We developed an artificial intelligent computer system with a deep neural network to analyze real-time video signals from the endoscopy station. This randomised controlled trial compared adenoma detection rate between computer-assisted colonoscopy and standard colonoscopy.

Colonoscopy is a primary screening and follow-up tool to detect colorectal cancer, a third leading cause of cancer death in Taiwan. Most colorectal cancers (CRCs) arise from preexisting adenomas, and the adenoma-carcinoma sequence offers an opportunity for the screening and prevention of CRCs. The removal of adenomatous polyps can lower the incidence of CRCs and result in reduced motality from CRCs. The adenoma detection rate, the proportion of screening colonoscopies performed by a endoscopist that detect at least one colorectal adenoma or adenocarcinoma, has been recommended as a quality indicator. The adenoma detection rate was inversely associated with the risks of interval colorectal cancer, advanced-stage interval cancer, and fatal interval cancer. However, adenoma detection rates vary widely among endoscopists in both academic and community settings. Polyp miss rates as high as 20% have been reported for high definition resolution colonoscopy. An improvement in adenoma detection rate at screening colonoscopy, translates into reduced risks of interval colorectal cancer and colorectal cancer death. Computer-aided detection of polyps might assist endoscopists to reduce the miss rate and enhance screening performance during colonoscopy. Computer-aided diagnosis and computer-aided detection are computerized systems that learn and inference in medical fields. Computer-aided diagnosis has been developed in colon polyp classification.

Computer-assisted image analysis has the potential to further aid adenoma detection but has remained underdeveloped. A notable benefit of such a system is that no alteration of the colonoscope or procedure is necessary. Machine learning with a deep neural network has been successfully applied to many areas of science and technology, such as object recognition and detection of computer vision, speech recognition, natural language processing. We developed an artificial intelligent computer system (PX-1) with a deep neural network to analyze real-time video signals from the endoscopy station. This randomised controlled trial compared ADR between computer-assisted colonoscopy and standard colonoscopy.

Inclusion Criteria:

Patients aged ≥20 years, scheduled for colonoscopy for one of the following indications for colonoscopy, were invited to participate in this study: polyp surveillance, changed bowel habits and/or bloody stools, bowel complaints, a positive family history for CRC, a positive FOBT, abdominal pain, diarrhoea, post-polypectomy surveillance.

Exclusion Criteria:

We excluded patients from this study if: (1) they had known colonic neoplasia or inflammatory or other significant colonic disease, such as patients specifically presenting for polypectomy; (2) there was open bleeding or they were receiving an emergency colonoscopy; (3) they had previously previous colonic resection; (4) they were in poor general condition (more than American Society of Anesthesiologists grade III); (5) they were receiving anticoagulant medication; (6) they had severe comorbidity, including end-stage cardiovascular, pulmonary, liver or renal disease); (7) they were not able or refused to give informed written consent; (8) following enrolment and randomisation to one of the arms, those subjects who had inadequate colon preparation or in whom the caecum could not be reached were also excluded.