ML 101: The Radiologist's Basic Guide
From IBM's Watson to CAD, most radiologists have heard of machine learning. But do you know how this technique is already used in the field? Plus, what does the future hold? The ACR Bulletin brings you FAQs so you can be sure to have the basics down pat.
What is Machine Learning?
A researcher named Arthur Samuel coined the term “machine learning” (ML) in 1959. Samuel defined ML as the “field of study that gives computers the ability to learn without being explicitly programmed.” In other words, ML is the notion that machines can, over time, learn to do what we as humans do. How? Computers can capture datasets, aggregate the information in those data, and then create predictions, explains Keith J. Dreyer, DO, PhD, vice chair of radiology at Massachusetts General Hospital in Boston and assistant professor of radiology at Harvard Medical School in Cambridge, Mass.
How Does It Work?
The idea that a machine can understand how to perform a new process without specific instruction once seemed impossible. What inner workings have made ML a reality? The answer is a series of algorithms, according to J. Raymond Geis, MD, FACR, assistant professor of radiology at University of Colorado School of Medicine in Aurora, Colo., and vice chair of the ACR’s Commission on Informatics. “Instead of being programmed to do a specific task, machines rely on these algorithms to incorporate statistical methods and learn to perform the task without being specifically shown how,” Geis explains. “While these algorithms have been known for decades, the huge increase in computer-accessible data and faster computing power now make them affordable to use in daily life.”
How Is It Being Applied?
One of the most famous examples of ML is IBM’s Watson, whom you may remember from the TV game show Jeopardy! Watson won a first place prize of $1 million against human competitors. More recently, you may have also seen a commercial about Watson sorting through hundreds of medical images to make a diagnosis. But ML is already used in medicine and radiology: the most common use being computer-aided detection, or CAD. CAD uses datasets to process an image and detect conspicuous sections where disease might be present. However, even with CAD, radiologists still typically interpret images and review and contextualize the information from CAD: “It needs to be an interactive process between human and machine,” says Dreyer.
What Does It Mean for Radiology?
There’s a large initiative by academic and commercial groups to develop more valuable ML products, so “you will see the use cases boom in the future,” says Geis. He adds that ML provides an opportunity to find new and useful patterns in datasets, such as electronic health records, radiologic images, or even genomic information. Within ML, radiology seems to have become a hot topic because of the vast quantities of already labeled images, such as those on social media platforms like Facebook, which allow us to test applicability. Even though Facebook images may be of friends or scenery, rather than radiologic images, they can serve as use cases to experiment with ML recognition — for instance, the way that Facebook can now “suggest” who to tag in an image.
Dreyer explains that ML can be used for two different tasks: detection and diagnosis. Detection simply refers to identifying the presence of a finding in a radiologic image. Diagnosis could take this a step further, analyzing possible clinical classifications to identify the object, similar to how a radiologist would read an image to determine what’s going on with a patient.
What’s the Difference Between Machine Learning vs. Artificial Intelligence vs. Deep Learning?
Machine learning refers to a computer’s ability to train itself without being programmed. Artificial intelligence is what results when machine learning is put into practice and is a more mainstream, popularized term. Another common term that’s been confused with ML is deep learning, which is actually a subset of ML. Deep learning involves multiple, complex levels of computation, explains Geis. “In a simple network, you have one input layer, a single ‘hidden’ or computational layer, and one output layer. In a deep network, you have multiple hidden layers.”
What Challenges Exist?
A recent article in the New England Journal of Medicine warned that ML could “displace much of the work of radiologists and anatomical pathologists” while improving diagnostic accuracy. Despite the use of CAD and the development of even better algorithms, many remain concerned about the use of ML in image interpretation. Dreyer explains, “There’s a lot of context and information that needs to be conveyed beyond a single simple interpretation based on one finding in a complex imaging exam. Even if an algorithm finds tuberculosis on a chest CT, that doesn’t mean other diseases aren’t present. There still needs to be a comprehensive interpretation.” Geis also adds that while many are optimistic about the benefits of ML, many roadblocks exist, including the need for more research and datasets. Additionally, ML brings up many issues that need to be thought through, such as legal risk, ethical issues, and an undefined regulatory framework.
What Does the Future Hold?
Could ML actually replace radiologists? “Not anytime soon,” according to Geis. “Machine learning will, however, arrive in the next few years in the form of advanced CAD, personalized exam protocols, and tools to help with specific clinical questions.” So what’s the key for radiologists to avoid being left behind? “To see the potential of ML in radiology, we need to watch the consumer markets to see how technologies are being applied,” says Dreyer. “Companies like Google, Facebook, and IBM are driving a lot of innovation, and it will be translated into the clinical domain.”
By Alyssa Martino, freelance writer for the ACR Bulletin