A new artificial intelligence (AI) algorithm evaluates colonoscopy exams with a level of accuracy equivalent to experts, scientists in Japan report.
The AI algorithm they developed may also help physicians predict whose patients with ulcerative colitis (UC) have entered in remission without need for a bowel biopsy.
Findings were reported in the study, “Development and Validation of a Deep Neural Network for Accurate Evaluation of Endoscopic Images From Patients with Ulcerative Colitis,” published in the journal Gastroenterology.
Evaluations of ulcerative colitis are based on a combination of colonoscopy and biopsy — which requires the removal of a small sample of bowel tissue — and used by clinicians to choose the best treatment regimens for patients and to monitor their response to therapy.
Remission, the disappearance of all signs of the disease detected either by colonoscopy (endoscopic remission) or biopsy (histological remission), is the main goal of treatment for UC. Both exams can also be used to predict patient clinical outcomes, including the likelihood of a worsening bout of the disease (relapse).
Patients with residual microscopic inflammation (only detectable on a biopsy analysis) are more likely to relapse, therefore histologic remission “may represent the ultimate therapeutic goal,” the researchers noted.
However, both colonoscopy and biopsy analyses require trained clinicians. And the interpretation of results can differ between examiners, which poses problems of reproducibility and consistency.
“The interpretation of endoscopic images is subjective and based on the experience of individual endoscopists, thereby making the standardization of evaluation and real-time characterization challenging,” Kento Takenaka, PhD, lead author of the study, said in a press release.
In searching for a reliable method, researchers at Tokyo Medical and Dental University, in collaboration with Sony, created an AI algorithm capable of automatically analyzing endoscopic images with a level of accuracy equivalent to that of expert endoscopists.
The algorithm also was created with the goal of predicting histological remission with no need for biopsies, which are costly and carry risks.
“AI medical analysis systems are characterized by a shorter screening time and the absence of fatigue, making it possible to receive results more immediately after the endoscopic examination,” the researchers wrote.
The team developed a deep neural network algorithm — a form of machine learning — using 40,758 images of colonoscopies and 6,885 biopsy results from 2,012 UC patients referred to a single center in Japan (trial number UMIN000031430).
This first set of images was used to train the new algorithm to rate each image according to the UC endoscopic index of severity (UCEIS) — a tool used by physicians to rate the severity of UC — as well as to predict endoscopic and histologic remission.
Endoscopic remission was defined as a UCEIS score of 0, whereas histological remission (biopsy analysis) was defined as a Geboes score, which grades inflammation in the large intestine, of 3 points or less.
After being trained, the algorithm was validated by comparing the model’s outputs (remission and severity scores) with 4,187 colonoscopy and 4,104 biopsy findings judged by experts in a different group of 875 patients.
Among endoscopists, a discrepancy rate of 27.1% in relation to UCEIS scores was recorded, while a discrepancy rate of 13.2% was noted among specialists in biopsy analysis (pathologists).
When compared to evaluations made by medical experts, the algorithm was highly sensitive (93.3% and 92.4%), specific (87.8% and 93.5.%), and had a considerable diagnostic accuracy (90.1% and 92.9%) for predicting endoscopic and histological remission, respectively.
“We found that the DNUC [deep neural network for evaluation of UC] achieved a level of accuracy that was equivalent to that of expert endoscopists,” Takenaka said. “Thus, our system was able to predict histologic remission from UC using endoscopic images only, as opposed to both histological and endoscopic data. This represents an important development given the costs and risks associated with biopsies.”
“Our results supported that the DNUC offered objective, consistent, and real-time endoscopic evaluation results,” the researchers added.
The team suggested situations in which the algorithm could be helpful, including as an alternative to central analysis during clinical trials, and for training junior gastroenterologists.