Artificial Intelligence Explained for Nonexperts

Abstract

Artificial intelligence (AI) has made stunning progress in the last decade, made possible largely due to the advances in training deep neural networks with large data sets. Many of these solutions, initially developed for natural images, speech, or text, are now becoming successful in medical imaging. In this article we briefly summarize in an accessible way the current state of the field of AI. Furthermore, we highlight the most promising approaches and describe the current challenges that will need to be solved to enable broad deployment of AI in clinical practice.

AI MSK clinical applications: spine imaging

“…AI is the broadest umbrella term for a series of computer algorithms that appear "intelligent" to humans and covers a diverse spectrum of algorithms [10,11].…”

“…Currently, the most dominant approach to AI is machine learning (ML) and builds on the principle of learning from data [11]. ML algorithms essentially require input data, e.g., images in the context of radiology.…”

“…ML algorithms are often categorized into unsupervised, supervised, and reinforcement learning methods, and each category requires a different type of input data [11]. Specifically, data can be labelled or unlabelled -unlabelled data contains purely the raw information (for example, a CT or MRI scan), whereas labelled data has been annotated or classified into a specific category (for example, a diagnosis or a segmentation has been made on the scan by a radiologist).…”

“…Recent developments in a subgroup of unsupervised learning (so-called self-supervised learning models) [13] have attempted to use partly labelled data to make predictions; however, the research is still in its infancy. Thus, supervised learning is still the most frequently used ML method in AI practice in general, and specifically for medical applications [11]. Supervised learning uses labelled data, and thus, it requires a professional radiologist or a separate pre-trained ML algorithm to create a labelled dataset.…”

“…The output of ML algorithms largely depends on the quality and selection of input features ("garbage in -garbage out" [14,15]). This is one of the major issues in applying ML for image classification tasks, as it is very difficult to generate the correct labels for images from unprocessed input pixels or voxels [11]. The difficulty of "supervising" this transformation is the most important trigger for the increasing popularity of a new class of ML called deep learning (DL), which aim to mimic more intelligent behavior through encompassing multiple stages, or layers, in an architecture known as a neural network (NN).…”