The AI revolution, with some important caveats
BY JEREMY R. GLISSEN BROWN, MD, MSC
In 2018, Japan’s Pharmaceutical and Medical Device Agency approved the first artificial intelligence (AI)–based tool, a computer-aided diagnosis system (CADx) for use in clinical practice.1 Since that time, we have seen regulatory approval for a variety of deep learning and AI-based tools in endoscopy and beyond. In addition, there has been an enormous amount of commercial and research interest in AI-based tools in clinical medicine and gastroenterology, and it is almost impossible to open a major gastroenterology journal or go to an academic conference without encountering a slew of AI-based projects.
Dr. Jeremy R. Glissen Brown
Many thought and industry leaders say that we are in the midst of an AI revolution in gastroenterology. Indeed, we are at a period of unprecedented growth for deep learning and AI for several reasons, including a recent shift toward data-driven approaches, advancement of machine-learning techniques, and increased computing power. There is, however, also an unprecedented amount of scrutiny and thoughtful conversation about the role AI might play in clinical practice and how we use and regulate these tools in the clinical setting. We are thus in a unique position to ask ourselves the essential question: “Are we on the cusp of an AI revolution in gastroenterology, or are we seeing the release of medical software that is perhaps at best useful in a niche environment and at worse a hype-driven novelty without much clinical benefit?” We will use the most popular use-case, computer aided detection (CADe) of polyps in the colon, to explore this question. In the end, I believe that deep-learning technology will fundamentally change the way we practice gastroenterology. However, this is the perfect time to explore what this means now, and what we can do to shape what it will mean for the future.
CADe: Promise and questions
CADe is a computer vision task that involves localization, such as finding a polyp during colonoscopy and highlighting it with a hollow box. CADe in colonoscopy is perhaps the most well-studied application of deep learning in GI endoscopy to date and is furthest along in the development-implementation pipeline. Because of this, it is an ideal use-case for examining both the evidence that currently supports its use as well as the questions that have come up as we are starting to see CADe algorithms deployed in clinical practice. It is honestly astounding to think that, just 5 years ago, we were talking about CADe as a research concept. While early efforts applying traditional machine learning date back at least to the 1990s, we started to see prospective studies of CADe systems with undetectable or nearly undetectable latency in 2019.2 Since that time we have seen the publication of at least 10 randomized clinical trials involving CADe.
CADe clearly has an impact on some of the conventional quality metrics we use for colonoscopy. While there is considerable heterogeneity in region and design among these trials, most show a significant increase in adenoma detection rate (ADR) and adenomas per colonoscopy. Tandem studies show decreases in adenoma miss rate, and at least one study showed a decrease in sessile serrated lesion miss rate as well. In one of the first randomized, controlled trials across multiple endoscopy centers in Italy, Repici and colleagues showed an increase in ADR from 40.4% in the control group to 54.8% in the CADe group (RR, 1.30; 95% confidence interval, 1.14-1.45).3 Because of pioneering trials such as this one, there are currently several CADe systems that have received regulatory approval in Europe, Asia, and the United States and are being deployed commercially.
It is also clear that the technology is there. In clinical practice, the Food and Drug Administration–approved systems work smoothly, with little to no detectable latency and generally low false-positive and false-negative rates. With clinical deployment, however, we have seen the emergence of healthy debate surrounding every aspect of this task-specific AI. On the development side, important questions include transparency of development data, ensuring that algorithm development is ethical and equitable (as deep learning is susceptible to exacerbating human biases) and methods of data labeling. On the deployment level, important concerns include proper regulation of locked versus “open” algorithms and downstream effects on cost.
In addition, with CADe we have seen a variety of clinical questions crop up because of the novelty of the technology. These include the concern that the increase in ADR we have seen thus far is driven in large part by diminutive and small adenomas (with healthy debate in turn as to these entities’ influence on interval colorectal cancer rates), the effect CADe might have on fellowship training to detect polyps with the human eye, and whether the technology affects sessile serrated lesion detection rates or not. The great thing about such questions is that they have inspired novel research related to CADe in the clinical setting, including how CADe affects trainee ADR, how CADe affects gaze patterns, and how CADe affects recommended surveillance intervals.
CADx, novel applications, and the future
Though there is not space to expand in this particular forum, it is safe to say that with the advancement of CADx in endoscopy and colonoscopy, we have seen similar and novel questions come up. The beautiful thing about all of this is that we are just scratching the surface of what is achievable with deep learning. We have started to see novel projects utilizing deep-learning algorithms, from detecting cirrhosis on ECG to automatically classifying stool consistency on the Bristol Stool Scale from pictures of stool. I ultimately do think that the deployment of AI tools will fundamentally change the way we practice and think about gastroenterology. We are at an incredibly exciting time where we as physicians have the power to shape what that looks like, how we think about AI deployment and regulation and where we go from here.
Dr. Glissen Brown is with the division of gastroenterology and hepatology at Duke University Medical Center, Durham, N.C. He has served as a consultant for Medtronic.
References
1. Aisu N et al. PLOS Digital Health. 2021 Jan 18. doi: 10.1371/journal.pdig.0000001.
2. Wang P et al. Gut. 2019 Oct;68(10):1813-9.
3. Repici A et al. Gastroenterology. 2020 Aug;159(2):512-20.e7.
