**3. Artificial Intelligence Paradigms: Machine Learning and Deep Learning**

Although the terms AI, ML, and DL are commonly used interchangeably, each term has its own specific definition. AI is the broad, umbrella term that encompasses both ML and DL, with DL being a subset of ML (Figure 1). Marvin Minsky, an early AI developer, described AI as "the science of making machines do things that would require intelligence if done by men" [55]. AI is the ability of any tool to accept inputs of prior knowledge, experience, goals, and observations and then create an output that implements an action. This definition covers a wide range of tools varying from a simple thermostat to a self-driving car. AI research often falls under the domain of computer science because AI tools perform many computations to create appropriate outputs [56].

**Figure 1.** Relationship between artificial intelligence, machine learning, and deep learning. Artificial intelligence is an umbrella term that includes machine learning and deep learning. Deep learning is a hyponym of machine learning.

Whereas AI typically entails a fixed, rules-based computational method, ML dynamically improves upon computational methods as data is input and trained. In traditional programming, a computer receives data and a program as inputs and then produces the output in a one-to-one manner. All improvements to the results derive from alterations to the program rules. In ML, a computer receives data and labels as inputs and then creates a program to refine the outputs. The computer learns by comparing its own outputs, also known as predictions, to data that has already been defined and associated with a label. Over time, the ML algorithm will improve upon its ability to create a program that can match its own output to a label. The effectiveness of the program is highly dependent on the quality and size of data that the ML algorithm receives as input.

The data types that can be input into an ML algorithm vary widely, encompassing digitized handwriting, text from documents, DNA sequences, facial images, and more. A ML algorithm can utilize this data to train and make predictions. Two of the most common ML implementations are classification and regression [57]. In classification, ML receives data and then decides upon a category for each item in the data. For example, ML could look at images and decide whether the image is a plane, car, or boat. In regression, ML receives data and then predicts a numerical value for each item in the data. Examples include predicting tomorrow's ambient temperature or the price of a stock.

Within the ML discipline, DL has garnered significant attention because of the groundbreaking results that it achieved in the ImageNet Large Scale Visual Recognition Challenge competition, where competitors developed algorithms using a subset of a public dataset of images [58]. DL has flourished with the rise of big data and faster hardware [39]. In traditional ML, the algorithm has features that it will extract from the data before training begins [57] (Figure 2). These features are constant and are based upon established rules. For example, the algorithm can look for eyes when trying to recognize a face or search for wings when identifying an airplane. By contrast, a DL algorithm does not require feature selection before training. DL simply receives input and learns its salient features during training (Figure 2). DL architecture is also notable because it is formed by many tiered layers, which resemble a brain's neuronal network. These layers enable DL to extract features from progressively smaller sizes of input data and allow for increased feature complexity [59]. Although various DL architectures exist, convolutional neural networks (CNN) are considered well suited for medical imaging. The overall goal of these techniques is to allow the machine to determine and optimize features automatically for evaluating and classifying images.

**Figure 2.** Machine learning versus deep learning used for multiparametric magnetic resonance imaging (mpMRI) sequence identification. In machine learning, the computer receives inputs of mpMRI images and goes through feature extraction specific to the different sequences of T2-weighted (T2W), diffusion-weighted imaging (DWI), and dynamic contrast-enhanced (DCE). Then, the computer is trained on additional images and is able to identify the correct sequence as an output. Deep learning differs from machine learning in that feature extraction and training can be done simultaneously to produce the output.

Medical imaging studies that use ML algorithms are frequently designed with three dataset types: training, validation, and test [60]. The study will first use training data as its input to develop an algorithm that produces the desired output. During this training period, the algorithm constantly uses validation data to provide correct feedback to modify itself. After the algorithm has finished

development, final performance is then assessed with test data. Because test data was not used during the algorithm training, it is an objective method to assess performance.
