Real-Time Artificial Intelligence Assistance for Peripheral Nerve Ultrasound Training in Radiology Residents: A Randomized Crossover Study (Preprint)

BACKGROUND Peripheral nerve ultrasound is technically demanding for novice operators and requires precise anatomical knowledge and image interpretation skills. Real-time artificial intelligence (AI) assistance may facilitate nerve recognition and support ultrasound training. However, its impact on trainee performance and perceived educational value remains insufficiently investigated. OBJECTIVE To evaluate whether real-time AI assistance improves peripheral nerve identification performance among radiology residents and to assess its potential educational benefits and limitations. METHODS This prospective randomized crossover study included 15 radiology residents with no previous experience in peripheral nerve ultrasound. Participants performed bilateral identification of the median, ulnar, and radial nerves, with and without real-time AI assistance using NerveTrack (Samsung). Each resident completed 12 randomized identification attempts, yielding 90 matched comparisons. The primary outcome was successful nerve identification within 60 seconds, assessed by senior musculoskeletal radiologists. Secondary outcomes included identification time, confidence, perceived difficulty, AI overlay accuracy, anatomical scanning location, and resident and faculty perceptions. RESULTS Fifteen residents completed 180 identification attempts. Successful nerve identification increased from 52.2% (47/90) without AI to 65.6% (59/90) with AI assistance (exact McNemar P=.036), although the resident-level sensitivity analysis was not statistically significant (P=.125). Mean identification time decreased from 34.1 (SD 9.7) to 25.3 (SD 8.1) seconds (P=.001). Confidence increased from 4.6 (SD 1.8) to 6.4 (SD 1.1) on a 0–10 scale (P<.001), while perceived difficulty decreased from 5.8 (SD 1.7) to 4.6 (SD 1.7) (P=.001). Appropriate anatomical scanning location remained unchanged (74.4% in both conditions). The largest numerical improvement in identification success was observed among first-year residents, although differences across training levels were not statistically significant. AI overlay accuracy varied according to anatomical target, highlighting the importance of expert supervision. CONCLUSIONS Real-time AI assistance was associated with faster peripheral nerve identification, increased confidence, and reduced perceived difficulty among radiology residents. Although identification success improved at the matched-attempt level, this finding was not statistically significant in the resident-level sensitivity analysis. These results suggest that AI-assisted ultrasound may support supervised training, but further studies are needed to determine whether these immediate performance benefits translate into sustained, independent ultrasound skills. CLINICALTRIAL No