Only 9% of AI-Generated Pediatric Neuroradiology Images Deemed Teachable

Only 8.6% of synthetic pediatric neuroradiology images generated by widely available AI applications were rated as suitable for teaching by >50% of pediatric radiologists. GPT-5 outperformed Gemini 3 Pro (12.7% vs 4.5%, P=0.031). Inaccuracies remain common, so these tools should…

• Reader study: 220 images from GPT-5 and Gemini 3 Pro across 22 pediatric neuroradiology scenarios, assessed by 8 fellowship-trained pediatric radiologists.
• GPT-5 produced more acceptable images (12.7% vs 4.5%, P=0.031), but overall quality remains poor; inaccuracies may impair learning.
• Single-center, limited to two specific AI models; rapidly evolving technology may quickly date results.

RadPigeon summaries are original and for information only. They are not clinical advice.