Last month, the FDA announced its first ever authorization of a machine learning tool for automatically detecting polyps from colonoscopy images.
In the realm of “AI that makes it real cancer patients”, I would classify this as a big and important milestone.
To make sense of this FDA authorization, consider first that the colonoscopy is an extremely effective procedure for detecting and preventing colorectal cancer. In a normal colonoscopy, a doctor examines the colon via a flexible tube and fiber optic camera, and the doctor relies on his or her own visual acumen. Suspicious polyps, which may indicate pre-cancerous lesions are detected by the physician and immediately removed.
By identifying and removing early stage polyps, colonoscopy has been shown to reduce the overall rate of colorectal cancer. However, colonoscopies are not perfect, and it is estimated that as many as 1/4 of all polyps are missed during routine colonoscopy.
Enter machine learning
Multiple groups have built deep learning models to automatically identify colorectal polyps from images. For an overview of several recent papers in the field, see several nice summaries here, here and here.
Into this fray, Cosmo Pharmaceuticals has built GI Genius. GI Genius is a computer aided detection (CADe) tool, built on convolutional neural networks. During colonoscopy, GI Genius highlights suspected polyps in real-time with a green bounding box. Doctors are then free to inspect the region further or ignore the signal. To get a better sense of how GI Genius works in clinical care, check out the YouTube video below of GI Genius in action on a real patient:
The FDA authorization of GI Genius was based on a randomized, multi-site clinical trial in Italy. The primary endpoint of the trial was the Adenoma Detection Rate (ADR), defined as the fraction of patients with at least 1 histologically proven adenoma or carcinoma. The trial had two arms: an experimental arm with GI Insight, and a control arm with no AI. The detection rate was significantly higher in the AI arm (54.8% v 40.4%), indicating that the AI assist resulted in more polyps being detected.
Notably, GI Genius has not been proven to show true clinical benefit. And, neither Cosmo nor the FDA make this assertion. It is true that GI Genius results in more polyps identified, but that does not necessarily mean that the tool is doing a better job at preventing colorectal cancer.
In fact, many of the polyps detected by GI Insight are extremely small <5 mm “diminutive” adenomas that may or may not progress to true colorectal cancer. Future long-term studies assessing rates of colorectal incidence will be needed. Nonetheless, this remains a significant milestone for AI in Cancer Care, and I suspect we are likely to see many more FDA authorizations over the next year.