Introduction to AI for medical imaging

Hello, my name is Alan McMillan, I'm a professor at the University of Wisconsin in Madison. I'm here today to talk to you about the use of artificial intelligence in medical imaging. Thank you for joining.

In this presentation, I will cover the following topics related to artificial intelligence applied to medical imaging. First, I will discuss the machine learning paradigm and provide a very brief overview of current techniques in deep learning that are used to solve various problems in medical imaging. Second, I will walk through an example of the development of machine learning model to detect brain tumors, attempting to point out key factors that should be addressed in developing a machine learning solution. Finally, I will discuss how machine and deep learning models should be evaluated and pertinent issues that should be considered in disseminating these findings. A disclaimer for this talk is that these concepts are intended to be presented at a very high level to provide a comprehensive introduction into the topic. There exists many dedicated sub-fields of research for many aspects that are covered herein, and thus, there is not sufficient time to delve into the specific details for every topic covered in this talk. Let's get started. Artificial intelligence,

machine learning, and deep learning are all interrelated. In fact, machine learning approaches can be described as a subset of artificial intelligence, and deep learning approaches can be described as a subset of machine learning. Artificial intelligence, first termed at a famous conference in Dartmouth in 1956, can be simply summed up as "the ability to learn without being explicitly programmed". AI methods can be categorized into two; first are general AI approaches, which would be the goal of developing an AI system that, as the name indicates, can be taught and applied to any problem or question. Narrow AI, on the other hand, refers to systems that are taught to have expertise in a limited domain. For nearly all of the applications that we are talking about today, these are applications of narrow AI. Solutions trained to solve very specific problems in medical imaging. Machine learning refers to a way and how an AI system can be trained. Machine learning is a generalized way of training where a system learns by being exposed to numerous examples of input data along with matching output. Deep learning is a further subset of machine learning methods. The key benefits of deep learning approaches relative to traditional machine learning approaches is the automated feature extraction capability. That is, with deep learning approaches, we do not need to tell the system what part of the data is important and what features need to be learned to learn the desired output. While this does not necessarily guarantee a better output or a better performing model, the burden of determining the important features of the data is eliminated, greatly simplifying the training process. Deep learning approaches themselves are commonly implemented in two ways; using supervised learning or unsupervised learning. The difference here with an example for supervised learning containing matched pairs of data. For example, including PET brain images of those with and without Alzheimer's disease and putting those into a deep learning model such that the model could predict whether a patient has or does not have Alzheimer's disease. Unsupervised learning would relax that constraint and determine a classification on its own. While this may have value in my previous example, perhaps at the current moment, it may be a bit hard to understand what that might mean as perhaps there could be a number of sub-classifications that the deep learning might determine that may not necessarily correlate with our interpretation of the clinical presentation. While there are some compelling demonstrations of newer techniques that are hybrid of supervised and unsupervised learning, most of what we talk about today will be supervised learning.