learner group (Residents, Students), and sample size (small sample size (n < 40), large sample size (n ≥ 40)). Subgroup analyses were performed using the meta package, with SMDs and 95% CIs reported for each subgroup. Differences between subgroups were evaluated using the chi-square interaction test (with corresponding degrees of freedom [df] and P-values) to determine whether these subgroup factors could explain the observed heterogeneity. Subgroup analyses were not performed for outcomes with a limited number of included studies. For Bayesian meta-analyses, the brms and tidybayes packages were employed. Parameter estimation was conducted via Markov Chain Monte Carlo (MCMC) simulations, with 60,000 total iterations (including 30,000 warm-up iterations) to ensure stable convergence of posterior distributions. Convergence was verified using R-hat values (R-hat < 1.05 indicating adequate convergence), and effect sizes were reported as posterior medians with 95% credible intervals (CrIs). Statistical significance was set at α = 0.05: for SMDs (theoretical scores, skill scores, image judgment time), significance was defined as 95% CIs (frequentist) or 95% CrIs (Bayesian) excluding zero; for RRs (teaching satisfaction), significance was defined as 95% CIs or 95% CrIs excluding 1. Drapery plots (generated using the meta package) were used to visually assess publication bias, characterize heterogeneity, and interpret the distribution of effect sizes, complementing quantitative heterogeneity evaluations. These plots display individual study effect sizes with 95% CIs, weight data points by statistical precision, and integrate multi-dimensional data to facilitate robust interpretation of teaching outcomes and detection of potential bias. Results A total of 81 records were identified via systematic searches of eight databases (PubMed, Web of Science, Cochrane Library, Embase, CNKI, VIP, Wanfang, SinoMed) in accordance with the PRISMA guidelines. After removing 22 duplicate records, 59 records underwent title and abstract screening. Among these, 42 were excluded due to irrelevant teaching content, 6 were excluded for non-randomized controlled trial (non-RCT) designs, and 4 were excluded for lack of key outcome measures. Finally, 7 full-text articles were retrieved for detailed eligibility assessment. As shown in Fig. 1 (PRISMA flow diagram), all 7 studies underwent full-text review and were included in the qualitative and quantitative syntheses. No articles were excluded during full-text evaluation. PRISMA flow diagram of study selection for AI-assisted CTA teaching interventions As summarized in Table 1, these 7 studies (published between 2021 and 2025) enrolled a total of 515 participants (258 in the AI-assisted teaching group and 257 in the control group), with sample sizes ranging from 13 to 120 [21–27]. All studies were conducted in China, and the AI interaction language used in all included studies was Chinese without any English or multilingual versions. The study populations included radiology residents, clinical medicine students, and medical imaging technology students. All included studies compared AI-assisted CTA teaching with traditional teaching approaches, and the primary outcome measures included theoretical examination scores, practical skills scores, imaging judgment time, and teaching satisfaction. Table 1Baseline characteristics of including studyAuthor / YearStudy DesignCountryTeaching LearnersSample Size (Intervention vs. Control)Age (years, mean ± SD ) (Intervention vs. Control)Outcome IndicatorsIntervention (with subgroup)Control teachingAI interaction languageDeng Gang 2025 [22]RCTChinaNeurosurgery residents32 vs. 3226.88 ± 1.24 vs. 26.72 ± 1.44Theoretical scores; Skill scores; Teaching satisfactionAI-assisted system (Head-Neck CTA)Traditional teachingChineseChen Suyue 2025 [21]RCTChinaResidents (Radiology, Neurology, Surgery)25 vs. 28NA vs. NASkill scores; image judgment timeAI-assisted system (Head-Neck CTA)Traditional teachingChineseZhang Jing 2024 [27]RCTChinaRadiology residents16 vs. 1728.31 ± 1.40 vs. 27.41 ± 2.09Theoretical scores; Skill scores; Teaching satisfactionAI + film rendering technology (Comprehensive CTA)Traditional teachingChineseWang DongXu 2024 [26]RCTChinaClinical medical students41 vs. 4321.73 ± 1.41 vs. 22.35 ± 1.65Skill scoresAI-assisted image diagnosis software (Comprehensive CTA)Traditional teachingChineseFeng Qi 2023 [24]RCTChinaRadiology residents7 vs. 627.14 ± 2.27 vs. 26.33 ± 1.37Skill scoresAI-assisted training (Coronary CTA)Traditional teachingChineseKang Shaolei 2023 [25]RCTChinaMedical imaging students60 vs. 6023.21 ± 1.34 vs. 23.58 ± 1.18Theoretical scores; Teaching satisfactionAI-assisted system (Coronary CTA)Traditional teachingChineseFan Wei 2021 [23]RCTChinaMedical imaging students30 vs. 3022.79 ± 3.06 vs. 23.03 ± 2.15Theoretical scores; Skill scores; image judgment time; Teaching satisfactionAI-assisted system (Coronary CTA)Traditional teachingChinese Baseline characteristics of including study The risk of bias was independently assessed by two reviewers using the Cochrane RoB 2.0 tool, with excellent inter-rater agreement confirmed (Kappa = 0.85). As presented in Table 2, two studies (Chen Suyue 2025 [21], Wang DongXu 2024 [26]) were classified as having a high risk of bias, primarily due to quasi-random sequence generation, inadequate allocation concealment, and lack of blinding for performance bias. The remaining five studies (Deng Gang 2025 [22], Zhang Jing 2024 [27], Feng Qi 2023 [24], Kang Shaolei 2023 [25], Fan Wei 2021 [23]) were rated as having some concerns, mainly due to unclear allocation concealment or detection bias. All studies adopted Chinese as the unified AI interaction language, and no regional or multilingual heterogeneity was found. Table 2Risk of bias assessment of included studies (Cochrane Risk of Bias 2.0 tool)Study (Author/Year)Random sequence generation (selection bias)Allocation concealment (selection bias)Blinding of participants and personnel (performance bias)Blinding of outcome assessment (detection bias)Incomplete outcome data (attrition bias)Selective reporting (reporting bias)Other sources of biasOverall risk of biasDeng G. 2025 [22]lowunclearhighunclearlowlowunclearsome concernsChen S.Y. 2025 [21]highhighhighunclearlowlowunclearhighZhang J. 2024 [27]lowunclearhighlowlowlowunclearsome concernsWang D.X. 2024 [26]highhighhighunclearlowlowunclearhighFeng Q. 2023 [24]unclearunclearhighunclearlowlowunclearsome concernsKang S.L. 2023 [25]unclearunclearhighunclearlowlowunclearsome concernsFan W. 2021 [23]unclearunclearhighunclearlowlowunclearsome concerns Risk of bias assessment of included studies (Cochrane Risk of Bias 2.0 tool) Notably, none of the included RCTs achieved an overall low risk of bias. Most studies consistently presented high performance bias and unclear allocation concealment, which constituted the main methodological limitations of the current evidence. These limitations may reduce the strength of the conclusions of this meta-analysis. Outcome synthesis Theoretical scores Four studies (n = 317) were meta-analyzed for theoretical scores (Fig. 2A). Substantial heterogeneity (I² = 91%, P < 0.0001) warranted a random-effects model. Frequentist analysis showed a pooled standardized mean difference (SMD) of 1.15 (95% CI [0.39, 1.91], P = 0.0030), with the 95% CI excluding 0, indicating that AI-powered teaching significantly improved theoretical performance versus traditional methods. Extreme heterogeneity (I² = 91%) was observed across the theoretical scores outcome. Forest plots of frequentist meta-analyses evaluating the effect of AI-assisted CTA teaching on (A) theoretical scores, (B) practical skill scores, (C) image judgment time, and (D) teaching satisfaction Bayesian analysis (Fig. 3A) demonstrated excellent convergence (R-hat values for overall pooled effect: 1.00019; between-study heterogeneity: 1.00005; all others < 1.05), confirming stable MCMC sampling. Individual study posterior median effect sizes ranged from 0.41 to 1.87, with one study (Wang DongXu 2024) showing a 95% credible interval (CrI) that included 0. The Bayesian pooled posterior median SMD was 0.96 (95% CrI [-0.01, 1.73]), where the 95% CrI slightly crossed the null value—contrasting with the frequentist 95% CI, which fully excluded 0. The 95% CrI for theoretical scores crossed zero. Wide variability in CrI ranges across studies reflected substantial between-study heterogeneity, consistent with an I² value of 91%. Forest plots of Bayesian meta-analyses evaluating the effect of AI-assisted CTA teaching on (A) theoretical scores, (B) practical skill scores, (C) image judgment time, and (D) teaching satisfaction Accordingly, the evidence supporting a beneficial effect of AI teaching on theoretical scores remains statistically uncertain under Bayesian modeling. Skill scores Six studies (n = 307) were analyzed for skill scores (Fig. 2B). High heterogeneity (I² = 82%, P < 0.0001) prompted use of a random-effects model. Frequentist analysis yielded a pooled SMD of 1.24 (95% CI [0.53, 1.94], P = 0.0006), with the 95% CI excluding 0, indicating that AI-assisted instruction significantly enhanced practical skill proficiency. High heterogeneity (I² = 82%) was observed across the skill scores outcome. Bayesian analysis (Fig. 3B) confirmed excellent convergence (R-hat values: overall pooled effect: 1.00016; between-study heterogeneity: 1.00018; all others < 1.05), with stable MCMC sampling. Individual study posterior median effect sizes spanned from 0.79 to 2.61, with all 95% CrIs excluding 0. The Bayesian pooled posterior median SMD was 1.07 (95% CrI [0.27, 1.75]), with the 95% CrI excluding 0, consistent with frequentist findings. Heterogeneous distribution of effect sizes aligned with I² = 82%. Image judgment time Two studies (n = 113) were included for image judgment time (Fig. 2C). Moderate heterogeneity (I² = 41%, P = 0.1914) justified a fixed-effects model. Frequentist analysis showed a pooled SMD of -2.70 (95% CI [-3.22, -2.18], P < 0.0001), with the 95% CI excluding 0, indicating that AI-powered teaching significantly reduced image judgment time, reflecting higher diagnostic efficiency. Bayesian analysis (Fig. 3C) used a fixed-effects model (consistent with I² = 41%) and achieved excellent convergence (R-hat for overall pooled effect: 1.00018; all others < 1.05). Posterior median effect sizes were − 3.12 and − 2.42, with both 95% CrIs excluding 0. The Bayesian pooled posterior median SMD was − 2.53 (95% CrI [-3.03, -2.04]), with the 95% CrI excluding 0, consistent with frequentist results. Relatively low between-study heterogeneity (I² = 41%) was reflected in clustered CrI ranges, confirming robust findings. Teaching satisfaction Three studies (n = 213) were meta-analyzed for teaching satisfaction using risk ratio (RR; Fig. 2D). High heterogeneity (I² = 87%, P = 0.0005) necessitated a random-effects model. Frequentist analysis yielded a pooled RR of 3.56 (95% CI [1.18, 10.71], P = 0.0241), with the 95% CI excluding 1, indicating significantly higher satisfaction in the AI group versus the traditional group. Extreme heterogeneity (I² = 87%) was observed across the teaching satisfaction outcome. Bayesian analysis (Fig. 3D) confirmed excellent convergence (R-hat values: overall pooled effect: 1.00013; between-study heterogeneity: 1.00029; all others < 1.05). Individual study posterior median RRs ranged from 1.88 to 5.94, with one study (Fan Wei 2021) showing a 95% CrI whose lower bound approached 1. The Bayesian pooled posterior median RR was 2.82 (95% CrI [0.81, 6.18]), where the 95% CrI included 1—contrasting with the frequentist 95% CI, which fully excluded 1. Notable between-study heterogeneity (I² = 87%) was reflected in wide CrI ranges across studies. Thus, the evidence for improved teaching satisfaction with AI teaching remains statistically uncertain based on Bayesian modeling. Publication bias and heterogeneity Funnel plots were constructed for visual inspection of potential publication bias across all four main outcomes and are provided in the supplementary materials (Figure S5). Because the number of included studies was small (2–6 studies per outcome), formal statistical tests for funnel plot asymmetry were not sufficiently reliable. Therefore, publication bias cannot be reliably excluded. Drapery plots (Fig. 4) were employed to visually evaluate publication bias and heterogeneity for all outcomes: (1) For theoretical scores (Fig. 4A), effect sizes and confidence intervals exhibited substantial variability, consistent with high heterogeneity (I2 = 91%), with no typical characteristics of small-study publication bias observed; (2) for skill scores (Fig. 4B), effect sizes were widely dispersed, reflecting high heterogeneity (I2 = 82%), and no clear signals of publication bias were detected; (3) for image judgment time (Fig. 4C), effect sizes were relatively clustered, consistent with moderate heterogeneity (I2 = 41%), with no extreme outliers or asymmetric distributions noted, and no obvious publication bias suggested; (4) for teaching satisfaction (Fig. 4D), risk ratios were heterogeneously distributed, consistent with high heterogeneity (I2 = 87%), with no distinct characteristics of publication bias shown. Drapery plots for publication bias and heterogeneity assessment for (A) theoretical scores, (B) practical skill scores, (C) image judgment time, and (D) teaching satisfaction Subgroup analyses To explore sources of heterogeneity in skill scores and address the potential influence of small-sample studies on pooled estimates, subgroup analyses were conducted, with results presented in Table 3, focusing on three dimensions: (1) CTA anatomical type (Head-Neck CTA, Comprehensive CTA, Coronary CTA); (2) learner group (Residents, Students); (3) total sample size (small sample size: n < 40; large sample size: n ≥ 40). Table 3Subgroup analysis of skill scoresSubgroupNo. of StudiesSample Size (Intervention vs. Control)SMD (95% CI)Heterogeneity TestOverall Effect TestTest for Subgroup DiffferencesSubgroup = CTA Typeχ2 = 1.62; df = 2 (P = 0.4443) Head-Neck CTA257 vs. 602.07 [1.11, 4.03]Tau2=1.877; χ2 = 16.31;df = 1 (P < 0.0001); I2 = 94%Z = 2.07; P = 0.0386 Comprehensive CTA257 vs. 600.77 [0.40, 1.15]Tau2=0; χ2 = 0; df = 1 (P = 0.9608); I2 = 0%Z = 4.02; P < 0.0001 Coronary CTA237 vs. 360.80 [0.32, 1.28]Tau2=0; χ2 = 0.26; df = 1 (P = 0.6098); I2 = 0%Z = 3.27; P = 0.0011 Total6151 vs. 1561.23 [0.61, 1.84]Tau2=0.652; χ2 = 27.53; df = 5 (P < 0.0001); I2 = 82%Z = 3.93; P < 0.0001Subgroup = Learner Typeχ2 = 1.92; df = 1 (P = 0.1656) Residents480 vs. 831.51 [0.50, 2.53]Tau2=0.891; χ2 = 20.95; df = 3 (P = 0.0001); I2 = 86%Z = 2.93; P = 0.0034 Students271 vs. 730.76 [0.42, 1.10]Tau2=0; χ2 = 0; df = 1 (P = 0.9586); I2 = 0%Z = 4.39; P < 0.0001 Total6151 vs. 1561.23 [0.61, 1.84]Tau2=0.453; χ2 = 27.48; df = 5 (P < 0.000)1; I2 = 82%Z = 3.93; P < 0.0001Subgroup = Sample Sizeχ2 = 0.90; df = 1 (P = 0.3419) Small sample size (n < 40)223 vs. 230.87 [0.25, 1.48]Tau2=0; χ2 = 0.18; df = 1 (P = 0.672); I2 = 0%Z = 2.77; P = 0.0056 Large sample size (n ≥ 40)4128 vs.1331.37 [0.54, 2.20]Tau2=0.628; χ2 = 26.83; df = 3 (P < 0.0001); I2 = 89%Z = 3.23; P = 0.0013 Total6151 vs. 1561.23 [0.61, 1.84]Tau2=0.652; χ2 = 27.53; df = 5 (P < 0.0001); I2 = 82%Z = 3.93; P < 0.0001 Subgroup analysis of skill scores CTA anatomical type (1) For Head-Neck CTA (2 studies, n = 117), AI-assisted teaching significantly improved skills (pooled SMD = 2.07, 95% CI [1.11, 4.03], P = 0.0386) with high heterogeneity (I²=94%, P < 0.0001); (2) for Comprehensive CTA (2 studies, n = 117), the pooled SMD was 0.77 (95% CI [0.40, 1.15], P < 0.0001) with no significant heterogeneity (I²=0%); (3) for Coronary CTA (2 studies, n = 73), the pooled SMD was 0.80 (95% CI [0.32, 1.28], P = 0.0011) with no significant heterogeneity (I²=0%), indicating significant coronary-specific skill improvements, with only two eligible studies included in this subgroup. No significant differences were observed across CTA anatomical types (χ²=1.62, df = 2, P = 0.4443). Learner group (1) For Residents (4 studies, n = 163), AI-assisted teaching significantly enhanced skills (pooled SMD = 1.51, 95% CI [0.50, 2.53], P = 0.0034) with high heterogeneity (I²=86%, P = 0.0001); (2) for Students (2 studies, n = 144), the pooled SMD was 0.76 (95% CI [0.42, 1.10], P < 0.0001) with no significant heterogeneity (I²=0%), indicating consistent benefits in foundational CTA skills for undergraduate learners. No significant differences were observed between the Resident and Student subgroups (χ²=1.92, df = 1, P = 0.1656), though a numerical trend suggested potentially larger effects in Residents. Sample size subgroup To evaluate the potential impact of small-sample studies on the pooled effect estimates, we performed a subgroup analysis stratified by total sample size. (1) For small sample size (n < 40, 2 studies, n = 46), AI-assisted teaching significantly improved skills (pooled SMD = 0.87, 95% CI [0.25, 1.48], P = 0.0056) with no significant heterogeneity (I²=0%, P = 0.672); (2) for large sample size (n ≥ 40, 4 studies, n = 261), the pooled SMD was 1.37 (95% CI [0.54, 2.20], P = 0.0013) with high heterogeneity (I²=89%, P < 0.0001). No significant differences were observed between the small and large sample size subgroups (χ²=0.90, df = 1, P = 0.3419), confirming that the overall pooled effect estimates remained robust and stable, and were not unduly influenced by the small-sample studies. These subgroup analyses partially explained the overall high heterogeneity (I²=82%), with CTA anatomical complexity, learner background, and sample size identified as key contributors. AI teaching demonstrated significant skill improvement across all subgroups, reflecting broad applicability in standardized CTA skill training. Sensitivity analyses Sensitivity analyses, including leave-one-out and exclusion of studies with high risk of bias, were conducted for the three main outcomes to evaluate the robustness of effect estimates. Overall, the results were consistent with the main meta-analytic findings, supporting the stability of the pooled effect estimates. Detailed results are presented in Supplementary Tables 6–11. Discussion This systematic review and meta-analysis synthesized data from 7 RCTs involving 515 participants to evaluate AI-assisted teaching in CTA education. Frequentist and Bayesian analyses consistently demonstrated that AI-powered instruction significantly improves practical skills and reduces image judgment time, with positive trends observed for theoretical knowledge and teaching satisfaction. Subgroup analyses identified CTA anatomical type and learner background as key sources of heterogeneity, confirming AI’s broad applicability across different training contexts. These findings align with the broader consensus that AI enhances radiology education while providing novel insights into its utility for CTA-specific training, a domain where targeted anatomical visualization and efficient image interpretation are paramount [6, 13]. Hedges’ g was used for all effect size calculations to minimize bias associated with small sample sizes. Core findings and mechanisms The significant improvement in practical skills (pooled SMD = 1.24) underscores AI’s strength in facilitating hands-on learning, and this is consistent with prior research showing AI-driven simulations enhance technical proficiency and that high-quality practical training can compensate for individual differences in visual-spatial abilities [7, 28]. AI tools achieve this by offering interactive visualization of complex vascular anatomy [29, 30], real-time feedback on image interpretation [3], and targeted practice of high-complexity cases [6], thereby addressing critical gaps identified in radiology training surveys, such as limited structured practice and variable feedback [2]. The marked reduction in image judgment time (pooled SMD=-2.70) highlights AI’s role in optimizing diagnostic efficiency, critical given the growing volume of imaging studies and time constraints for trainees [31, 32]. Notably, shortened image interpretation time does not independently equate to improved diagnostic accuracy. AI-assisted CTA education requires a balanced trade-off between diagnostic speed and interpretive precision, and future outcome assessments should integrate both efficiency metrics and diagnostic accuracy indicators to fully reflect clinical diagnostic quality. For theoretical scores and teaching satisfaction, Bayesian analyses showed greater uncertainty, with 95% CrIs crossing the null value. This was due to high heterogeneity (I²=91% and 87%), which arose from three key factors: ① variations in AI tool functionality—some tools offer 3D visualization and real-time feedback, while others only provide basic annotation [33], with interactive tools better facilitating active learning while basic tools yield inconsistent outcomes [30]. ② differences in training duration: longitudinal intensive training ≥ 4 weeks consolidates skills via repeated AI interaction, whereas short-term interventions only achieve superficial improvement [34]. ③ lack of standardized assessment: diverse evaluation tools including validated scales and self-designed questionnaires result in inconsistent outcome measurement [30]. Radiology postgraduates report positive attitudes but limited use due to accuracy concerns [4]. Frequentist analyses still showed significant benefits, supporting AI’s potential to enhance knowledge acquisition and learner experience when properly implemented [35]. Subgroup analysis insights Subgroup findings clarified that CTA anatomical type influences AI’s effectiveness: Head-Neck CTA showed the largest effect size (SMD = 2.07) but extreme heterogeneity (I²=94%), while Comprehensive and Coronary CTA demonstrated consistent, homogeneous benefits (I²=0%). This variation likely reflects the anatomical complexity of head-neck vasculature, and AI tools may better address the steep learning curve for this region by integrating 3D anatomical labeling and targeted error correction [29], aligning with clinicians’ observations that CT-based anatomy teaching benefits from 3D insight and studies showing interactive 3D post-processing improves visual-spatial abilities and diagnostic skills [1, 28]. Residents exhibited larger effect sizes (SMD = 1.51) than students (SMD = 0.76), though both groups benefited significantly. This trend may reflect residents’ greater exposure to clinical workflows, enabling them to better integrate AI tools into practice [3], as well as their higher familiarity with radiological concepts that AI tools build upon [4]. However, the absence of significant subgroup differences confirms AI’s utility for both foundational and advanced CTA training, which represents a key advantage for curriculum integration and addressing the need for standardized training across learner levels [31, 35]. Subgroup analysis stratified by sample size further revealed that AI-assisted teaching retained significant beneficial effects in both small-sample (n < 40) and large-sample (n ≥ 40) subgroups, with no significant between-subgroup difference observed. These results indicate that the overall pooled effect estimates for skill scores were robust and not unduly influenced by the inclusion of small-sample studies, supporting the stability of the main findings. Given the very high heterogeneity in the Head-Neck CTA subgroup (I² = 94%), all subgroup results should be considered exploratory and hypothesis-generating rather than confirmatory. Comparison with existing literature Our findings align with and build upon existing research on AI in radiology education. For instance, a systematic review by Keshavarz et al., reported that 92.9% of studies showed AI improves trainees’ knowledge and skills—consistent with our observation of significant practical skill gains [6]. Shah et al., additionally showed that AI-enhanced simulation cases improve educational value without increasing time burden, mirroring our finding of reduced image judgment time [7]. Notably, our focus on CTA-specific training addresses a gap identified in scoping reviews [12], which emphasized the need for targeted evaluations of AI education in subspecialty contexts. Our dual frequentist-Bayesian approach adds rigor by quantifying uncertainty, and this represents an advance over prior meta-analyses that often rely solely on frequentist methods [31], while also aligning with calls for rigorous outcome assessment in AI education research [36]. The uncertainty observed for theoretical scores and satisfaction highlights the need for standardized AI tool design and outcome measurement [37], a point echoed in prior calls for consistent AI curriculum frameworks and competency-based AI education [3, 35]. Parallel developments in clinical cardiac imaging have demonstrated that advanced large language models can reliably support coronary CTA reporting in accordance with CAD-RADS 2.0 and significantly improve the diagnostic accuracy of radiologists in cardiac case interpretation. These clinical advances strongly reinforce the value of integrating AI-assisted training in CTA education, as modern radiology trainees will increasingly practice within LLM-supported clinical workflows, making aligned educational and clinical AI strategies essential for sustainable skill development [14, 15]. Strengths and limitations This study’s strengths include adherence to PRISMA guidelines, registration in PROSPERO, comprehensive searches of Chinese and English databases, and integration of Bayesian and frequentist analyses. Sensitivity analyses including leave-one-out and exclusion of high-risk-of-bias studies were also performed to assess result robustness, and overall findings were consistent with the main meta-analytic results, supporting the stability of the pooled estimates. Subgroup analyses and drapery plots further enhance result interpretability. However, limitations exist: only 7 small-scale studies were included, all with high or unclear bias risk (a common issue in educational interventions [6]), and high heterogeneity persists for some outcomes, which is consistent with the broader landscape of AI in radiology education, characterized by heterogeneous study designs and limited large-scale trials [12]. Furthermore, although funnel plots were constructed for the assessment of publication bias, the small number of included studies (2–6 studies per outcome) meant that formal statistical tests for asymmetry were not reliable. As such, publication bias cannot be reliably excluded, which may affect the interpretation of the pooled effect sizes. Additionally, the exclusion of non-RCTs and gray literature may introduce selection bias, and we did not assess trainee gender or racial disparities—factors that influence radiology training experiences [38]; and may interact with AI tool utilization [39]. Furthermore, all included studies were conducted in China using Chinese-language AI systems, which may limit the generalizability of findings to other regions and linguistic populations. Practical implications and future directions AI-assisted teaching should be integrated into CTA curricula to address anatomical complexity and efficiency demands, especially considering the low satisfaction with traditional radiology training and the positive attitudes toward AI among trainees [2, 4]. Specialty-specific AI modules (Head-Neck CTA) and stage-matched training (residents vs. students) can reduce heterogeneity [3], while aligning with core AI competencies for radiologists, which include clinical implementation, algorithm evaluation, and error handling [35]. Future research should prioritize large-scale, multi-center RCTs with rigorous design [31], standardized AI tool reporting [37], and longitudinal assessment of clinical outcomes [7], thus addressing gaps identified in scoping reviews [36]. Additionally, research should explore AI’s role in addressing diversity gaps in radiology training and incorporate trainee feedback to optimize tool usability [4, 39]. Cost-effectiveness analyses and mixed-methods studies incorporating learner perspectives will further support widespread adoption, as will the development of web-based AI applications for accessible perception training [29, 40]. In conclusion, AI-powered instruction may improve CTA teaching outcomes, particularly practical skills and diagnostic efficiency. It should be emphasized that the present findings are limited to short-term educational outcomes and do not support inferences regarding long-term knowledge retention, clinical diagnostic performance, or patient-level outcomes. However, due to the low-to-moderate certainty of evidence rated by the GRADE approach, as well as the methodological limitations of included studies (e.g., high risk of bias, unclear allocation concealment), the above findings should be interpreted with caution. While heterogeneity and small sample sizes warrant caution, subgroup analyses provide actionable insights for curriculum optimization. Future large-scale, high-quality, double-blind RCTs with adequate allocation concealment are needed to validate these findings. Such trials will help establish AI’s role in standardized, efficient, and inclusive radiology education while addressing long-standing training challenges and aligning with evolving clinical practice demands [2]. Acknowledgements The authors would like to thank all researchers involved in the original studies included in this systematic review and meta-analysis. Abbreviations Artificial Intelligence Computed Tomography Angiography Randomized Controlled Trial Standardized Mean Difference Risk Ratio Grading of Recommendations Assessment, Development and Evaluation International Prospective Register of Systematic Reviews Liang Yuan is an Associate Chief Physician, Associate Professor, and Master’s Supervisor at the Department of Radiology, University-Town Hospital of Chongqing Medical University. His current research interests include medical imaging education, with a focus on the design and outcome evaluation of AI-powered instructional methods for radiology training, particularly in Computed Tomography Angiography (CTA). Authors’ contributions L.Y. designed the study. X.Y.X. and Y.Q.W. collected the data, performed the literature screening and data extraction. L.Y. analyzed the data and drafted the manuscript. All authors contributed to the revision of the manuscript and approved the final version. Funding The authors declare that no funding was received for this study. Data availability All data analysed during this systematic review and meta-analysis are included in this published article and its supplementary information files. Raw extracted data are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate Not applicable. This study is a systematic review and meta-analysis based on published literature, and does not involve original human or animal experiments. 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. Liang Yuan and Xingyi Xie contributed equally to this work. References Associated Data Data Availability Statement All data analysed during this systematic review and meta-analysis are included in this published article and its supplementary information files. Raw extracted data are available from the corresponding author on reasonable request.