Does artificial intelligence for classifying ultrasound imaging generalize between different populations and contexts?

In a recent publication in this journal, Drukker et al.1 reviewed the role of Artificial Intelligence (AI) in ultrasound imaging in Obstetrics and Gynecology. The authors describe the application of AI algorithms for automated detection and classification of standard planes, among other applications.

One particular challenge is that large amounts of ultrasound images are required to train these AI algorithms. The way the resulting algorithms are trained carries a risk of introducing bias. Secondly, a potential problem arises in the application of algorithms outside the context of the population of data on which the algorithms were trained and validated.

There is currently insufficient evidence that AI algorithms generalize from the population they are trained on to other populations. It is problematic for the general adoption of research results if AI algorithms fail to generalise across different settings. We here describe how an AI algorithm developed in a British setting and using data from a British population in 2016 performs with images obtained in a Danish population across two fetal medicine centers between 2009 and 2017.