Combining Radiological Expertise with Artificial Intelligence (AI) in Diagnostic Imaging

AI tools for diagnostic imaging could reduce costs while providing more equitable access to high-quality radiology services. Despite rapid technical progress, however, human input is going to be indispensable in the foreseeable future for a variety of technical, legal and regulatory reasons. This raises important questions about how humans and AI can and should collaborate. To understand such collaboration, we propose to (i) investigate whether human radiologists suffer from automation bias or neglect when combining clinical contextual information with AI predictions, (ii) quantify the value of contextual information not available to AI tools, and (iii) design human-AI collaboration setups that improve diagnostic accuracy. The proposed research uses a novel experimental design and a model of human decision-making to study the interpretation of chest X-rays by tele-radiologists. The experiment randomizes tele-radiologists into varying information environments and elicits their diagnostic assessments and recommend treatment/follow-up conditions. The information environments vary whether AI predictions are provided, whether contextual information is hidden or revealed, and the form of AI prediction.