User login
WASHINGTON –
When tested on 36,076 archived colonoscopy images, the polyp recognition program developed through machine learning had a sensitivity for polyp detection of 98%, a specificity of 93%, a positive predictive value of 0.758 and a negative predictive value of 0.995, Priyam V. Tripathi, MD, said at the annual Digestive Disease Week.® The program also showed an area under the receiver operator characteristic curve of 0.99, indicating nearly perfect ability to discriminate between images of polyps and normal colonic tissue, said Dr. Tripathi, a gastroenterologist at the University of California, Irvine.
She and her associates initially developed the polyp-recognition program with machine learning engineering that involved 4,088 images of polyps and 4,553 images of normal tissue drawn from the extensive colonoscopy video archive maintained at UC Irvine. Refinement of the program continues as it undergoes further use. The program can review 98 images a second, making it more than fast enough to aid during real-time colonoscopy examinations, Dr. Tripathi explained in a video interview. As an operator withdraws the colonoscope and views the images, the program is designed to monitor the pictures along with the operator and trigger alerts that flag high-probability lesions by framing them in a colored box on the screen. The operator can then examine these sites with more attention and decide whether they warrant biopsy or polypectomy.
A second validation study used the program to review 20 archived colonoscopy videos along with an expert panel. During the original examinations, the operators of these 20 procedures identified 28 polyps. The expert review confirmed these 28 and identified eight additional polyps. The researchers then assessed the same videos with the recognition program and confirmed the original 28 plus the added eight and also found nine additional polyps that had been missed twice by clinicians. Dr. Tripathi and her associates recently published results from this validation study (Gastroenterology. 2018 Jun 18. doi: 10.1053/j.gastro.2018.06.037).
The next step is a prospective, multicenter study to compare the adenoma detection rate of operators aided by the recognition program with their detection rate without the program, she said.
“The adenoma detection rate is the key marker,” noted William E. Karnes, MD, a gastroenterologist at UC Irvine and senior investigator on these studies. “If the adenoma detection rate rises, we won’t know whether it’s the artificial intelligence that’s the reason, or whether it’s the artificial intelligence watching the operator” and motivating the gastroenterologist to do a more thorough job, Dr. Karnes noted in an interview. “But it doesn’t matter as long as the software is easy to use. It can potentially close the gap in adenoma detection rates. There are a lot of missed polyps” in routine practice right now.
SOURCE: Tripathi PV et al. DDW 2018. Presentation 133.
WASHINGTON –
When tested on 36,076 archived colonoscopy images, the polyp recognition program developed through machine learning had a sensitivity for polyp detection of 98%, a specificity of 93%, a positive predictive value of 0.758 and a negative predictive value of 0.995, Priyam V. Tripathi, MD, said at the annual Digestive Disease Week.® The program also showed an area under the receiver operator characteristic curve of 0.99, indicating nearly perfect ability to discriminate between images of polyps and normal colonic tissue, said Dr. Tripathi, a gastroenterologist at the University of California, Irvine.
She and her associates initially developed the polyp-recognition program with machine learning engineering that involved 4,088 images of polyps and 4,553 images of normal tissue drawn from the extensive colonoscopy video archive maintained at UC Irvine. Refinement of the program continues as it undergoes further use. The program can review 98 images a second, making it more than fast enough to aid during real-time colonoscopy examinations, Dr. Tripathi explained in a video interview. As an operator withdraws the colonoscope and views the images, the program is designed to monitor the pictures along with the operator and trigger alerts that flag high-probability lesions by framing them in a colored box on the screen. The operator can then examine these sites with more attention and decide whether they warrant biopsy or polypectomy.
A second validation study used the program to review 20 archived colonoscopy videos along with an expert panel. During the original examinations, the operators of these 20 procedures identified 28 polyps. The expert review confirmed these 28 and identified eight additional polyps. The researchers then assessed the same videos with the recognition program and confirmed the original 28 plus the added eight and also found nine additional polyps that had been missed twice by clinicians. Dr. Tripathi and her associates recently published results from this validation study (Gastroenterology. 2018 Jun 18. doi: 10.1053/j.gastro.2018.06.037).
The next step is a prospective, multicenter study to compare the adenoma detection rate of operators aided by the recognition program with their detection rate without the program, she said.
“The adenoma detection rate is the key marker,” noted William E. Karnes, MD, a gastroenterologist at UC Irvine and senior investigator on these studies. “If the adenoma detection rate rises, we won’t know whether it’s the artificial intelligence that’s the reason, or whether it’s the artificial intelligence watching the operator” and motivating the gastroenterologist to do a more thorough job, Dr. Karnes noted in an interview. “But it doesn’t matter as long as the software is easy to use. It can potentially close the gap in adenoma detection rates. There are a lot of missed polyps” in routine practice right now.
SOURCE: Tripathi PV et al. DDW 2018. Presentation 133.
WASHINGTON –
When tested on 36,076 archived colonoscopy images, the polyp recognition program developed through machine learning had a sensitivity for polyp detection of 98%, a specificity of 93%, a positive predictive value of 0.758 and a negative predictive value of 0.995, Priyam V. Tripathi, MD, said at the annual Digestive Disease Week.® The program also showed an area under the receiver operator characteristic curve of 0.99, indicating nearly perfect ability to discriminate between images of polyps and normal colonic tissue, said Dr. Tripathi, a gastroenterologist at the University of California, Irvine.
She and her associates initially developed the polyp-recognition program with machine learning engineering that involved 4,088 images of polyps and 4,553 images of normal tissue drawn from the extensive colonoscopy video archive maintained at UC Irvine. Refinement of the program continues as it undergoes further use. The program can review 98 images a second, making it more than fast enough to aid during real-time colonoscopy examinations, Dr. Tripathi explained in a video interview. As an operator withdraws the colonoscope and views the images, the program is designed to monitor the pictures along with the operator and trigger alerts that flag high-probability lesions by framing them in a colored box on the screen. The operator can then examine these sites with more attention and decide whether they warrant biopsy or polypectomy.
A second validation study used the program to review 20 archived colonoscopy videos along with an expert panel. During the original examinations, the operators of these 20 procedures identified 28 polyps. The expert review confirmed these 28 and identified eight additional polyps. The researchers then assessed the same videos with the recognition program and confirmed the original 28 plus the added eight and also found nine additional polyps that had been missed twice by clinicians. Dr. Tripathi and her associates recently published results from this validation study (Gastroenterology. 2018 Jun 18. doi: 10.1053/j.gastro.2018.06.037).
The next step is a prospective, multicenter study to compare the adenoma detection rate of operators aided by the recognition program with their detection rate without the program, she said.
“The adenoma detection rate is the key marker,” noted William E. Karnes, MD, a gastroenterologist at UC Irvine and senior investigator on these studies. “If the adenoma detection rate rises, we won’t know whether it’s the artificial intelligence that’s the reason, or whether it’s the artificial intelligence watching the operator” and motivating the gastroenterologist to do a more thorough job, Dr. Karnes noted in an interview. “But it doesn’t matter as long as the software is easy to use. It can potentially close the gap in adenoma detection rates. There are a lot of missed polyps” in routine practice right now.
SOURCE: Tripathi PV et al. DDW 2018. Presentation 133.
REPORTING FROM DDW 2018
Key clinical point: New software provides real-time aid to colonoscopists for identifying adenomatous polyps.
Major finding: The software was 98% sensitive, 93% specific, and had a negative predictive value of 0.995 for ruling out adenomatous polyps.
Study details: Single-center review of 36,076 images of polyps and normal colonic tissue.
Disclosures: The software development has no commercial funding. Dr. Tripathi had no disclosures. Dr. Karnes is cofounder of Docbot/Qualoscopy, a company that markets colonoscopy software.
Source: Tripathi PV et al. DDW 2018. Presentation 133.
