Five years ago, artificial intelligence was barely a blip on radiology’s radar screen. But, today, you’re hard-pressed to find an industry conference that doesn’t include multiple sessions on how these technological advancements are influencing daily practice or how they can make diagnosis easier.
Growth has been exponential, according to industry experts, and there’s no sign it’s slowing down. Based on data from Tractica, a marketing intelligence firm that focuses on human interaction with technology, the artificial intelligence (AI) industry is set to balloon from $2 billion to $3 billion in 2016 to more than $60 billion within the decade. And, medical imaging’s portion of that will likely be $19 billion by 2025.
But, what does that mean for you? What do you need to know about AI now, what should you keep your eyes open for over the next five to 10 years, and what’s being done to move in that direction? The pervasive fear has been AI would begin to chip away at jobs, slowly eroding the provider population. While that concern isn’t coming true, the technology is set to become an effective partner.
“Artificial intelligence won’t necessarily replace radiologists, but it will replace radiologists who don’t use artificial intelligence,” says Raymond Liu, MD, associate radiologist at Massachusetts General Hospital. “The idea is that artificial intelligence will be an augmentation tool.”
Combining the prowess of AI and radiologists can create a hybrid intelligence that could lead to higher diagnostic accuracy and improved safety standards. They could also serve as effective decision support systems, facilitating diagnoses and reducing physician burnout.
To date, AI efforts have mainly focused on machine learning, the ability of a machine to intake data and learn how to evaluate it. The next steps will be far more complex with deep learning, says Bradley Erickson, MD, an oncology radiologist at the Mayo Clinic in Rochester, Minn.
“We will almost be limited by imagination. Traditional machine learning, for example, has been very focused on detecting cancer,” he says. “Deep learning can do that, but it can do other things we weren’t even thinking about.”
Here are some developments to follow:
Comprehensive AI: In many ways, machine learning has been narrow AI — tools that have had a limited application, such as identifying pneumonia. The next step, Liu says, will be creating tools that help you read and analyze full chest X-rays. In another example, the Shirley Eye Institute at the University of California at San Diego has developed technology that can identify if the eye lens is captured in a CT scan. Avoiding the lens is critical because it is particularly sensitive to radiation. A tool that can immediately identify this problem and alert providers or technologists can be vital to improving patient care.
Non-interpretive AI: The majority of attention with AI in radiology has been given to software and other tools that enhance diagnostic and therapeutic capabilities. However, the technology can also be applied to non-interpretive tasks, such as improving operational workflow, finance management, and quality improvement. For example, FlowSigma, led by Erickson, automates and integrates tasks to save you time.
“AI is largely used in radiology for interpretation of images and identifying pattern recognition. When you think about value imaging, interpretation and acquisition is just one step in the process,” Liu says. “But, the value chain also includes scheduling, post-interpretative tasks, and connecting with communication.”
Pattern AI: AI has been effective in identifying patterns in imaging, but work is ongoing to further augment what the technology is able to perceive. For example, Erickson says, radiation oncologists currently use imaging to focus on histology and tumor grading, but new tools can find molecular markers that are better prognostic and therapeutic indicators.
“We can actually predict those molecular markers with high accuracy, such as 90 percent to 95 percent from routine brain MRI using deep learning,” he says. “This is something, as a radiologist, I would not get close to accurately predicting. This is something new algorithms are finding that we don’t currently perceive. It’s a really exciting thing.”
Ultimately, says Erickson, who’s lab is working on this type of tool, this brand of AI will reveal details about medicine that providers don’t see. Siemens Healthineers also offers a pattern recognition tool, called Automatic Landmarking and Parsing of Human Anatomy (ALPHA), that offers faster anatomical structures, improving workflow.
Explainable AI: This next wave of AI will be designed to give providers a way to understand how the deep learning tools reach their conclusions, making the findings even more reliable, Erickson says. It’s currently being explored by the Department of Defense to improve targeting systems, but the application for radiology is there.
For example, those explainable AI tools that identify molecular markers to indicate disease in tumors, but that are imperceptible to the human eye, Erickson says, pinpoint these markers, giving providers a greater insight into tumor biology and revealing unseen invasiveness. Explainable tools could also fine-tune algorithms used for diagnosis, such as ones used to identify pneumothorax.
Support For AI Advancement
Navigating the fast-changing path of implementing AI in your practice will likely be fraught with challenges and some confusion. Both industry leadership and academic are taking steps to help guide you as you determine which tools to employ and how best to use them.
First, the American College of Radiology (ACR) has launched the Data Science Institute (DSI). The goal is to promote industry standards and transparency and provide clinically relevant case studies in medical imaging, interventional radiology, and radiation oncology that outline the best ways to use AI, Liu says. The DSI also aims to create ways to monitor AI effectiveness and address regulatory, legal, and ethical issues. Ultimately, the hope is the Food & Drug Administration (FDA) will embrace and more easily approve AI technology.
“It’s very early, but I think the FDA is trying to remake itself and trying to be flexible and facile with newer developments,” Liu says. “We can’t predict what regulation aspects will loosen up. It will be a moving target as the FDA adapts itself to newer technologies.”
Additionally, Massachusetts General and Brigham & Women’s Hospitals created the Center for Clinical Data Science. By combining two large-volume data sets to train AI tools, both academic medical institutions aim to find opportunities for diagnostic, therapeutic, population health, and personal genetics. The intent, Liu says, is to bolster industry-clinical partnerships with vendors to provide widely-applicable, clinically-useful solutions.
Before you invest in and implement these types of tools — once they all become available — be sure you’ve prepared yourself by setting up the best workflow management possible, Erickson says. Without the right software in place, you won’t be able to effectively and efficiently put more comprehensive AI tools into place. These new partners won’t take away your responsibilities as a radiological provider, but they will shift how you handle your workload.
“Just like Excel spreadsheets changed the way accountants work, deep learning AI will prove the same thing is true in radiology,” Erickson says. “Algorithms will tee up some of the routine things for us so we’re more efficient with mundane tasks and routine screenings. It will give us more time to focus on the unusual and atypical scans.”