User login
A deep-learning algorithm for automatic polyp detection during colonoscopies showed both high sensitivity and high specificity, according to a study presented at the World Congress of Gastroenterology at ACG 2017.
The algorithm was set up using a retrospective set of 5,545 images annotated by colonoscopists, while the validation set used for the study consisted of 27,461 colonoscopy images from 1,235 patients, according to Pu Wang, MD, of the Sichuan Academy of Medical Sciences & Sichuan Provincial People’s Hospital, Chengdu, China, and his associates.
At the high-sensitivity operating point, the algorithm had a sensitivity of 94.96% and a specificity of 92.01%, and at the low–false positive rate operating point, the algorithm had a sensitivity of 92.35% and a specificity of 97.05%. The area under the curve was 0.958 in a receiver operating characteristic curve analysis.
In subgroup analyses of flat polyps, polyps less than or equal to 0.5 cm, and isochromatic polyps, the algorithm had areas under the curve of 0.943, 0.957, and 0.957 respectively.
The algorithm reported results in 60-80 ms, offering real-time assistance with polyp detection during colonoscopies, Dr. Wang and his colleagues noted.
The study was not funded by industry grants, and no disclosures were reported.
A deep-learning algorithm for automatic polyp detection during colonoscopies showed both high sensitivity and high specificity, according to a study presented at the World Congress of Gastroenterology at ACG 2017.
The algorithm was set up using a retrospective set of 5,545 images annotated by colonoscopists, while the validation set used for the study consisted of 27,461 colonoscopy images from 1,235 patients, according to Pu Wang, MD, of the Sichuan Academy of Medical Sciences & Sichuan Provincial People’s Hospital, Chengdu, China, and his associates.
At the high-sensitivity operating point, the algorithm had a sensitivity of 94.96% and a specificity of 92.01%, and at the low–false positive rate operating point, the algorithm had a sensitivity of 92.35% and a specificity of 97.05%. The area under the curve was 0.958 in a receiver operating characteristic curve analysis.
In subgroup analyses of flat polyps, polyps less than or equal to 0.5 cm, and isochromatic polyps, the algorithm had areas under the curve of 0.943, 0.957, and 0.957 respectively.
The algorithm reported results in 60-80 ms, offering real-time assistance with polyp detection during colonoscopies, Dr. Wang and his colleagues noted.
The study was not funded by industry grants, and no disclosures were reported.
A deep-learning algorithm for automatic polyp detection during colonoscopies showed both high sensitivity and high specificity, according to a study presented at the World Congress of Gastroenterology at ACG 2017.
The algorithm was set up using a retrospective set of 5,545 images annotated by colonoscopists, while the validation set used for the study consisted of 27,461 colonoscopy images from 1,235 patients, according to Pu Wang, MD, of the Sichuan Academy of Medical Sciences & Sichuan Provincial People’s Hospital, Chengdu, China, and his associates.
At the high-sensitivity operating point, the algorithm had a sensitivity of 94.96% and a specificity of 92.01%, and at the low–false positive rate operating point, the algorithm had a sensitivity of 92.35% and a specificity of 97.05%. The area under the curve was 0.958 in a receiver operating characteristic curve analysis.
In subgroup analyses of flat polyps, polyps less than or equal to 0.5 cm, and isochromatic polyps, the algorithm had areas under the curve of 0.943, 0.957, and 0.957 respectively.
The algorithm reported results in 60-80 ms, offering real-time assistance with polyp detection during colonoscopies, Dr. Wang and his colleagues noted.
The study was not funded by industry grants, and no disclosures were reported.
FROM WORLD CONGRESS OF GASTROENTEROLOGY
Key clinical point:
Major finding: The deep-learning algorithm had a sensitivity of 95% and a specificity of 92% when adjusted to the high-sensitivity operating point.
Data source: A set of 27,461 colonoscopy images from 1,235 patients.
Disclosures: The study was not funded by industry grants, and no disclosures were reported.
