GI disease screening with artificial intelligence is close

SAN FRANCISCO – As a tool for the screening and diagnosis of diseases in the gastrointestinal (GI) tract, artificial intelligence (AI) is advancing rapidly, according to a review of this technology presented at the 2019 AGA Tech Summit, sponsored by the AGA Center for GI Innovation and Technology.

Much of the focus of the update was on screening colonoscopy, but the same principles are relevant and being pursued for other GI conditions, such as dysplasia screening in patients with Barrett’s esophagus and the assessment of mucosal healing in inflammatory bowel disease, according to Michael F. Byrne, MD, a clinical professor in the division of gastroenterology at Vancouver General Hospital.

“There are many technologies [to improve screening and diagnosis of GI diseases], but I believe these will struggle if they do not also have some kind of built-in machine intelligence,” Dr. Byrne said. In addition to his practice in gastroenterology, Dr. Byrne is CEO of Satis Operations and founder of AI4GI, a commercial joint venture focused on clinical applications of AI in colon polyp disease.

In this context, AI is being built on the principle of deep learning, which employs neural networks or a set of algorithms that permits a computer to recognize patterns when “trained” with data. In the machine learning process, the computer can use a large number of features in the task of discrimination.

This might suggest that AI could, in turn, train physicians to recognize the same features, but this underestimates the complexity and sophistication of machine learning, according to Dr. Byrne. The current status of machine learning for screening colonoscopy underscores this point.

“A computer can consider a thousand features when evaluating a polyp, which is way beyond what we can do,” Dr. Byrne said. Even with advances to improve visualization in screening colonoscopy, such as improved resolution and better lighting, the reason that AI is expected to prevail is that “the human eye is just not accurate enough.”

Many groups have developed advanced machine learning systems for screening colonoscopy. Dr. Byrne reviewed some of the early work done in Japan and that performed with a system in development by his group. In a study with the AI4GI model, published recently in Gut (2019;68:94-100), greater than 94% accuracy was achieved in distinguishing adenomas from hyperplastic polyps using histopathology as a gold standard.

Because of the ability of machine learning to see what the human eye cannot, Dr. Byrne predicts that AI-centric classification will replace current polyp classification systems, which could offer categories that are more clinically useful and reliable.

However, the work in screening colonoscopy is just the beginning, according to Dr. Byrne. “The opportunity of machine learning goes way beyond polyps.”

Recognizing dysplasia associated with Barrett’s esophagus has parallels with identifying adenomatous polyps in screening colonoscopy, but Dr. Byrne also discussed machine learning as an “optical biopsy” for evaluating the mucosa of patients with IBD. No longer a screening approach, the characterization of IBD tissue could help with therapeutic decisions.

With an AI approach to optical biopsy, “there is a great opportunity to assign an inflammatory burden in IBD,” he suggested, explaining how evidence of disease activity could guide escalation or de-escalation of treatment within the context of the treat-to-target approach to prolonging remission.

Overall, there is abundant evidence that “optical biopsy is feasible,” Dr. Byrne said. He indicated that clinical applications are approaching quickly. While he acknowledged that the technology “will need a human in the loop” as it enters clinical practice initially, he believes that this technology will play a significant role in GI practice because of the clear limitations of the human eye in assessing endoscopic images of GI tissue.

SAN FRANCISCO – As a tool for the screening and diagnosis of diseases in the gastrointestinal (GI) tract, artificial intelligence (AI) is advancing rapidly, according to a review of this technology presented at the 2019 AGA Tech Summit, sponsored by the AGA Center for GI Innovation and Technology.

Much of the focus of the update was on screening colonoscopy, but the same principles are relevant and being pursued for other GI conditions, such as dysplasia screening in patients with Barrett’s esophagus and the assessment of mucosal healing in inflammatory bowel disease, according to Michael F. Byrne, MD, a clinical professor in the division of gastroenterology at Vancouver General Hospital.

“There are many technologies [to improve screening and diagnosis of GI diseases], but I believe these will struggle if they do not also have some kind of built-in machine intelligence,” Dr. Byrne said. In addition to his practice in gastroenterology, Dr. Byrne is CEO of Satis Operations and founder of AI4GI, a commercial joint venture focused on clinical applications of AI in colon polyp disease.

In this context, AI is being built on the principle of deep learning, which employs neural networks or a set of algorithms that permits a computer to recognize patterns when “trained” with data. In the machine learning process, the computer can use a large number of features in the task of discrimination.

This might suggest that AI could, in turn, train physicians to recognize the same features, but this underestimates the complexity and sophistication of machine learning, according to Dr. Byrne. The current status of machine learning for screening colonoscopy underscores this point.

“A computer can consider a thousand features when evaluating a polyp, which is way beyond what we can do,” Dr. Byrne said. Even with advances to improve visualization in screening colonoscopy, such as improved resolution and better lighting, the reason that AI is expected to prevail is that “the human eye is just not accurate enough.”

Many groups have developed advanced machine learning systems for screening colonoscopy. Dr. Byrne reviewed some of the early work done in Japan and that performed with a system in development by his group. In a study with the AI4GI model, published recently in Gut (2019;68:94-100), greater than 94% accuracy was achieved in distinguishing adenomas from hyperplastic polyps using histopathology as a gold standard.

Because of the ability of machine learning to see what the human eye cannot, Dr. Byrne predicts that AI-centric classification will replace current polyp classification systems, which could offer categories that are more clinically useful and reliable.

However, the work in screening colonoscopy is just the beginning, according to Dr. Byrne. “The opportunity of machine learning goes way beyond polyps.”

Recognizing dysplasia associated with Barrett’s esophagus has parallels with identifying adenomatous polyps in screening colonoscopy, but Dr. Byrne also discussed machine learning as an “optical biopsy” for evaluating the mucosa of patients with IBD. No longer a screening approach, the characterization of IBD tissue could help with therapeutic decisions.

With an AI approach to optical biopsy, “there is a great opportunity to assign an inflammatory burden in IBD,” he suggested, explaining how evidence of disease activity could guide escalation or de-escalation of treatment within the context of the treat-to-target approach to prolonging remission.

Overall, there is abundant evidence that “optical biopsy is feasible,” Dr. Byrne said. He indicated that clinical applications are approaching quickly. While he acknowledged that the technology “will need a human in the loop” as it enters clinical practice initially, he believes that this technology will play a significant role in GI practice because of the clear limitations of the human eye in assessing endoscopic images of GI tissue.

SAN FRANCISCO – As a tool for the screening and diagnosis of diseases in the gastrointestinal (GI) tract, artificial intelligence (AI) is advancing rapidly, according to a review of this technology presented at the 2019 AGA Tech Summit, sponsored by the AGA Center for GI Innovation and Technology.

Much of the focus of the update was on screening colonoscopy, but the same principles are relevant and being pursued for other GI conditions, such as dysplasia screening in patients with Barrett’s esophagus and the assessment of mucosal healing in inflammatory bowel disease, according to Michael F. Byrne, MD, a clinical professor in the division of gastroenterology at Vancouver General Hospital.

“There are many technologies [to improve screening and diagnosis of GI diseases], but I believe these will struggle if they do not also have some kind of built-in machine intelligence,” Dr. Byrne said. In addition to his practice in gastroenterology, Dr. Byrne is CEO of Satis Operations and founder of AI4GI, a commercial joint venture focused on clinical applications of AI in colon polyp disease.

In this context, AI is being built on the principle of deep learning, which employs neural networks or a set of algorithms that permits a computer to recognize patterns when “trained” with data. In the machine learning process, the computer can use a large number of features in the task of discrimination.

This might suggest that AI could, in turn, train physicians to recognize the same features, but this underestimates the complexity and sophistication of machine learning, according to Dr. Byrne. The current status of machine learning for screening colonoscopy underscores this point.

“A computer can consider a thousand features when evaluating a polyp, which is way beyond what we can do,” Dr. Byrne said. Even with advances to improve visualization in screening colonoscopy, such as improved resolution and better lighting, the reason that AI is expected to prevail is that “the human eye is just not accurate enough.”

Many groups have developed advanced machine learning systems for screening colonoscopy. Dr. Byrne reviewed some of the early work done in Japan and that performed with a system in development by his group. In a study with the AI4GI model, published recently in Gut (2019;68:94-100), greater than 94% accuracy was achieved in distinguishing adenomas from hyperplastic polyps using histopathology as a gold standard.

Because of the ability of machine learning to see what the human eye cannot, Dr. Byrne predicts that AI-centric classification will replace current polyp classification systems, which could offer categories that are more clinically useful and reliable.

However, the work in screening colonoscopy is just the beginning, according to Dr. Byrne. “The opportunity of machine learning goes way beyond polyps.”

Recognizing dysplasia associated with Barrett’s esophagus has parallels with identifying adenomatous polyps in screening colonoscopy, but Dr. Byrne also discussed machine learning as an “optical biopsy” for evaluating the mucosa of patients with IBD. No longer a screening approach, the characterization of IBD tissue could help with therapeutic decisions.

With an AI approach to optical biopsy, “there is a great opportunity to assign an inflammatory burden in IBD,” he suggested, explaining how evidence of disease activity could guide escalation or de-escalation of treatment within the context of the treat-to-target approach to prolonging remission.

Overall, there is abundant evidence that “optical biopsy is feasible,” Dr. Byrne said. He indicated that clinical applications are approaching quickly. While he acknowledged that the technology “will need a human in the loop” as it enters clinical practice initially, he believes that this technology will play a significant role in GI practice because of the clear limitations of the human eye in assessing endoscopic images of GI tissue.