A Quick Primer on AI Terminology

A Quick Primer on AI Terminology

Becoming familiar with concepts that underpin the technology

Author: Dr. Will Guest, member of the CAR AI Working Group


Dr. Will Guest

Radiologists are accustomed to using a specialized vocabulary to describe findings precisely and communicate them to other radiologists, referring physicians, and patients. As artificial intelligence (AI) technology expands and eventually becomes a part of clinical workflow, radiologists will need to become familiar, at least in a general sense, with the concepts that underpin the technology. Understanding the terminology of AI is a good place to start, so here are a few phrases that may soon enter the radiology lexicon:

Machine learning has become something of a catch-all term for computer automation of a cognitive function previously performed by humans. However, in distinction from a computer program that performs a fixed set of operations on input data, a machine learning algorithm should incorporate the ability to improve its performance at its assigned task as more data is made available to it. This is accomplished by the iterative tuning of parameters within the algorithm to optimize the goodness of fit between the attribute of interest in the data (i.e. the features) that the algorithm seeks to classify and the output (i.e. the prediction) of the algorithm. This process may occur either via supervised learning, in which a test set of data has human-applied labels giving the desired output of the algorithm, or unsupervised learning, in which no labels are applied to the data and the algorithm seeks to cluster or organize the data to reveal underlying patterns.

A variety of computational tools are available to implement the machine learning concept, but the technique with the greatest current interest is the artificial neural network, named from the analogy with interconnected neurons in a biological nervous system. The building-block of a neural network is the artificial neuron, which is a relatively simple function that takes a variety of inputs (either from the data supplied to the network, or from other artificial neurons), and determines whether the neuron should activate (and in some models, how much it should activate) by weighting these inputs against a bias that determines the excitability of the neuron. The output of this neuron may become the input for other neurons or be aggregated with the output of other neurons (via mathematical operations like a softmax function that amplifies the strongest signal from the group of neurons) to produce the final output of the network.

The complexity of neural networks arises from the large number of constituent neurons and the organization of neurons within the network. In a common neural network design, neurons are organized into groups called layers, which receive inputs from neurons in an earlier layer and send their outputs to neurons in a later layer (without direct connections between neurons in the same layer). Such a network design is called a feedforward neural network in that there is a unidirectional flow of information in the network, from the network input through a sequence of intermediate layers, to the final output of the network. In contrast, a network in which there is bidirectional flow of information between neurons is called a recurrent neural network. The intermediate layers that neither directly accept the input data for the network nor produce its final output are termed hidden layers. A network with several hidden layers is said to be a deep network.

In image analysis, convolutional neural networks have become a popular tool for their ability to identify image features over a range of size scales. A neuron in such a network processes inputs from a group of pixels, termed its receptive field (like ganglion cells in the retina), and applies mathematical filters called convolutions that emphasize certain attributes of the image in the receptive field (like lines, edges, or textures). Neuron layers are arranged in a hierarchical fashion, so that early layers have small receptive fields and later layers process pooled input to cover a larger part of the image (such as lesions and entire organs). Deep networks that exhibit this hierarchical organization are considered examples of deep learning.

An unusual but powerful neural network architecture is the autoencoder, which seeks to re-generate as its output the image it received as input. This may seem like a pointless exercise, but the significance arises from configuring the network so that one of the intermediate hidden layers has fewer neurons than either the input or the output layer. The neurons in this hidden layer are trained to respond to the most descriptive attributes of the image, which enables these models to perform feature extraction. Importantly, because the input and output of autoencoders are the same, they do not require human-labelled training data sets and are therefore capable of unsupervised learning.

The applications of machine learning in radiology generally fall into a few categories: detection, in which the goal is to identify a structure within an image (for example a lung nodule); classification, in which an image or lesion within an image is assigned to a category (for example, is pulmonary embolism present or not present on this CT scan?); segmentation, in which a structure of interest is isolated from the remainder of the study (such as defining the boundary of an organ); and registration, which seeks to optimize the alignment between images from different studies (either different time points or different modalities) to ease comparison between the two.

This is just a brief overview of some of the concepts and terminology that play a role in AI radiology applications. I hope it whets your appetite and inspires you to explore this burgeoning field!