Back to all papers

Widely available, general-purpose generative AI applications for radiology education: determining the teaching quality of synthetic pediatric neuroradiology images.

July 6, 2026pubmed logopapers

Authors

Reith TP,D'Alessandro DM,Modi KA,Sabawi M,Zadeh CS,Sebaaly MG,Lai LM,Sato TSP,Kao SC,Sato Y,D'Alessandro MP

Affiliations (3)

  • Department of Radiology, University of Iowa Hospitals and Clinics, Iowa City, IA, USA. [email protected].
  • Department of Radiology, Northwestern University Feinberg School of Medicine, 676 N. St. Clair St., Suite 800, Chicago, IL, 60611, USA. [email protected].
  • Department of Radiology, University of Iowa Hospitals and Clinics, Iowa City, IA, USA.

Abstract

Artificial intelligence (AI) is increasingly incorporated into medical education. Widely available, general-purpose generative AI applications are increasingly capable of producing high-fidelity synthetic medical images, which may have potential educational uses. To assess the ability of widely available, general-purpose generative AI applications to produce synthetic pediatric neuroradiology images of teaching quality. Ten significant diagnostic errors in pediatric neuroradiologic emergencies were identified and divided into 22 individual clinical scenarios. Two widely available, general-purpose generative AI applications, GPT-5 (OpenAI, San Francisco, CA) and Gemini 3 Pro with the Nano Banana Pro image-generation engine (Google, Mountain View, CA), were prompted to generate five separate images for each scenario, yielding a total of 220 images. Eight fellowship-trained pediatric radiologists independently assessed each image for teaching quality. Ratings were recorded and analyzed. Differences between applications were evaluated using Fisher's exact test. Overall, 8.6% of generated images (19/220) were deemed suitable for teaching by more than half of the pediatric radiologists. GPT-5 produced a higher proportion of acceptable images meeting this >50% criterion compared to Gemini 3 Pro (12.7% [14/110] vs. 4.5% [5/110]; P=0.031). At present, widely available, general-purpose generative AI applications should not be used to produce synthetic radiological images for educational purposes. Inaccuracies and inconsistencies limit reliability and may actively impair learning.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAI Slice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.