EndoBRAIN — The Approval Billed as Japan's First AI Medical Device

EndoBRAIN is software that, during a colonoscopy, shows as a number how likely a polyp is to be neoplastic (a growth that should be removed) or non-neoplastic (one that need not be), to support the endoscopist's diagnosis. On 6 December 2018 it was approved as a Class III (highly controlled) medical device under Japan's Pharmaceuticals and Medical Devices Act. Cybernet Systems, the marketing authorization holder, later wrote that this was the first approval of an AI-equipped medical device under the Act. It is the product usually called Japan's first AI medical device. Olympus handled sales and launched it in Japan on 8 March 2019.

This was a Japanese approval under Japanese law. Earlier the same year, in April, the US FDA had authorized IDx-DR, which screens for diabetic retinopathy. IDx-DR was designed to give a result without a doctor reading the image; EndoBRAIN was designed to support a doctor's judgement, so the two play different roles.

Who built it

According to a joint release of 10 December 2018 from Showa University, Nagoya University and the Japan Agency for Medical Research and Development (AMED), the work was led by Professor Shin-ei Kudo's group at the Digestive Disease Center of Showa University Northern Yokohama Hospital. Kensaku Mori's laboratory at Nagoya University's Graduate School of Informatics built the diagnostic engine, and the medical-imaging software company Cybernet Systems was the industrial partner. The three had worked on it since 2013, funded by an AMED programme for AI-based endoscopic diagnosis software aimed at reducing colorectal cancer.

How it works

EndoBRAIN connects directly to the endoscopy unit and analyses images taken with Olympus's Endocyto, an endoscope with up to 520-fold optical magnification that lets doctors see cell nuclei in real time during the procedure. From those images the software outputs the likelihood that a polyp is neoplastic or non-neoplastic. Olympus said it supports two viewing modes: narrow-band imaging (NBI), which emphasises surface blood vessels, and dye staining.

Notably, the AI was not deep learning. AMED's release describes it as a support vector machine, a method first published in 1963, and lists as advantages over deep learning that its training and output are easier for engineers to understand and adjust. Images labelled "neoplastic" or "non-neoplastic" were used to train it, a textbook case of supervised learning. The release also states plainly that the product does not learn from images taken after release, so its performance does not change.

Accuracy: two sets of figures

Two sets of accuracy figures were published.

The five sites were Showa University Northern Yokohama Hospital, the National Cancer Center Hospital and National Cancer Center Hospital East, Shizuoka Cancer Center and Tokyo Medical and Dental University Hospital. Olympus labelled its study retrospective; it does not show how real-time use changes outcomes for patients.

What "first" covers

The wording differs from source to source:

• Cybernet Systems (January 2021): obtained approval under the Act in December 2018 as the first AI-equipped medical device.
• Olympus (February 2019): the first AI product in the endoscopy field to receive regulatory approval in Japan.
• The developers' review paper (2020): Japan's first regulatory approval of an AI-based diagnostic support system.

Olympus's version is limited to endoscopy. This page treats the claim as that of the maker and the developers.

What changed for clinics and for policy

The developers' goal was colorectal cancer prevention. The joint release notes that colorectal cancer was the leading cause of cancer death among Japanese women and the third among men, and that removing neoplastic polyps endoscopically is known from overseas studies to cut colorectal cancer deaths by 53–68%. Doctors therefore have to tell, during the examination, which polyps need removing and which do not. EndoBRAIN supplies a number to support that call, and the release says it is expected to improve the accuracy of doctors' endoscopic diagnosis.

Final responsibility stays with the doctor. The release positions EndoBRAIN as software that assists the doctor's diagnostic prediction. The follow-up product EndoBRAIN-EYE, approved on 24 January 2020 as a Class II device, helps doctors find lesions; Cybernet says it deliberately does not point to where a lesion is, and only warns with a sound and a screen colour, to leave room for the doctor's own diagnosis. It uses deep learning and was trained on about 3.95 million images.

Two policy questions surfaced. The first was insurance. In their 2020 review the developers, led by Yuichi Mori, described insurance coverage as essential for adoption after approval and noted that it had not yet been achieved for AI medical devices. The second was updates. Cybernet said EndoBRAIN-EYE does not improve itself through autonomous learning after release, and that any improvement would be filed with the PMDA and released as a new version only after approval. Reviewing software that does not learn after release is a point it shares with IDx-DR in the US, which was locked before its trial.

The first EndoBRAIN also depended on images from the specialised Endocyto scope. EndoBRAIN-EYE was described as working with Olympus's general colonoscopes of high-definition quality or better, which widened its reach.

The series afterwards

In its 2021 release Cybernet said the line-up now let AI support the whole colonoscopy workflow, from detecting and characterising lesions to choosing treatment. It noted that the performance study for EndoBRAIN-Plus was retrospective.

Japan's wider AI policy is covered on the AI Promotion Act page. Another side of medical AI, a commercial prediction model that fell well short when validated externally, is on the Epic Sepsis Model page. For the surrounding period, see the AI timeline and the AI and society timeline.