Human Evaluation of Synthetic Videostroboscopic Laryngeal Images Generated Using StyleGAN3.

Abstract

To evaluate the perceptual realism of synthetic videostroboscopic laryngeal images and determine the relationship between training duration and clinician classification accuracy. Synthetic images were generated using StyleGAN3 from two age-stratified datasets: Dataset A (≥ 65 years) and Dataset B (< 65 years). A total of 114 clinicians evaluated a randomized, balanced-controlled 36-image survey drawn from a 144-image pool containing real frames and five StyleGAN3 training intervals (5120-25,000 kimg). Accuracy was analyzed across image classes, experience levels, and viewing devices. Clinicians demonstrated significantly higher mean accuracy identifying true clinical frames than synthetic generations (70.9% vs. 57.8%, p < 0.001). Overall survey accuracy was 59.9% ± 14.2%, with realism peaking at 20,120 kimg (44.8% accuracy). Accuracy dropped significantly between 5120 kimg (79.6%) and 10,120 kimg (59.1%, p < 0.001), with diminishing returns thereafter. No significant accuracy difference existed between Datasets A and B (p = 0.27). The effect of specialty reached borderline significance (p = 0.059), though limited by severe subgroup imbalances (e.g., n = 5 fellows, n = 2 residents). Computer users (63.9%) were significantly more accurate than phone users (56.2%, p = 0.014). A confounder analysis confirmed specialty and device choice were statistically independent (p = 0.164). StyleGAN3 produces synthetic laryngeal images demonstrating high perceptual realism. The significant real-versus-synthetic accuracy gap highlights that while true frames remain distinct, synthetic generations achieve profound ambiguity. Perceptual realism plateaus early (10,120-20,120 kimg), suggesting moderate training durations can produce human-perceived realism while reducing computational and environmental demands. N/A.