AI is about to change radiology faster than you think — here’s why
New forces are gathering to accelerate AI’s plunge deep into the heart of radiology
I’m a big fan of automation, but I’m also a radiologist that loves humans. For the past twenty years, I’ve worked on brain-machine interface technology and I share Elon Musk’s enthusiasm for supercharging human brains to meet the rising challenge of artificial intelligence. Just a few months ago, my team fully automated CME for radiologists by launching Orbit, the browser plugin that lets radiologists earn all of their CME for the searching and learning we do all the time in the reading room.
But the rise of AI in radiology comes from a different culture, focused on disrupting (read: destroying) industries with automation. When I visited a prominent Silicon Valley radiology AI startup in 2016, the attitude from the entire team was focused on “destroying radiology.” They derided input from expert radiologists, and appointed a radiation oncologist to oversee the medical considerations.
Geoff Hinton’s comments from a 2016 machine learning conference in Toronto represents the resolute confidence of a machine learning luminary — “I think if you work as a radiologist, you’re like the coyote that’s already over the edge of the cliff… within five years, deep learning is going to do better than radiologists… it might be ten years.”
There are lots of reasons we should dismiss the hubris that AI will replace radiologists.
- Much of our work integrates imaging with knowledge of pathophysiology and clinical medicine, and machines that learn images paired with diagnoses won’t capture the richness of our training experience.
- Radiologists see imaging as a snapshot in the story of a patient that extends forward and backwards in time, adding the complexity of modeling spatiotemporal processes. We make decisions based on data that’s a mixture of images, lab values, written text, and unconstrained conversations — curating all of this heterogeneous data and then designing algorithms that make sense of it is hard.
- The radiologist frequently consults with referring physicians on the next step in diagnosis or management in relation to patient goals, social norms, and other context that requires solving entirely separate machine learning problems.
- Image guided interventions require robotics, which is an order of magnitude more complex than software-based AI.
But don’t let these arguments sedate you. AI will start to impact the radiology job market long before it replaces radiologists. Here’s why:
1. FDA approval may not be necessary
It turns out that FDA approval is probably not necessary for AI to enter our practices. A prime opportunity for AI is triage. By sorting and assigning reading lists, AI will be able to increase the efficiency and accuracy of radiologists. Here’s how.
Image analysis will rate the complexity of studies. Initially, these studies might be assigned in first-come-first-serve fashion to the subspecialists in the group, with urgent cases (detecting head bleeds, dissections, etc.) popping to the top of the list. But with assistance, algorithms will record turnaround times and accuracy of diagnoses for each provider based on those initial case assignments. Over the course of weeks, algorithms will learn which radiologists in the group perform best at particular times of day, based on subject matter and complexity of cases.
We already do some of this triage manually, using administrative assistants to assign cases or occasionally getting lucky with technologists that flag time-sensitive diagnoses. We even have bitter RVU-focused turf battles over whether the neuro or MSK section will read spine. But AI will apply the principles of speed and quality (or any other values it’s asked to optimize) relentlessly and uniformly. A modest 10% increase in efficiency could substantially impact the job market, unlocking the power of 11 radiologists for every group of 10.
2. Radiologists are aligning with AI companies
In the intervening two years since the unvarnished proclamations from Geoff Hinton, AI companies have learned to warm up to radiologists. This is because radiologists will generate the data sets that will ultimately replace them. Powerscribe was originally trained on data from the medical transcriptionists it replaced. Medical coding software was trained on the output from human medical coders it replaced.
Diverse and cleanly labeled training data is the fuel that accelerates the progress of AI, and radiology groups are the refineries that generate this fuel. Compared with Hinton’s comments, words from Facebook Chief AI Scientist Yann LeCun in October 2017 about AI in radiology are relatively seductive, talking about helping radiologists by taking care of the mundane work and freeing radiologists to care for more complex cases.
Realizing the incredible financial and technological potential of AI, radiology departments around the country are quietly forming partnerships with AI ventures from traditional large tech companies and startups alike. Some startups fantasize about running their own radiology practices, and using their radiologists both to generate their own data in-house and to showcase the power of their algorithms. Academics are pioneering structured reporting practices with the goal of paving the path for machines to learn from these reports.
The motives to engage in these partnerships are varied, driven by financial gain, academic prestige, and in some cases the noble goal of making patient care safer and more accessible. Regardless of the motive, aligning radiologists with the mission of training AI will accelerate the impact of AI on our profession.
3. Consolidation will potentiate rapid adoption
The basic challenge in spreading any software is getting people to use it. With the consolidation of radiology practices and hospital systems, decision-making about software purchases is also being concentrated. Technology companies will have to convince fewer people to have their AI software adopted.
Furthermore, the people making decisions about whether to adopt AI will be focused on broad goals like efficiency and quality of care, not the specific goal of preserving radiologist jobs. That means that AI companies won’t need to convince radiologists that AI is good for radiology. They’ll just need to convince CEOs that AI will make more money for the enterprise.
Bottom line on the radiologist job market
AI is coming for radiology, and a confluence of factors are already accelerating its arrival.
- For patients, AI will mean faster reporting, uniformity of care, and price relief.
- For existing full-time radiologists and departments that co-invest with AI companies, there will be substantial short-term financial gains.
- For new graduates and locum tenens radiologists, even 10% gains in efficiency will leave weekend and night radiology work more competitive than ever.
And AI is only one part of the jobs equation that also includes case volume and reimbursement, encompassing factors like demographic trends, evidence-based and consensus recommendations, capitation, and vertically-integrated healthcare systems. All of these factors bubble and churn in the mix that modulates the fate of our profession.
References
R.A. Washington, “Why scan-reading artificial intelligence is bad news for radiologists,” The Economist, Nov 29, 2017. [link]. TL;DR: even the smallest advance in AI can strongly impact the radiology job market.
J.H. Thrall et al., “Artificial Intelligence and Machine Learning in Radiology: Opportunities, Challenges, Pitfalls, and Criteria for Success,” JACR, March 2018. [link]. TL;DR: rosy opinion piece about AI out of MGH Radiology — supposedly everything is going to be ok.
Ram Srinivasan MD PhD is founder and CEO of Orbit, the revolutionary new plugin that radiologists across the United States have started using to complete all of their CME requirements including SA-CME from the searching and learning they’re already doing in the reading room. He received his PhD in Electrical Engineering and Computer Science from MIT, and completed his MD at Harvard Medical School. Ram also maintains an active radiology practice in Palo Alto and teaches Core Physics Review, the leading performance-focused radiology physics course for the ABR Core Exam.
