Subscribe to RSS
DOI: 10.1055/s-0043-1766048
Can artificial intelligence (AI) aid in the sizing of colorectal polyps in real-time?
Aims “Resect and discard” strategy for colorectal polyps is based on the negligible risk of cancer in diminutive (<5mm) polyps and thus, accurately sizing the polyp during endoscopy is an important decision-making process. We aim to develop and validate an AI system using a deep convolutional neural network to classify the polyps into non-diminutive and diminutive polyps.
Methods Using VGG-16 architecture, 63,063 images of 512 polyps were used to train AI system for sizing. In phase 1 of the study, the system was tested on pre-recorded white light images of the polyps. In phase 2, the system was tested on prospective real-time colonoscopies done by expert endoscopist. Ground truth in both phases was based on the consensus of 3 expert endoscopists (>1000 lifetime colonoscopies).
Results 292 polyps (106 non-diminutive and 186 diminutive) were included in Phase 1 study. Sensitivity, specificity, and accuracy of AI system classifying into non-diminutive polyps were 93.40%, 79.03% and 84.25% respectively.
In Phase 2, 135 polyps (75 non-diminutive (median size of 10mm) and 60 diminutive (median size of 4mm)) from 38 colonoscopies were included. The sensitivity, specificity, and accuracy of AI system sizing into non-diminutive polyps were 82.67%, 88.33% and 85.19% respectively.
Conclusions Our data from both phases of the study demonstrates that AI-based sizing is effective in differentiating non-diminutive from diminutive polyps. This can facilitate the introduction of “resect and discard” strategy ([Table 1]).
Publication History
Article published online:
14 April 2023
© 2023. European Society of Gastrointestinal Endoscopy. All rights reserved.
Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany