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.

Fig. 1

Relationship between different fields: artificial intelligence (AI), machine learning (ML), neural networks (NNs), deep learning (DL), and convolutional neural networks (CNNs).

Fig. 2

Illustration of a simple neural network with fully connected layers. The numbers on the edges are the learnable parameters of the network, and the numbers inside neurons represent their activations. The input to the network is a vector [1, 0, 2] that is transformed by the first layer into [1, 3], then by the following layer into [0, 6], and finally into [6, 0] by the output layer. For illustration purposes, we are assuming the activation values for each neuron of the network are computed as a simple weighted average of the values in the neurons in the preceding layer. In reality, after computing the weighted average, some nonlinear function is applied. This is necessary to allow the network to represent complicated transformations between the input and the output.

Fig. 3

Illustration of a simple convolutional network for a one-dimensional input. The two unique types of layers illustrated here are the convolutional layer and the pooling layer. The special property of the convolutional layer is that it involves applying the same filters at different locations. The convolutional layer in this network contains two filters that transform the input into two feature maps. The pooling layer illustrated above is a max pooling layer that simply computes the maximum of the activations of the neurons in the preceding layer. The same ideas can be easily generalized to two-dimensional and three-dimensional inputs.