Artificial intelligence accurately classifies small colorectal polyps

Reuters Health Information: Artificial intelligence accurately classifies small colorectal polyps

Artificial intelligence accurately classifies small colorectal polyps

Last Updated: 2017-10-31

By Scott Baltic

NEW YORK (Reuters Health) - An artificial intelligence (AI) system was able to equal the performance of experienced endoscopists, and surpass that of novice endoscopists, in evaluating digital images of small (<5 mm) colorectal polyps, according to research from Taiwan.

The AI system classified polyps as neoplastic or hyperplastic with an accuracy of 90.1%. By comparison, two endoscopists with at least five years of experience achieved accuracies of 90.5% and 87.0%. Four novice endoscopists with only one year of experience ranged from 88.0% to 80.3% accuracy, two of them significantly less accurate than the AI system.

AI also performed more rapidly than the endoscopists (mean classification time, 0.45 sec vs. 1.54 sec for the experts and 1.77 sec for the novices).

The system, called DNN-CAD (computer-aided diagnosis with a deep neural network), was developed through a collaboration between physicians and computer scientists. It is based on Google's TensorFlow, an open-source software library for machine intelligence.

The report was published online on October 14 in Gastroenterology.

"DNN-CAD can offer consistently high diagnostic accuracy with very fast diagnostic time, and avoid errors or inconsistency resulting from human fatigue or disagreement between different endoscopists," corresponding authors Dr. Vincent S. Tseng, of the Institute of Data Science and Engineering at National Chiao Tung University, Hsinchu, Taiwan, and Dr. Peng-Jen Chen, of Tri-Service General Hospital, National Defense Medical Center, Taiwan, told Reuters Health by email.

The research group, they added, plans to develop a real-time version of DNN-CAD for clinical use, with extended classifications of colorectal polyps, such as hyperplastic polyps, adenoma/carcinoma in situ or advanced cancer, to help endoscopists make proper treatment decisions.

The researchers first trained the DNN-CAD by exposing it to 2,157 narrow-band images of both neoplastic and non-neoplastic polyps, from which the software learned to distinguish between the two. They then tested the software with 284 separate new images, which the six endoscopists also evaluated. Histologic findings were the standard of comparison.

"Pre-cancerous colon polyps can be removed during colonoscopy, arresting the natural progression from polyp to cancer and preventing cancer altogether," Dr. Swati G. Patel, director of the Gastrointestinal Cancer Risk and Prevention Clinic, University of Colorado Anschutz Medical Center, explained in an email to Reuters Health. She was not involved in the current study.

Polyps <5 mm are those most often found during routine colonoscopy, she said. Although a large portion of them (hyperplastic polyps) do not harbor any pre-cancerous potential and the rest (adenomas) rarely harbor malignancy, they are routinely removed to confirm histology.

However, Dr. Patel added, if the polyp histology can be determined in real time during colonoscopy, endoscopists can leave hyperplastic polyps in place and remove only the precancerous polyps, reducing costs and minimizing risk to the patient.

She noted that while narrow-band imaging allows an adequately trained endoscopist to highlight superficial blood-vessel patterns and distinguish hyperplastic from adenomatous polyps, not all endoscopists reach expert societies' performance benchmarks - and that even those who do, need ongoing monitoring to make sure they maintain proficiency.

The researchers "have done remarkable work to automate narrow-band imaging interpretation of diminutive colorectal polyps. . . . This system surpassed the benchmarks set forth by the American Society for Gastrointestinal Endoscopy and performed better than the endoscopists who participated in the study," Dr. Patel said.

The technology, she concluded, "can help take the human error out of real-time characterization of colorectal polyps and has the potential to be cost-saving without compromising patient care." The main remaining question, she says, is whether the computer-aided diagnostic platform can be made available in routine endoscopy practices.


Gastroenterology 2017.

© Copyright 2013-2019 GI Health Foundation. All rights reserved.
This site is maintained as an educational resource for US healthcare providers only. Use of this website is governed by the GIHF terms of use and privacy statement.