Deep learning applied to medical image analysis: Epistemology and data.
Deep learning (DL) is changing the way we analyze medical images. What do these changes consist of? I will introduce some interpretations according to which AI epistemology is moving towards a new paradigm called the "fourth paradigm" in which theory, hypothesis and experiment are going to be unified through the data. In this context, the famous statement "Correlation is enough" seems to be an effective way to describe this evolution. This change raises several issues related to medical image analysis that can be discussed. One of the most interesting I want to speak about is data. Data are a main part of DL development especially because DL algorithms are not explainable. Medical images data sets are scarce, underpopulated and usually do not contain acquisition and reconstruction parameters, which could be helpful to harmonize multicentric data. It is only knowing data characteristics that we can define the boundaries in which an algorithm can properly work. Mean performances on very diverse data are not reliable and correlation is not enough if data are complex and models are opaque. The research of causation is an interesting point to be discussed among physicists.