Exploring the use of AI-generated counterfactual chest X-rays to enhance diagnostic learning in medical education

Tools

Mohr, Greta and Zhu, Yifei and Ye, Xujiong and Lennon, Marilyn and MacLellan, Calum and Maclay, John and Lowe, David J. and Sainsbury, Christopher and Dong, Feng and Lagnado, David (2026) Exploring the use of AI-generated counterfactual chest X-rays to enhance diagnostic learning in medical education. BMC Medical Education, 26 (1). 1331. ISSN 1472-6920 (https://doi.org/10.1186/s12909-026-09597-7)

[thumbnail of Mohr-etal-BMCME-2026-Exploring-the-use-of-AI-generated-counterfactual-chest-X-rays-to-enhance-diagnostic-learning]
Preview
 Text. Filename: Mohr-etal-BMCME-2026-Exploring-the-use-of-AI-generated-counterfactual-chest-X-rays-to-enhance-diagnostic-learning.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo
Download (2MB)| Preview

Abstract

Accurate interpretation of chest X-rays is a critical clinical skill, yet radiology training in medical education remains limited and often fails to provide broad exposure to a wide variety of conditions, including those that are rare, subtle, or easily confused with one another. Although artificial intelligence (AI) has shown impressive performance in medical image classification, its potential to actively improve clinical education has not been fully realised. In this study, we introduce AI-generated counterfactual chest X-rays-real patient images digitally altered to show a different but still realistic condition, while keeping the same patient's anatomy. These counterfactuals create 'what if' scenarios within the same patient and offer a novel tool to enhance diagnostic learning, support decision-making, and improve confidence calibration in medical trainees and professionals.Forty-two participants, including medical students and doctors, completed a four-part online study involving diagnostic classification, comparison of counterfactual image pairs, image authenticity judgements, and clinical treatment planning. All image comparisons focused on three clinical conditions: healthy lungs, pneumonia, and pleural effusion. Statistical analyses included t-tests, ANOVAs, Pearson correlations, and linear regression.Results showed that counterfactual images meaningfully supported diagnostic learning. Participants demonstrated improved accuracy and confidence over time, particularly when distinguishing pleural effusion from healthy lungs. Confidence became more closely aligned with accuracy, and participants were better able to recognise condition-specific differences. These comparisons also revealed areas of diagnostic weakness that would be difficult to detect through conventional instruction alone. While real images led to higher raw diagnostic accuracy, counterfactuals proved highly effective in promoting learning and reflective reasoning. Image authenticity influenced clinical decision-making, and initial confidence was a stronger predictor of final treatment confidence than diagnostic correctness.These findings highlight the strong potential of counterfactual imaging to enhance radiology education. By enabling learners to engage with rare, complex, or demographically underrepresented cases in a controlled and consistent way, AI-generated counterfactuals offer a powerful and scalable tool to improve diagnostic preparedness and clinical confidence in future healthcare professionals. [Abstract copyright: © 2026. The Author(s).]

ORCID iDs

Mohr, Greta, Zhu, Yifei, Ye, Xujiong, Lennon, Marilyn ORCID logoORCID: https://orcid.org/0000-0003-3271-2400, MacLellan, Calum, Maclay, John, Lowe, David J., Sainsbury, Christopher, Dong, Feng and Lagnado, David;
CORE (COnnecting REpositories)