Secondary Logo

Journal Logo

Can Artificial Intelligence (AI) Achieve Real-Time ‘Resect and Discard‘ Thresholds Independently of Device or Operator?

221

Zachariah, Robin MD1; Ninh, Andrew2; Dao, Tyler3; Requa, James2; Karnes, William MD1

American Journal of Gastroenterology: October 2018 - Volume 113 - Issue - p S129
ACCEPTED: COLORECTAL CANCER PREVENTION
Free

1. University of California Irvine Medical Center, Orange, CA;

2. Docbot, Irvine, CA;

3. Docbot, Orange, CA

Introduction: A scope- and operator-independent method to optically diagnose diminutive polyps could make “Resect and Discard” a reality and save over $1 billion/year in the US alone. The American Society for Gastrointestinal Endoscopy’s (ASGE) Preservation and Incorporation of Valuable endoscopic Innovations (PIVI) “Resect and Discard” guideline requires a > 90% negative predictive value (NPV) for diminutive adenomas and > 90% concordance in recommended surveillance intervals comparing optical pathology to histology. Convolutional neural networks (CNN) have potential to predict polyp pathology and meet PIVI guidelines independently of operator or scope manufacturer.

Methods: We developed an optical pathology (OP) model using a CNN built on Tensorflow and pretrained on ImageNet. 5456 high quality images of unique adenomas and serrated polyps of known locations, size, and light source (white light [WL] or narrow band imaging [NBI]) were extracted from our endoscopic database. Images were partitioned into 5 equal-sized subsamples for 5-fold cross validation with training (80%) and validation (20%). An Adam optimizer generates a probability between 0-0.5 (serrated) and 0.5-1 (adenoma). Surveillance intervals were calculated based on US Multi-Society Task Force guidelines, comparing OP vs. true pathology (TP).

Results: Among polyps throughout the colon, NPV for adenomas was 92% (WL) and 93% (NBI). Surveillance interval concordance between OP and TP for screening and surveillance cases was 93% and 96%, respectively. Among diminutive polyps (< 6 mm) throughout the colon, NPV for adenomas was 91% (WL) and 92% (NBI). Surveillance concordance was 93% and 96% for screening and surveillance cases, respectively. Among diminutive polyps in the left colon, NPV improved to 97% (WL) and 95% (NBI). The model processes > 90 frames per second and provides real-time feedback during colonoscopy using a conventional desktop and graphics processing unit.

Conclusion: Without stringencies (no unclassifiable polyps), our optical pathology model meets “Resect and Discard” PIVI guidelines and provides operator-independent and real-time feedback during colonoscopy. Accuracy is unaffected by light source, suggesting it may work well with any scope manufacturer. Adenoma detection rate and surveillance recommendations at the time of colonoscopy are added benefits. Blinded multicenter studies utilizing multiple scope manufacturers are needed to validate these potentials.

© The American College of Gastroenterology 2018. All Rights Reserved.