Multi-view AI output variability as a teaching resource in a micro-course for thyroid TI-RADS interpretation among ultrasound residents: a randomized controlled trial.

Authors

Abstract

Thyroid TI-RADS interpretation requires integration of sonographic features across image planes, yet routine teaching often emphasizes the final category more than the cross-plane reasoning process. We evaluated whether a structured micro-course using multi-view AI output variability as a teaching resource could improve TI-RADS interpretation among ultrasound residents. In this single-center randomized controlled trial, 52 ultrasound residents were randomized 1:1 to either a structured micro-course using multi-view AI output variability as a teaching resource or time-matched conventional TI-RADS teaching. The primary outcome was participant-level TI-RADS interpretation accuracy at baseline (T0), immediately after teaching (T1), and 4 weeks later (T2). Secondary outcomes were inter-learner consistency, representative plane hit score, and self-efficacy. Justification quality in discordant cases was analyzed exploratorily. Baseline characteristics were comparable between groups. Accuracy showed a significant group-by-time interaction [<i>F</i> (2, 100.0) = 10.86, <i>P</i> < 0.001]. There was no between-group difference at T0 (<i>β</i> = 0.08, 95% CI: -0.74 to 0.89; <i>P</i> = 0.851), but the intervention group performed better at T1 (<i>β</i> = 1.69, 95% CI: 0.88-2.51; <i>P</i> < 0.001). The between-group difference at T2 was not significant (<i>β</i> = 0.73, 95% CI: -0.08 to 1.55; <i>P</i> = 0.078). At T1, the intervention group also had a higher representative plane hit score (15.00 [12.00-17.00] vs. 13.00 [11.00-15.00]; <i>P</i> < 0.001), higher inter-learner consistency at T1 and T2 (ΔAC2 = 0.08 and 0.09, respectively; both 95% CIs excluded zero), and a greater increase in self-efficacy (<i>β</i> = 0.27, 95% CI: 0.09-0.45; <i>P</i> = 0.006). Exploratory analysis showed more favorable justification quality in the intervention group (OR: 10.26, 95% CI: 5.36-19.67; <i>P</i> < 0.001). A structured micro-course using multi-view AI output variability as a teaching resource improved immediate TI-RADS interpretation accuracy, but the between-group advantage was not sustained at 4 weeks, despite persistently higher inter-learner consistency. These findings support the short-term educational value of the structured micro-course rather than an independent effect of AI output variability and warrant evaluation of repeated or booster instruction.

Tags

Topics