Human-AI Complementarity in Diagnostic Radiology: The Case of Double Reading

Abstract We apply the concept of robustness from the philosophy of science to human–AI collaboration in diagnostic radiology, introducing diagnostic complementarity as a way to understand how radiologists and AI systems can productively work together in the context of double-reading. Diagnostic complementarity refers to the idea that two readers (e.g. radiologist and AI) have different diagnostic strengths and limitations, such that their combined performance exceeds that of either working alone. We argue that state-of-the-art AI diagnostic systems—convolutional neural networks (CNNs)—realize diagnostic complementarity with radiologists, and thus can serve as effective second readers for radiologist first readers in double-reading. In the course of making our argument, we clarify diagnostic complementarity and its epistemic benefits, explain how it can be discovered through theoretical and empirical means, and show how our points can be used to evaluate not only the utility of CNNs as second readers, but also of future, more advanced AI systems.