What AI is learning or not learning from medical images

The availability of large public datasets and the increased amount of computing power have shifted the interest of the medical community to AI algorithms. However, it can still fail in practice.

This is because there is often no evaluation of different types of data, for example between men and women, people from different demographic regions, or even different hospitals. This leads to AI being biased.

AI can also learn irrelevant features of the image. For example, how the frame of the x-ray looks, leading to incorrect diagnoses if the hospital changes it's x-ray machine.

In this talk we will explain why AI can fail in practice, and give examples from many medical applications. The lecture does not go deeply into technical aspects of AI, but explains some general concepts that should be suitable for many audiences.

(Foto: Shutterstock)