Common Applications of ML Algorithms
Radiologist’s Intro to Machine Learning — 10 Part Series
Part 4
Authors: Danilo Pena, Dr. Ty Vachon
Editor: Dr. Michael Doxey
The use of machine learning is more ubiquitous than you might think. From determining whether or not a call or email is spam, to recognizing the boxes around your friends’ faces and actually determining who they are, or targeted marketing for advertising, companies are quick to use their data to increase revenue, traction, and attention on their webpage.
These moments in time are a part of a greater technology narrative that is quickly disrupting how businesses do businesses. Radiology is no different.
This article will introduce some of the real-world applications of machine learning. But before we do that, let’s take a look at this image below from Invidia.
As you can see, deep learning is a subset of machine learning, and machine learning is a subset of artificial intelligence. This natural progression of sub-fields can be seen as one field building upon another, and everything that is done around image recognition can trace back its roots to the early days of artificial intelligence. It is important to understand this hierarchy as many people (including us in this series) throw these words around interchangeably — when in fact, they have a more nuanced relationship.
Image recognition is one of the main advances in artificial intelligent algorithms. Facebook can detect your face, your friend’s face, and probably even your dog, and then assign the appropriate name tag!
Image recognition or analyzing objects within images has advanced rapidly due to the use of a subset of learning algorithms called convolutional neural networks. We don’t go in depth in this article about what that means, but just remember that it is a way to find details that make up an object.
For example, if you would like to detect a face — the algorithm would first need to learn how to detect edges. Once the edge detection gets good, you can maybe detect angles or arcs. Then, you can detect shapes such as ovals or spheres.
This build-your-face type of learning, if you will, is what happens underneath the hood of these convolutional neural networks (otherwise known as CNNs). These networks are a subset of deep learning algorithms. So, when you or your organization starts to add these functions to the workflow, there is a high chance that the developers use CNNs for the imaging use cases! These use cases can include anything from face detection, to object detection or even image segmentation.
Speech recognition is a very fun and interesting application of deep learning. Speech is one of those natural steps towards interacting with technology since we can speak faster than we type, and it is the natural way of communicating.
With the advent of Siri, Alexa, Google Home and others, many companies across the world are using machine learning and deep learning to take in vast amounts of speech data to tease out the nuances in sound.
Remember that there always needs to be a ground truth to the problem. Speech is a very easy problem at that point since people can be hired to physically listen to hours and hours of sound bites, transcribe them, and push this knowledge into an algorithm. From there, the algorithm will learn the nuances in sound, voice inflection, and relate these different features to probabilistically determine what word is the most likely.
Recommendation systems are a third set of use cases for machine learning. These applications have been the bread and butter for many companies. When we talk about recommendation systems, we are referring to the targeted advertising on your Facebook page, the recommended products to buy on Amazon, and even the recommended movies or shows to watch on Netflix.
These systems are powered by machine learning algorithms that have detected nuances in human behavior — whether it be purchasing food or watching the last season of The Office. Moreover, these systems have not only proven to work but are now the backbone to a lot of dollars spent on advertising and marketing. Whatever ad you click or how long you spend watching a show are small pieces to the larger algorithm that is running these companies recommendation systems.
In this article, we scratched the surface on the multitude of applications that machine learning and artificial intelligence has. It is only a matter of time when manual transactions will be automated. However, with all of the wonderful pros that we have outlined, there are barriers that need to be figured out before more progress is to ensue. These barriers, specific to radiology, will be discussed in a later section.
For now, let’s dig into some common machine learning topics to get you up to speed. This include feature extraction, labeled data, and validation sets! All explained in Article 5.
