Abstract
Artificial intelligence (AI)-assisted systems are increasingly used in coronary computed tomography angiography (CCTA), but their role in resident training and subsequent unaided interpretation remains unclear. This multicenter randomized pilot feasibility study evaluated a functionally standardized AI-assisted CCTA learning workflow with longitudinal unaided assessment in radiology resident training. Ten novice radiology residents from three centers were randomized 1:1 to conventional workstation-based or AI-assisted workflow training. Locally available AI platforms were required to meet predefined core-function criteria. After standardized instruction, residents completed a 4-week learning phase with next-day standardized disclosure of invasive coronary angiography (ICA) reference results and no additional expert case-by-case teaching. Formal assessments were independently performed without any support at baseline, after 2 and 4 weeks of learning, and 1 week after learning completion. Feasibility outcomes included completion, protocol adherence, data completeness, multicenter implementation, and collection of prespecified exploratory metrics. Exploratory metrics included diagnostic time, confidence, inter-rater consistency, agreement with ICA, and diagnostic discrimination for ICA-defined stenosis ≥70%. All residents completed training and scheduled assessments, with no withdrawals or missing assessment data. The workflow was implemented across all centers, and all exploratory metrics were collected at each time point, with descriptive learning-related changes in both groups. In conclusion, functionally standardized AI-assisted CCTA learning with longitudinal unaided assessment was feasible in structured multicenter resident training, supporting larger trials of training effectiveness.