Cross-cultural adaptation and psychometric evaluation of the Turkish ATRAI-14 among radiology residents: a multicenter national survey

Abstract

Background

Artificial intelligence (AI) is increasingly integrated into radiology, and residents’ attitudes are central to its adoption. The ATRAI-14 measures radiologists’ attitudes toward AI implementation, but its performance in other linguistic and health-system contexts is unknown. We aimed to adapt it into Turkish, examine whether its three-domain structure replicates in a national cohort of radiology residents, and describe their attitudes.

Methods

We surveyed radiology residents from 25 institutions across 17 cities in Türkiye in a national, cross-sectional, web-based study. The questionnaire was translated into Turkish and content-validated by eight expert radiologists (item-content validity index 0.875). Psychometric evaluation comprised confirmatory and exploratory factor analysis, Cronbach’s α and McDonald’s ω, item–total correlations, and criterion validity against a visual analog scale. Group differences and predictors of the total score were examined non-parametrically and by regression.

Results

Among 286 participants, the three-factor model fitted inadequately (CFI = 0.82, RMSEA = 0.075). Only Trust reached borderline-acceptable internal consistency (α = 0.69); Implementation Perspectives (α = 0.07) and Hopes and Fears (α = 0.06) lacked cohesion, and two items loaded negatively. Criterion validity for the total score was acceptable (ρ = 0.56), and the mean total score (20.99 ± 3.60) was 2.99 points above the scale midpoint of 18, indicating a modestly positive attitude. Residents favored AI as a collaborative aid (90.6% would re-examine a study when AI disagreed; item T4) but were uncomfortable with autonomous AI (22.0%; item T1). Most expected reduced workload (79.7%; item H4), yet 60.8% feared diminished prestige (item H1), and only 21.7% had access to AI tools at work. Attitudes did not differ by AI access, seniority, institution type, or research involvement.

Conclusions

The three-domain structure only partially replicated: Trust performed acceptably, whereas the other two require cross-cultural refinement before the total score can be used reliably. Because these two domains did not form coherent constructs, their findings should be read as observations about individual items rather than as measurements of the intended constructs, and should not be used to compare groups. With that caveat, the cohort was cautiously optimistic, valuing AI as a collaborator while resisting its autonomy. Few residents had used AI in practice; incorporating supervised, hands-on AI experience into residency training could help align expectations with practice — a hypothesis warranting direct evaluation rather than a conclusion from cross-sectional data.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12909-026-10415-3.

Introduction

The advent of artificial intelligence (AI), driven by rapid advancements in deep learning and big data analytics, has precipitated a paradigm shift in modern medicine, with radiology positioned at the epicenter of this transformation [1–3]. Over the last decade, AI applications have evolved from experimental algorithms into clinical tools spanning image reconstruction, workflow optimization, quantitative analysis, and lesion detection [4, 5]. While these technologies promise greater diagnostic accuracy, shorter turnaround times, and relief from rising workload, they have simultaneously ignited a global debate about the future role of the profession [6].

The disruption carries a dichotomy of “hope” and “fear”: AI is envisioned as a powerful “second reader” that augments human intelligence, yet concerns about professional displacement, algorithmic bias, and liability continue to provoke anxiety within the medical community [7, 8]. The successful implementation of AI in clinical practice, therefore, relies not only on the technical validity of the algorithms but fundamentally on the “human factor”—specifically, the readiness, trust, and sociotechnical perceptions of the radiologists who will wield these tools [9].

Radiology residents are likely to form a substantial part of the workforce in an increasingly hybrid human–AI practice environment, and their attitudes today may shape how AI is adopted in the years ahead. Although current residents are often termed “digital natives,” general technological fluency does not equate to the technical and ethical literacy that clinical AI integration requires, leaving a gap between confidence and competence [10, 11]. Many residents, while recognizing AI’s growing inevitability, still report inadequate preparedness, largely reflecting the absence of structured AI education within residency training [11, 12]. At the same time, residents’ attitudes toward AI are not uniform; perceptions appear heterogeneous and seem to be shaped more by prior exposure, self-reported knowledge, and institutional context than by training seniority alone [13]. Clarifying these attitudes—separating rational skepticism from non-evidence-based fear—is therefore essential to inform targeted educational strategies and support the safe, clinically responsible integration of AI into residency training. High algorithmic performance alone may not translate into clinical success, since bridging the so-called ‘AI chasm’ depends substantially on the willingness and sociotechnical readiness of radiologists to adopt these tools [3, 14, 15].

To date, numerous surveys have attempted to gauge radiologists’ attitudes toward AI. However, the majority of these studies have been concentrated in North America and Western Europe, leaving the perspectives of radiologists in dynamic, high-volume healthcare environments such as Türkiye largely underrepresented [16, 17]. Furthermore, a significant methodological limitation in the existing literature is the widespread use of non-standardized, ad-hoc questionnaires, which hampers the reproducibility of data. As highlighted by recent reviews, there is a critical need for standardized metrics to enable meaningful cross-cultural comparisons [18]. Addressing this gap, Vasilev et al. developed the Attitude Towards AI in Radiology (ATRAI-14) scale, a multidimensional psychometric instrument designed to quantify radiologists’ attitudes across distinct conceptual domains [19].

The ATRAI-14 distinguishes itself from earlier ad-hoc instruments through its structured, domain-based design, intended to capture the multifaceted nature of human–AI interaction. Comprising items across three scored sub-domains—“Trust,” “Implementation Perspectives,” and “Hopes and Fears”—the scale moves beyond simple acceptance metrics, enabling a granular analysis of how radiologists perceive algorithmic reliability, workflow integration, and future professional security. However, the psychometric properties of any instrument are established within a specific linguistic, professional, and health-system context, and cannot be assumed to transfer unchanged to a new population. Established guidelines for the cross-cultural adaptation of measurement instruments therefore recommend that an adapted scale be independently re-examined in its target setting rather than treated as valid by default [20, 21]. This makes the ATRAI-14 both a promising candidate for use among Turkish radiology residents and an instrument whose behavior in this context requires explicit evaluation.

Two gaps therefore remain. First, it is not known whether the scored structure of the ATRAI-14 holds outside the population and health system in which it was developed, since its measurement properties have not been examined in another language or training context. Second, the domain-specific attitudes of Turkish radiology residents — their trust in AI output, their views on implementation, and their expectations for the profession — have not been characterised with a multidimensional instrument. This national, cross-sectional survey therefore had two aims: first, to adapt the ATRAI-14 into Turkish and examine whether its three-domain structure replicates in a large cohort of radiology residents; and second, to characterize these residents’ attitudes toward AI and explore whether background factors — seniority, institutional setting, and practical AI exposure — are associated with those attitudes.

Methods

Study design and participants

This national, cross-sectional, web-based survey was conducted between 20 and 31 December 2025. The study adhered to the Declaration of Helsinki and was approved by the Institutional Clinical Research Ethics Committee (date: 18 December 2025; decision no: 2025-74). Electronic informed consent was obtained from all participants before they accessed the questionnaire.

Participants were radiology residents in active training, recruited from 25 institutions (university hospitals, training and research hospitals, and city hospitals) across 17 cities representing all seven geographical regions of Türkiye. Residents from all training years (postgraduate years 1–5) were eligible. Sampling was non-probability (convenience) in nature, since no central roster of radiology residents exists in Türkiye from which a random sample could be drawn; recruitment was instead directed deliberately across the three main institution types and all seven geographical regions in order to reduce single-centre and single-region bias. Because the invitation was relayed to residents within each department rather than issued to a fixed roster, the number of residents who received or viewed it could not be determined and a conventional response rate cannot be calculated. All questionnaire items were configured as mandatory in the survey platform, so every submitted response was complete; consequently there were no partial submissions, no record was excluded for missing data, and the 286 analysed responses represent all voluntary participants.

Data collection

Data were collected through a structured online questionnaire (Google Forms) distributed to radiology residents across the participating centers. To promote participation and sample representativeness, recruitment proceeded through institutional gatekeepers rather than by open online posting. For each participating department the investigators contacted, by telephone, the faculty member responsible for residency training and, where that was not possible, the department’s resident representative. These individuals were informed about the purpose and content of the study and were asked to pass the invitation and the survey link to the residents in their department, each of whom then decided individually whether to take part; distribution within departments used the professional channels already in place, including institutional messaging groups. Participation was entirely voluntary. On accessing the survey link, all participants were presented with an electronic informed consent form, and only those who provided consent could proceed to the questionnaire. The survey platform was configured not to collect e-mail addresses, names, or any other identifying information, and no data that could identify an individual respondent or institution were recorded; responses were held in an access-restricted account available only to the investigators and were analysed in aggregate. Residents could leave the questionnaire at any point without submitting it, and no reminder was directed at any identifiable individual. A total of 286 radiology residents completed the survey in full and were included in the final analysis. The participant selection process and study flow are summarized in Fig. 1.

Fig. 1.

Participant flow. The invitation was passed to residents by the training supervisor, or the resident representative, of each participating department, so no fixed sampling frame existed and the number of residents who received or viewed it could not be determined; a conventional response rate therefore cannot be calculated. All questionnaire items were configured as mandatory, so every submitted response was complete and no record was excluded for missing data

Sample size

The target sample size was determined based on the psychometric recommendation of 10–15 participants per item for scale validation studies [22]. With 286 participants completing the 14-item scale, our study achieved a subject-to-item ratio of approximately 20:1. This ratio exceeds the recommended minimum criteria, providing sufficient statistical power for the evaluation of the scale’s psychometric properties and factor analysis [23].

Survey instrument and adaptation

The Turkish version of the questionnaire, together with an English translation, and the original English ATRAI-14 items with their scoring algorithm, are openly available in a Zenodo repository (https://doi.org/10.5281/zenodo.19039735). The original items and scoring were reproduced with minor formatting changes from Vasilev et al. [19] under a CC BY 4.0 licence.

Demographic items such as gender and exact age were intentionally omitted, both to protect respondent anonymity and because the original instrument was likewise designed without identifying demographic items. The first section (professional profile) captured residency year (seniority), institution type, active involvement in research, and current access to AI tools in clinical practice. The second section comprised the ATRAI-14 scale, originally developed by Vasilev et al. [19]. This instrument comprises 14 items distributed across four domains: Familiarity (items F1, F4), capturing prior experience with AI; Trust (items T1, T2, T3, T4, T6), reflecting reliance on AI for autonomous or semi-autonomous tasks; Implementation Perspectives (items I1, I2, I4, H6), addressing the practical integration, workflow, and financial aspects of AI; and Hopes and Fears (items H1, H4, H5), concerning future expectations regarding job security, prestige, and remuneration. By the developers’ design, only the latter three domains, comprising 12 items, contribute to the total score; the two Familiarity items are not scored and serve solely to characterise respondents by their prior experience with AI. All psychometric analyses (confirmatory and exploratory factor analysis, internal consistency, and criterion validity) were therefore conducted on the 12 scored items, and the two Familiarity items were used descriptively only. Item labels follow the original instrument, in which the numbering is non-consecutive because candidate items were dropped during scale development, and in which item H6 belongs to the Implementation Perspectives domain despite its prefix.

Since the original scale was in English, it was first translated into Turkish by the authors. To ensure semantic clarity, cultural appropriateness, and content validity, the translated draft was independently reviewed by an expert panel of eight senior radiologists (four associate professors and four senior specialists) working in five different institutions (three university hospitals and two city hospitals). Each expert rated every item on a 4-point relevance scale (1 = not relevant; 2 = somewhat relevant/requires major revision; 3 = quite relevant/requires minor revision; 4 = highly relevant). This yielded a mean Item-Content Validity Index (I-CVI) of 0.875, equivalent to a scale-level content validity index averaging method (S-CVI/Ave) of 0.875, with seven of eight experts rating all items as highly relevant, exceeding the recommended threshold of 0.78 [24]. Minor wording refinements were made based on the panel’s feedback before the questionnaire was finalized; for example, the description of AI tools in the Familiarity item was broadened to span quantitative analysis and triage systems through to generative models, and the wording of items T1 and T2 was adjusted to reduce ambiguity regarding autonomous functionality and algorithmic output. The background section was adapted to the Turkish training context (e.g., institution types were specified as university, training and research, and city hospitals; seniority was recorded by residency year). The Turkish version as administered was independently back-translated into English by a bilingual radiologist (professor of radiology, 17 years of experience, with an international English-language publication record and prior training abroad) who had taken no part in the forward translation and who confirmed in writing that they had not seen the original English instrument. All 14 items, the global rating item, and every response option were back-translated and compared with the source. No unintended semantic drift was identified in the scored items; the deliberate wording refinements described above, together with two culturally required localisations, are documented in Additional file 3.

Scoring

Scoring strictly followed the original ATRAI-14 algorithm described by Vasilev et al. [19]. For single-answer items, the score equalled the numeric label of the selected option; for the three multi-select items (I1, I4, H6), the number of selected functional options was linearly converted to a 0–4 score per the original conversion tables. The total score was computed using the developers’ weighted formula, S = (T1 + T2+T3 + T4+T6) × 0.6 + (I1 + I2+I4 + H6) × 0.75 + (H1 + H4+H5), yielding a range of 0–36, with higher scores indicating a more positive attitude toward AI. Descriptive interpretation of the total score is anchored to this theoretical range: its midpoint is 18, so a mean above 18 denotes an attitude that is on balance positive and a mean below 18 one that is on balance negative. Because the aim was to test the published instrument as it stands, the developers’ scoring algorithm was applied without modification. The published algorithm does not specify how to score respondents who selected both a substantive option and the “none” or “undecided” alternative within a multi-select item (27 of 286 respondents, 9.4%). These responses were scored on their substantive selections, with substantive options taking precedence; no responses were excluded, since such patterns are interpretable rather than erroneous (for example, endorsing specific useful AI functions while doubting widespread implementation). A sensitivity analysis applying the opposite rule produced materially identical results (total score 20.78 versus 20.99; Implementation α − 0.03 versus 0.07; total-score criterion validity ρ = 0.56 under both rules) and altered no conclusion.

Statistical and psychometric analysis

Analyses were performed and figures generated in Python (v3.12; SciPy, statsmodels, factor_analyzer, semopy, Matplotlib), and reported in line with COSMIN recommendations for studies on measurement properties. Continuous variables were summarized as mean ± SD and median (IQR); categorical variables as counts and percentages. Normality was assessed with the Shapiro–Wilk test and visual inspection; as scores were not normally distributed, non-parametric tests were used throughout.

Structural validity

Confirmatory factor analysis (CFA) tested the original three-factor model, with fit evaluated using the comparative fit index (CFI; ≥0.90 acceptable), Tucker–Lewis index (TLI; ≥0.90), root mean square error of approximation (RMSEA; ≤0.08 acceptable), and goodness-of-fit index (GFI). Standardized loadings ≥ 0.40 were considered adequate. Because the confirmatory model did not fit adequately, an exploratory factor analysis (EFA; principal-axis factoring) was conducted using an oblique (promax) rotation, chosen a priori because the ATRAI-14 domains are conceptually related facets of a single overarching construct and were therefore expected to correlate rather than to be orthogonal; sampling adequacy was confirmed by the Kaiser–Meyer–Olkin measure and Bartlett’s test of sphericity, and the number of factors was determined by parallel analysis. The exploratory analysis was pre-specified as a diagnostic step to be undertaken only if the confirmatory model failed, in order to characterise how the structure departed from the original and to locate the misfitting items. It was therefore conducted in the same sample as the confirmatory test and is reported as exploratory and hypothesis-generating rather than as independent confirmation; because such an analysis capitalises on chance, the resulting solution may be sample-specific and requires replication in an independent cohort

Internal consistency

Reliability was quantified using Cronbach’s α (with bootstrap 95% confidence intervals, 2000 resamples) and McDonald’s ω, reported for the total scale and each domain separately; ω is reported alongside α because it does not assume tau-equivalence and is preferable for multidimensional instruments [25]. Conventional interpretive bands for α were applied (≥ 0.70 acceptable; 0.60–0.69 questionable; 0.50–0.59 poor; <0.50 unacceptable). Corrected item–total correlations (Spearman’s ρ; ≥0.30 considered adequate) and the average inter-item correlation (AIIC; recommended 0.15–0.50) were computed per domain.

Criterion validity

Spearman’s rank correlation (ρ) was calculated between the total ATRAI-14 score and a self-reported visual analog scale (VAS, 0–10; observed range 1–10) of overall attitude toward AI, and repeated for each domain.

Group comparisons and predictors

Two-group comparisons used the Mann–Whitney U test with Cohen’s d and rank-biserial correlation as effect sizes; comparisons across institution types used the Kruskal–Wallis test with η²H. The “Other” institution category (n = 2) was excluded from subgroup comparisons owing to insufficient cell size but retained as a covariate in the multivariable model. A multivariable linear regression (heteroscedasticity-robust HC3 standard errors) modelled the total score on AI access, seniority, research activity, and institution type, reporting standardized coefficients. Effect sizes are reported with 95% confidence intervals; intervals for rank-biserial correlations were obtained by bootstrap resampling (5000 replicates). Statistical significance was set at two-sided p < 0.05.

Results

Participant characteristics

A total of 286 radiology residents from 25 institutions across 17 cities completed the survey. Senior residents (years 3–5) comprised 54.9% (n = 157) of the sample and junior residents (years 1–2) 45.1% (n = 129). The largest institutional group worked at city hospitals (47.2%), followed by university hospitals (28.7%) and training and research hospitals (23.4%). While 43.7% of residents were actively involved in AI-related research, only 21.7% reported having access to AI tools in their daily clinical practice. Full participant characteristics are presented in Table 1.

Table 1.

Socio-demographic characteristics of the participants (n = 286)

Characteristics	Groups	n	%
Seniority	Junior (Year 1–2)	129	45.1
	Senior (Year 3–5)	157	54.9
Institution Type	City Hospital	135	47.2
	University Hospital	82	28.7
	Training & Research Hospital	67	23.4
	Other (Private/State)	2	0.7
AI Access at Work	No	224	78.3
	Yes	62	21.7
Scientific Research Activity	No	161	56.3
	Yes	125	43.7
Total		286	100.0

AI Artificial intelligence

Cross-cultural adaptation and content validity

The expert panel review yielded an Item-Content Validity Index (I-CVI) of 0.875, exceeding the 0.78 threshold for satisfactory content validity. Minor semantic refinements were implemented based on the panel’s feedback, after which the final Turkish version was judged to have satisfactory face and content validity.

Structural validity and internal consistency

Confirmatory factor analysis (CFA) of the original three-factor model showed inadequate fit (CFI = 0.82, TLI = 0.77, RMSEA = 0.075, GFI = 0.75; Table 2). Examination of the standardized loadings showed that the Trust items T1–T3 loaded adequately (0.63–0.78), whereas T4 (0.34) and T6 (0.36) loaded weakly. Within the Implementation domain, I1 (0.72) and H6 (0.43) loaded adequately, while I2 (0.23) was very weak and I4 loaded negatively (− 0.41). In the Hopes and Fears domain, H1 (0.59) and H5 (0.66) loaded adequately, whereas H4 loaded negatively (− 0.32).

Table 2.

Psychometric properties of the Turkish ATRAI-14 (n = 286)

Parameter	Value	Reference threshold
Structural validity (CFA, three-factor model)
Comparative fit index (CFI)	0.82	≥ 0.90
Tucker–Lewis index (TLI)	0.77	≥ 0.90
RMSEA	0.075	≤ 0.08
Goodness-of-fit index (GFI)	0.75	≥ 0.90
Internal consistency and criterion validity, by domain	α (95% CI)	ρ with VAS
Trust (5 items)	0.69 (0.61–0.74)	0.45***
Implementation (4 items)	0.07 (–0.13–0.22)	0.27***
Hopes & Fears (3 items)	0.06 (–0.20–0.25)	0.40***
Total scale (12 items)	0.56 (0.46–0.64); ω = 0.57	0.56***
Composite reliability and variance extracted, by domain	CR	AVE
Trust (5 items)	0.70	0.34
Implementation (4 items)	0.23	0.23
Hopes & Fears (3 items)	0.29	0.29

CFA confirmatory factor analysis, RMSEA root mean square error of approximation, VAS visual analog scale; ω McDonald’s omega, CR composite reliability (≥ 0.70 acceptable), computed from the standardized CFA loadings and numerically equivalent to congeneric ω, which is shown by domain in Fig. 3; AVE average variance extracted (≥ 0.50 acceptable). Average inter-item correlations: Trust 0.30, Implementation ≈ 0.00, Hopes & Fears 0.01. Item communalities, standardized CFA loadings and the full EFA loading matrix are provided in Additional file 1 and Fig. 2. ***p < 0.001

Given the inadequate confirmatory fit, an exploratory factor analysis (EFA) was performed (Kaiser–Meyer–Olkin = 0.70; Bartlett’s test of sphericity χ² = 515.5, p < 0.001). Parallel analysis supported a three-factor solution, which explained 34.5% of the total variance. As shown in the loading matrix (Fig. 2; Additional file 1), the first factor was defined by the Trust items T1–T3, the second grouped the Implementation items I1 and H6 together with T6 (with I4 loading negatively), and the third was defined by H1 and H5. Three items (T4, I2, H4) did not load on any factor at the 0.30 threshold. Taken together with the inadequate confirmatory fit and the low domain reliabilities, this modest proportion of explained variance indicates a weak latent structure rather than a robust multidimensional one.

Fig. 2.

Exploratory factor analysis loading matrix (promax rotation, three-factor solution). Cell values are item loadings; bold values denote the primary loading (≥ 0.30) for each item. Color strips on the left indicate the original conceptual domain of each item. Note the negative loading of item I4 on the second factor and of item H4 on the third

Internal consistency mirrored this structural pattern (Table 2; Fig. 3). Reliability of the total 12-item scale was low (Cronbach’s α = 0.56, 95% CI 0.46–0.64; McDonald’s ω = 0.57). Domain-level reliability was markedly heterogeneous: the Trust domain showed borderline-acceptable internal consistency (α = 0.69; average inter-item correlation [AIIC] = 0.30), whereas the Implementation (α = 0.07; AIIC ≈ 0.00) and Hopes and Fears (α = 0.06; AIIC = 0.01) domains showed essentially no internal cohesion. Corrected item–total correlations reached the conventional 0.30 threshold only for items T1, T2 and T3.

Fig. 3.

Internal consistency for each domain and the total scale: solid bars show Cronbach’s α with 95% confidence intervals and hatched bars show McDonald’s ω. Reference lines mark the conventional acceptable (0.70) and poor (0.50) thresholds. Only the Trust domain approached the acceptable threshold; the Implementation and Hopes and Fears domains showed negligible internal consistency

Criterion validity

The total ATRAI-14 score correlated significantly with the self-reported visual analog scale (VAS) attitude score (Spearman’s ρ = 0.56, p < 0.001; Table 2). At the domain level, correlations with the VAS were 0.45 for Trust (p < 0.001), 0.40 for Hopes and Fears (p < 0.001), and 0.27 for Implementation (p < 0.001); the Trust domain thus showed the strongest domain-level correlation with the VAS.

Attitudes toward artificial intelligence

The mean total ATRAI-14 score was 20.99 ± 3.60 (observed range 6.75–30.90; possible range 0–36). Among the weighted domains, Implementation scored highest (8.19 ± 1.71), followed by Trust (6.87 ± 2.05) and Hopes and Fears (5.93 ± 1.54) (Table 3). Dispersion varied systematically across items: the widest standard deviations occurred for items requiring judgements about health-system financing and professional trajectory (I2, SD 1.39; T2, SD 1.17), and the narrowest for concrete, clinically anchored items (T4, SD 0.73; H5, SD 0.75), a pattern that parallels the domain-level reliability findings. The item-level profile is shown in Fig. 4.

Table 3.

Domain scores and group comparisons of the total ATRAI-14 score (n = 286)

Weighted domain scores (mean ± SD)	Value	Statistic
	Trust	6.87 ± 2.05	—
Implementation	8.19 ± 1.71	—
Hopes & Fears	5.93 ± 1.54	—
Total score	20.99 ± 3.60	—
Group comparisons (total score)	mean ± SD	p; effect size
AI access: yes (n = 62) vs. no (n = 224)	21.56 ± 3.64 vs. 20.84 ± 3.58	0.119; d = 0.20 (95% CI − 0.08 to 0.48)
Seniority: junior (n = 129) vs. senior (n = 157)	20.82 ± 3.61 vs. 21.14 ± 3.60	0.305; d = − 0.09 (95% CI − 0.32 to 0.14)
Research: yes (n = 125) vs. no (n = 161)	21.22 ± 3.45 vs. 20.82 ± 3.71	0.286; d = 0.11 (95% CI − 0.12 to 0.35)
Institution type (three groups)	—	0.821; η²H ≈ 0

SD standard deviation, d Cohen’s d; η²H, eta-squared for the Kruskal–Wallis test. Two-group comparisons used the Mann–Whitney U test; the three-group comparison used the Kruskal–Wallis test. Item-level scores are shown in Fig. 4; full multivariable regression results are provided in Additional file 2

Fig. 4.

Item-level attitude scores. A Mean ± SD for each scored item (possible range 0–4); bars are coloured by conceptual domain and the dashed line marks the neutral midpoint (score = 2). B The full response distribution for each item, shown as the percentage of respondents selecting each score. A stacked distribution is shown because several items have an interquartile range spanning a single category, for which box geometry would be uninformative

A clear contrast emerged between collaborative and autonomous use of AI within the Trust domain. Residents were highly willing to re-examine a study when an AI tool contradicted their reading (item T4, 3.36 ± 0.73; 90.6% answered “yes” or “mostly yes”), yet comfort with fully autonomous AI was low: only 22.0% reported they would feel comfortable trusting an autonomous AI to interpret images without radiologist oversight (item T1).

Within the Hopes and Fears domain, an efficiency–devaluation tension was apparent. While 79.7% of residents expected AI to reduce their workload (item H4), 60.8% feared that widespread AI adoption would diminish the profession’s prestige (item H1), and 39.5% anticipated a decrease in salary (item H5).

Group comparisons and predictors

Attitude scores did not differ significantly by residency seniority (junior 20.82 ± 3.61 vs. senior 21.14 ± 3.60; p = 0.305, Cohen’s d = − 0.09, 95% CI − 0.32 to 0.14), institution type (Kruskal–Wallis H = 0.40, p = 0.821, η²H ≈ 0), or research involvement (p = 0.286, d = 0.11, 95% CI − 0.12 to 0.35) (Table 3). Residents with access to AI tools scored slightly higher than those without (21.56 ± 3.64 vs. 20.84 ± 3.58), but this difference did not reach statistical significance (p = 0.119) and the effect size was small (d = 0.20, 95% CI − 0.08 to 0.48). The confidence interval for this comparison includes the null value, and its width reflects the imprecision arising from the unbalanced subgroup sizes (n = 62 vs. n = 224); an adequately powered future test of an effect of this magnitude would require approximately 270 participants per group. Sub-domain analysis showed no significant differences by AI access in any domain (Trust p = 0.135; Implementation p = 0.887; Hopes and Fears p = 0.146). The two AI-access groups were also compared on the recorded background characteristics. They did not differ by seniority (Fisher exact p = 0.194), research involvement (p = 0.193), or pre-residency radiology experience (p = 0.772), but did differ by institution type (χ²(2) = 9.44, p = 0.009, Cramér’s V = 0.18; the two residents in the “Other” category were excluded from this comparison, as from the other subgroup analyses). Among residents with access to AI tools, 59.0% worked in city hospitals, 27.9% in training and research hospitals and 13.1% at university hospitals, whereas among those without access the corresponding proportions were 44.4%, 22.4% and 33.2%. Access therefore appeared to track institutional infrastructure rather than trainee seniority or academic orientation.

In a multivariable linear regression adjusting for AI access, seniority, research activity, and institution type, none of the predictors was significantly associated with the total score, and the model explained only 1.0% of the variance (R² = 0.010; model p = 0.766). Access to AI retained a positive but non-significant coefficient (B = 0.67, β = 0.08, p = 0.208). Full regression results are provided in Additional file 2.

Discussion

In this national survey of 286 radiology residents, the Turkish ATRAI-14 did not behave as a unified scale: only the Trust domain showed acceptable internal consistency, while the Implementation Perspectives and Hopes and Fears domains lacked cohesion, and the original three-factor structure was not reproduced. Alongside this partial psychometric replication, the survey revealed a coherent attitudinal profile — residents endorsed AI as a collaborative aid but resisted its autonomy, anticipated efficiency gains yet feared professional devaluation, and reported limited clinical access to AI despite considerable research interest.

Why did the Trust domain hold together when the other two did not? A likely reason is that the Trust items ask about something concrete and familiar from daily work, namely whether a radiologist would rely on a particular kind of AI output. Residents can answer such items from direct clinical reasoning, which likely explains their internal consistency and their close tracking of the overall VAS rating (Trust-domain ρ = 0.45; total-score ρ = 0.56). The Implementation Perspectives domain, by contrast, asks respondents to judge financing models, infrastructure readiness, and the future role of the profession — matters that depend on health-system context and policy awareness rather than personal clinical experience. In a setting where only 21.7% of residents reported access to AI tools, responses to these items may reflect speculation rather than grounded judgment, producing the near-zero internal consistency we observed. The cross-cultural literature anticipates precisely this pattern: items tied to concrete behavior tend to transfer across languages more faithfully than items tied to abstract or system-dependent constructs, which is why independent re-validation is recommended whenever an instrument is moved to a new linguistic and organizational context [20, 21, 26].

Cross-cultural non-equivalence is, however, only one of several explanations that merit consideration, and we do not present it as the sole account. Measurement and structural factors are likely to contribute: the Implementation and Hopes-and-Fears domains each bundle conceptually heterogeneous content — who should fund AI, which functions are most useful, which modalities will be affected, and what role the radiologist will play — and Cronbach’s α presumes interchangeable, unidimensional items, an assumption such multi-topic domains do not meet irrespective of language. Methodological factors also apply, since the exploratory solution was derived in the same sample as the confirmatory test and the sampling adequacy was only mediocre. Translation-related factors cannot be excluded with certainty, although the back-translation reported here identified no unintended semantic drift in the scored items. Population-specific factors are plausible as well: with only 21.7% of residents reporting access to AI tools, system-level and career-level items may have been answered from expectation rather than experience. Finally, the scoring logic itself contributes, since the inverse conversion applied to item I4 means that anticipating broader AI impact lowers the score. We regard these explanations as complementary rather than competing, and the relative contribution of each cannot be resolved by the present design. Two items, I4 and H4, loaded negatively on their assigned factors, meaning they relate to their domain in the direction opposite to the original design. This is not merely a statistical artifact but is informative about how Turkish residents interpreted these questions. A reverse-coding error can be excluded as an explanation: in the source instrument every scored item is oriented in the same direction, with higher values denoting a more positive attitude, and the developers reported no significant negative inter-item correlations among the scored domains [19]; the published scoring was applied here without modification and verified item by item, so no reverse-coded item exists that could have been mishandled. Item H4 asks whether AI will change radiologists’ workload; in the original instrument, a reduced workload is scored as a positive attitude [19]. Notably, H4 was already the weakest item in the original validation, with a loading of only 0.35 [19], so its instability in a new population is perhaps unsurprising. In our cohort, the interpretation appears to have shifted: a reduction in workload may be read not as a welcome relief but as an early signal of professional redundancy, aligning it with fear rather than optimism. Item I4 follows a counterintuitive scoring logic, in which selecting fewer affected modalities yields a higher score. Residents who anticipated that AI would affect a broader range of modalities were therefore assigned lower scores, even though identifying more affected modalities may reflect greater, rather than lesser, engagement with the technology. These reversals suggest that the items most sensitive to professional anxiety and to the scoring direction itself are the least portable across cultures.

Placed alongside the original validation, the divergence is instructive. Vasilev and colleagues reported a Cronbach’s alpha of 0.78 and a confirmatory model with excellent fit (CFI 0.95, RMSEA 0.049) in a sample of 90 predominantly experienced radiologists, 77% of whom had access to AI tools [19]. Our larger sample of residents, working in a system with far lower AI penetration, yielded a total alpha of 0.56 and inadequate confirmatory fit. The two studies also differed in their central tendency, with a higher median total score in our cohort than in the original. This gap likely has several sources. Seasoned radiologists and residents differ in their clinical exposure and sense of professional security, and the radiologists in the original study worked in a setting where AI tools were far more available than in ours. There may also be genuine cross-cultural differences in how professional prestige, remuneration, and workload are understood. Rather than indicating a flawed instrument, these differences underscore that psychometric properties established in one population cannot be assumed to hold in another, and that the ATRAI-14 requires domain-level refinement before it can serve as a validated scoring instrument for Turkish residents. Accordingly, findings from the Implementation Perspectives and Hopes and Fears domains should be read as observations about individual items rather than as measurements of the constructs they were designed to capture, and domain-level scores from them should not be used to compare groups.

Beyond its measurement behavior, the survey offers a substantive portrait of where Turkish radiology residents stand. The clearest signal was a preference for AI as an assistant rather than an autonomous agent. Residents were nearly unanimous in saying they would re-examine a study when an AI tool disagreed with their reading, yet only about a fifth were comfortable with AI interpreting images without oversight. This pattern echoes findings from resident surveys in other countries, where enthusiasm for AI as a support tool coexists with caution about ceding diagnostic authority [13, 27]. A second, more striking signal was an efficiency–devaluation tension within the Hopes and Fears domain: most residents expected AI to lighten their workload, yet a majority simultaneously feared a decline in the profession’s prestige, and a substantial minority anticipated lower earnings. The same respondents who welcomed relief from routine tasks appeared to worry that the relief itself might erode their standing. This ambivalence resonates with the fear-of-replacement literature, in which apprehension about AI is closely tied to how much residents know about it and how it is framed during training [28, 29].

These observations carry practical implications for radiology residency education, which is the central concern of programs preparing the next generation of radiologists. The gap between high research engagement (43.7%) and low clinical access to AI (21.7%) suggests that, for many Turkish residents, AI remains an academic subject rather than a daily tool. Curricula that move AI from the seminar room to supervised clinical use may do more to shape balanced, realistic attitudes than theoretical instruction alone, a position consistent with calls to integrate structured, hands-on AI training into residency programs [11, 30–32]. Addressing the professional-anxiety dimension matters just as much. Education that openly discusses how AI is likely to reshape the radiologist’s role, without necessarily replacing it, may help turn vague apprehension into informed engagement. Because attitudes did not differ meaningfully by seniority, institution type, or research involvement in our cohort, such educational efforts may be broadly applicable across training stages rather than targeted to a particular subgroup.

The strengths of this study lie primarily in the scope and composition of its sample. To our knowledge, this is the first cross-cultural adaptation of the ATRAI-14 and the largest survey of radiology residents’ attitudes toward AI conducted in Türkiye, drawing 286 respondents from 25 institutions across 17 cities and all seven geographical regions. This breadth contrasts with the original validation, which was based on 90 predominantly senior radiologists from a single metropolitan health system [19], and it lends greater confidence and generalizability to our psychometric observations: the finding that two domains lack internal consistency is far more difficult to attribute to sampling limitations in a large, geographically diverse cohort than it would be in a small one. Moreover, by focusing specifically on residents and capturing a range of training years, the study characterizes attitudes during the formative period in which professional identity and openness to new technology are still taking shape — a population of particular relevance to medical educators, yet underrepresented in the original instrument, where residents constituted only 17% of respondents [19]. Including residents across all training years also let us examine whether attitudes vary with seniority. Junior and senior residents reported closely similar scores, with no significant difference between them, suggesting that attitudes toward AI are already established early in training rather than evolving with experience. This comparison would not have been possible in the original validation sample, which consisted predominantly of experienced radiologists [19].

Limitations

The study has several limitations. Its cross-sectional design precludes any causal inference, including about the relationship between AI access and attitude, which did not reach statistical significance. While the overall sample was large and geographically diverse, the specific subgroup reporting AI access was comparatively small (n = 62 versus n = 224), and the confidence intervals around the corresponding effect sizes were correspondingly wide. Recruitment through professional networks and voluntary participation introduces a risk of self-selection bias, whereby residents more interested in AI may have been more inclined to respond; the same limitation was acknowledged in the original instrument’s development [19]. Because the invitation was relayed within departments rather than to a fixed roster, the number of residents who received it could not be determined and a response rate could not be calculated; participation was also self-initiated, so the reported attitudes may not fully represent the broader population of Turkish radiology residents. Because age and gender were deliberately not collected, the study cannot assess whether attitudes toward AI vary with either, although both are recognised correlates of technology adoption and acceptance; this also limits direct comparison with prior surveys that report them, and future studies should collect these variables under appropriate confidentiality safeguards. Several measurement properties were not evaluated and remain unknown for the Turkish version: test–retest reliability and therefore temporal stability; measurement invariance across seniority, institution type, or level of AI exposure; responsiveness to change following training or exposure; and convergent and discriminant validity against other established instruments. Criterion validity was assessed only against a single self-reported global rating rather than against behavioural or objective indicators. Item re-specification and alternative scoring approaches — for example a Trust-only or unidimensional score — are a logical next step, but would require an independent sample. Back-translation was carried out during revision rather than as part of the original adaptation protocol, and neither cognitive interviewing nor pilot testing was performed; the adaptation therefore falls short of a full Beaton-type procedure, and this should be weighed when interpreting the findings. Finally, the psychometric findings themselves constrain any future use of the scale: with two of three scored domains lacking internal consistency, the total score should be interpreted with caution in this population. Beyond refining the culturally unstable items and adding test–retest assessment, the more important challenge is to understand why attitudes tied to implementation and career prospects are so difficult to measure consistently across health systems. This is a question that matters well beyond the Turkish setting.

Conclusions

Among Turkish radiology residents, the three-domain structure of the ATRAI-14 transferred only in part. The Trust domain performed well, whereas the Implementation Perspectives and Hopes and Fears domains did not cohere. This dissociation carries a message that reaches beyond a single country. Attitudes rooted in direct clinical experience appear to travel well across cultures, but those bound up with health-system context and professional uncertainty may not, and an instrument that measures them cannot be assumed to remain valid once it is moved to a new setting. A further signal runs through the data. These residents are keenly interested in AI yet have had little chance to use it in practice. Closing that distance by building supervised, hands-on AI experience into training could help the next generation form its views through experience rather than expectation — a possibility that warrants direct evaluation rather than a conclusion that follows from cross-sectional data.

Abbreviations

AI

Artificial intelligence

ATRAI-14

Attitude Toward Radiology Artificial Intelligence (14-item)

AIIC

Average inter-item correlation

CFA

Confirmatory factor analysis

CFI

Comparative fit index

CI

Confidence interval

EFA

Exploratory factor analysis

GFI

Goodness-of-fit index

I-CVI

Item-content validity index

IQR

Interquartile range

KMO

Kaiser–Meyer–Olkin

RMSEA

Root mean square error of approximation

SD

Standard deviation

TLI

Tucker–Lewis index

VAS

Visual analog scale

α

Cronbach’s alpha (coefficient of internal consistency)

ω

McDonald’s omega

ρ

Spearman’s rank correlation coefficient

Authors' contributions

Authors’ contributions [EO] conceived and designed the study, performed the cultural adaptation, analysed and interpreted the data, and drafted the manuscript. [MB] contributed to data collection and data interpretation, and critically revised the manuscript. All authors read and approved the final manuscript.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The Turkish-adapted questionnaire, the original English ATRAI-14 items, and the scoring algorithm are openly available in the Zenodo repository (https://doi.org/10.5281/zenodo.19039735). The analysis code that scores the questionnaire and reproduces every reported statistic directly from the raw responses is provided as Additional file 4. The dataset analysed during the current study is not publicly available, because participants were informed at the point of consent that their responses would not be shared with third parties; it is available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki and approved by the Kastamonu University Clinical Research Ethics Committee (approval date: 18 December 2025; reference no: 2025-74). All participants provided electronic informed consent before accessing the questionnaire; participation was voluntary and consent was indicated by an affirmative response without which respondents could not proceed. The measures taken to protect respondent anonymity are described in the Methods.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.