S133 Video Validation of a Novel Multi-Class Neural Network for In-situ Polyp Pathology Characterization

Abstract

Introduction: Polyp pathology is essential in determining malignancy risk, surveillance intervals, and quality measurements in colonoscopy. The majority of polyps are resected and characterized manually by pathologists, which has a large cost and time burden. In-situ polyp characterization can potentially reduce costs and facilitate point-of-care decision-making. We previously developed and validated an optical pathology deep neural network for distinguishing adenomas from serrated polyps. Our goal was to create and validate a multiclass neural net to further distinguish serrated polyps into categories of hyperplastic and sessile serrated polyps. Methods: Our training set included 1,000 unique polyp video clips and 25,000 freeze-frame polyp images annotated for polyp pathology, including nearly 6000 unique adenomas, 3000 unique hyperplastic polyps, and 1500 unique sessile serrated polyps. A 6-step multi-stage training pipeline was devised to detect and extract polyp pathology features from individual frames and classify polyp pathology using Multiple Instance Learning. For validation, we utilized a fresh set of deidentified 171 video clips of unique polyps with known pathology. Results: The multi-class optical pathology system correctly characterized 94/99 (95%) polyps as adenoma and 51/59 (86.4%) as serrated. The negative predictive value (NPV) was 91.1%, positive predictive value (PPV) was 92.2%, and accuracy was 91.8%. Among serrated polyps, 16/16 (100%) were correctly characterized as SSP and 31/35 (88.5%) as HPs, with an NPV of 100%, PPV of 80%, and accuracy of 92.3%. Conclusion: Our optical pathology neural network accurately characterizes adenomas, HPs, and SSPs in-situ with an accuracy of >90%. Next-step prospective multicenter validation studies will determine if in-situ polyp characterization can automate ADR recording, provide accurate point-of-care surveillance intervals, and meet thresholds to reduce pathology cost through resect-and-discard or leave-alone strategies, with a potential cost savings of more than $1 billion per year.Figure 1.: Real-time sample images from the optical pathology system with confidence level. A) Hyperplastic polyp, B) SSA, C) AdenomaTable 1.: 3x3 confusion matrix comparing AI optical path results with true pathology results, including performance parameters.