The electric dream
The long-term goal for AI in colonoscopy is automated polyp detection, a so-called optical biopsy, but that vision lies well in the future, said Mr. Schwartz. The primary issue is that only still images are available as training sets, and these don’t capture the diversity of patients, endoscopy systems, and operators that will be required to create a robust, generalizable polyp detection system. Existing efforts have shown promise on training sets, but struggle in real-world tests. “AI is good at tricking you into thinking it’s a working system when it’s only looking at retrospective data,” said Mr. Schwartz.
Olympus signed an agreement last year with ai4gi, a commercial initiative applying deep learning to gastrointestinal cancer, to combine its AI systems with Olympus’ colonoscopy line, but Mr. Heine agreed that optical biopsies won’t appear any time soon: “We’re not ready right now to launch anything that’s making a diagnosis claim. It’s not about optical biopsies at this point. It’s about supporting the physician,” he said.
Along with improving video capture and quality-control efforts, Mr. Schwartz believes that Virgo’s systems can help solve the problem of limited training data. By capturing and storing video data from a wide range of procedures, it is generating a resource that could boost the field and may one day make optical biopsies a reality. “It becomes the training set to build the AI video systems of the future,” he said.
Mr. Heine is an employee of Olympus. Dr. Tucker-Schwartz is an employee of NinePoint Medical. Mr. Schwartz is an employee of Virgo.
