Radiology Reimagined: Advancing Clinical Practice Through AI Innovation (2025)-"Rad Training, Bias, and Bias Mitigation"

"Rad Training, Bias, and Bias Mitigation" – Dr. Stavroula Kyriazi

Pdf Summary

This presentation by Dr. Stavroula Kyriazi addresses the critical role of AI training, bias, and bias mitigation in radiology. AI literacy is essential for providers and users to understand AI systems' opportunities, risks, and impacts, as defined by the AI Act. Effective training covers fundamental AI concepts—such as machine learning, data curation, model validation, governance, and human-AI interaction—to ensure appropriate technical, interpretive, legislative, and ethical use.<br /><br />Training goals focus on equipping radiologists and developers with knowledge of AI algorithms, integration, limitations, and potential biases, fostering patient trust and data privacy. Formats include one-pagers, videos, interactive online modules, case examples, and direct Q&A sessions. Unilabs follows a training pipeline with core vendor-neutral education and solution-specific training, both mandatory before AI deployment, supported by continuous monitoring and refresher sessions.<br /><br />A major concern is bias throughout the AI life cycle—especially automation bias, where radiologists may either over-rely on or dismiss AI outputs, leading to clinical errors or skill deterioration. Studies show automation bias can cause radiologists to accept incorrect AI recommendations, measured by metrics like Error Concordance Rate (ECR) and Discrepancy Adoption Rate (DAR). Surveys suggest best practice involves reviewing AI results alongside CT images collaboratively.<br /><br />Mitigation strategies include comprehensive training highlighting AI limitations and false-positive/negative rates, monitoring AI usage patterns, and maintaining radiologist accountability through annotations and feedback loops. Despite ongoing research and lack of formal guidelines, raising awareness and implementing bias mitigation are vital.<br /><br />Key takeaways emphasize that tailored, locally contextualized training is crucial for safe AI integration in radiology, addressing bias risks and promoting responsible human-machine collaboration to improve clinical outcomes effectively.

Keywords

AI training in radiology

bias mitigation

automation bias

AI literacy

machine learning

model validation

human-AI interaction

radiologist education

AI deployment

clinical error prevention