Association of AI-Assisted Training with Learning Outcomes in Lung Cancer Puncture Ablation: A Retrospective Before-and-After Study.

BACKGROUND: AI-assisted training has been introduced into interventional radiology education, but its association with learning outcomes in lung cancer puncture ablation remains insufficiently defined. METHODS: This single-center retrospective before-and-after study included 69 early-career interventional physicians who underwent six-month refresher training between January 2020 and December 2024. Trainees were allocated by training period: conventional senior physician-guided training before August 2022 (n=38) and AI-assisted training after implementation of the AI planning/visualization system (n=31). De-identified patient-level ablation outcomes were extracted from the institutional electronic medical database under ethics approval and waiver of written informed consent. RESULTS: Compared with conventional training, AI-assisted training was associated with shorter time to competency (median 43 vs 84 days; absolute difference, -41 days) and fewer supervised cases before independent performance (21 vs 31.5 cases; absolute difference, -10.5 cases). Radiation exposure during training was lower in the AI-assisted group (12.8 ± 2.20 vs 21.1 ± 3.78 mSv; mean difference, -8.3 mSv; approximate 95% CI, -9.7 to -6.9). Initial independent procedure time was also shorter (48.7 ± 8.28 vs 62.2 ± 5.92 min; mean difference, -13.5 min; approximate 95% CI, -17.0 to -10.0). Complete ablation rates were similar (80.7% vs 81.3%). Pneumothorax (5.0% vs 12.5%; risk difference, -7.6 percentage points; approximate 95% CI, -10.1 to -5.1) and hemoptysis (12.7% vs 16.9%; risk difference, -4.2 percentage points; approximate 95% CI, -7.6 to -0.9) were less frequent in the AI-assisted cohort. CONCLUSION: AI-assisted training was associated with faster acquisition of procedural independence, lower occupational radiation exposure, shorter early procedure times, and fewer selected minor-to-moderate complications, while maintaining comparable ablation efficacy. Because of the retrospective non-randomized before-and-after design, these findings should be interpreted as associations and require confirmation in prospective multicenter studies.