COMMENTARY

AI in GI: The Future of Diagnosing Disease in Gastroenterology

Digestive Disease Week (DDW) 2019

Alok S. Patel, MD; Sushovan Guha, MD, MA, PhD

Disclosures

July 18, 2019

Alok S. Patel, MD: Hi, everyone. This is Dr Alok Patel with Medscape. We're here at Digestive Disease Week 2019 in San Diego, talking about all types of fascinating gastrointestinal (GI) topics. Today I'm joined by Dr Sushovan Guha, a gastroenterologist and physician executive director at Banner Digestive Diseases Institute in Phoenix, Arizona. We are going to talk about artificial intelligence (AI) in GI.

Dr Guha, last year a headline all over the GI world was about how AI can help high-performing colonoscopists identify 20% more polyps than they would without AI.[1] In layman's terms, can you explain how AI works with colonoscopy?

Sushovan Guha, MD, MA, PhD: There has been a lot of excitement in the field since this publication last year about the high-performing colonoscopists. But this technology is going to help even the novices with minimal training improve their adenoma detection rates. So how does it work? As you know, AI-assisted colonoscopy is a system through which a computer analyzes high-quality images; then, when it detects a polyp, it can tell you what type of polyp it is.

Patel: It's only looking at the images the colonoscopist takes.

Guha: Exactly. You have to take high-quality images. That is the key. Once we take the high-quality images, we can train the computer. That's called machine learning. We feed the computer thousands of images, and it begins to learn and re-learn and gets better at detection every time. Once the computer learns what to look for, that's what we call computer-aided detection (CAD). Algorithms have been developed based on the surface characteristics of the polyps in the colon—either the pit patterns or the vascularity on the surface of the polyp—whereby the computer can then diagnose. That's called computer-aided diagnosis (CADx). This year at DDW, I was fortunate to chair a few sessions on this topic.

Patel: The results from that early trial last year were promising. We read that there was almost a 99.5% negative predictive rate with this AI technology. So it seems that there's a bright future with polyp detection in colonoscopies. Looking ahead, aside from colonoscopy, is there any talk about using AI in other GI diseases—for example, those in the esophagus?

Guha: Yes, certainly. The colonoscopy field has progressed further in this over the past 18 months or so using the AI technology with machine learning. At the same time, people are also looking at other disease processes in the GI tract, such as esophageal diseases. For example, how do we accurately diagnose Barrett esophagus? Barrett esophagus is a metaplastic condition. The question for most gastroenterologists is, when does it become dysplastic, which is the early cancer stage. That's when we can take certain actions to address it. So that is a key question. Can AI learn dysplasia? That's the holy grail. It hasn't accomplished this yet, but some high-quality images are being developed.

For this technology, remember that right now, the computer only sees the surface. It also has to learn the edge and the depth—that is, the three dimensions. So all of those things will come with more and better high-quality images. There are techniques that enable us to see the depth in the esophagus—for example, using confocal laser endomicroscopy. Another technology, called endocytoscopy, also shows the depth. Thus, once we feed the computer all of those images, it will probably learn how to detect in a three-dimensional fashion.

Patel: A lot of promise, a lot of things we could do. How far are we from seeing this in clinical practice? What are the barriers?

Guha: There are quite a few barriers. These are all promising technologies in the GI tract, including the colon, the esophagus, and the stomach and eventually the small bowel, the biliary tract, and the pancreas. As you know, however, we need to be cognizant of several things before we move forward into the clinical arena.

We have to work out the legal aspect. With this technology, the computer is making certain decisions. Who is liable in case of errors? I know the sensitivity and specificity are pretty high, but they are still not 100%. There will be some miss rates. The question is, who is to blame if the computer misses the cancer or the physician misses the cancer? All those legal questions will need to be worked out.

We also need to work out the validation. How do we validate the thousands of algorithms people are working on? The Center for GI Innovation and Technology within the American Gastroenterological Association (AGA) is spearheading this validation of technologies. We need to know that there is some basic standardization of all of these algorithms. Then the US Food and Drug Administration (FDA) will come in and conduct its own validation and approve the technology. Right now, this is neither a device nor a drug; it is in a gray zone. It's a computer-based algorithm. But we are getting there. The FDA announced a couple of months back that they are trying to come up with a white paper on how to approve these algorithms, maintain the intellectual property, and ask companies to conduct validation trials. I have been talking to our societies, the AGA and the American Society for Gastrointestinal Endoscopy (ASGE), to develop a library of high-quality images.

Patel: At the end of the day, we have a long way to go, but the computer is only as good as the operator.

Guha: Exactly—and as good as the images it is fed.

Patel: Dr Guha, thank you so much.

Comments

3090D553-9492-4563-8681-AD288FA52ACE
Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as:

processing....