Explainable AI in nuclear medicine.

Abstract

In this short communication, we consider the need for explainable AI from the perspective of a large multi-disciplinary research project for predicting cachexia in cancer patients. In a series of meetings, comprising expertise from medicine, data science, sociology, and philosophy, project participants discussed the need for explainability. We distinguish between contexts in which a black box AI tool undertakes tasks that users can perform or validate themselves and contexts in which this is not the case. We conclude that explanations are likely required when a black box AI tool undertakes tasks that users cannot perform or validate themselves. If the user can verify outputs manually, documented reliability and accuracy may suffice, but explainability can still add value when outputs are uncertain or errors occur. More generally, close collaboration among physicians, AI developers, and other stakeholders is crucial to ensure that AI tools are trustworthy and useful in clinical practice.