Machine Learning Performance Metrics and Diagnostic Context in Radiology

Abstract

In this pilot study data gathered from interviewing specialists in radiology is combined with an assessment of the way machine learning metrics are used in studies of radiological work. It argues that situated context of use should be an important contributor to the design of machine learning applications in radiology. The article shows how radiologists see their professional practice as utilizing a wider range of expert knowledge than many existing studies on machine learning in radiology allow for. The article describes a case study drawn from radiology practice in a major Danish hospital and discusses a widely cited study on machine learning in radiological work. The study connects current understandings of appropriate metrics used by machine learning researchers with professional radiologists' understanding of their diagnostic work. This comparison helps identify gaps in understanding between these two communities and suggests how they might be addressed.