AI in nuclear medicine - what, why and how?

Abstract

There have been various attempts to define artificial intelligence (AI), and none is sufficiently precise but at the same time universally applicable. However, in the context of medical imaging, the term machine learning (ML), which is generally considered a subset of AI [1], may describe most applications more appropriately. Here, “learning” relates to the capability of systems to identify complex relationships between data and to predict outcomes in new and unknown data with similar characteristics. With the computing power available today, ML has advanced from classical ML methods, such as decision trees or support vector machines, to more complex architectures, such as deep learning. This uses “deep” artificial neural networks, which are characterized by multiple layers of artificial neurons [2]. In several applications in medical imaging, deep learning has been found to be equivalent or superior to classical ML methods [3, 4, 5], and it is now the most commonly used ML approach for such tasks. Deep neural networks, and especially convolutional neural networks (yet another subset), are inherently useful for the numerous “visual tasks” involved in image analysis.