Polyp characterization using deep learning and a publicly accessible polyp video database

Abstract

Objectives: Convolutional neural networks (CNN) for computer‐aided diagnosis of polyps are often trained using high‐quality still images in a single chromoendoscopy imaging modality with sessile serrated lesions (SSLs) often excluded. This study developed a CNN from videos to classify polyps as adenomatous or nonadenomatous using standard narrow‐band imaging (NBI) and NBI‐near focus (NBI‐NF) and created a publicly accessible polyp video database. Methods: We trained a CNN with 16,832 high and moderate quality frames from 229 polyp videos (56 SSLs). It was evaluated with 222 polyp videos (36 SSLs) across two test‐sets. Test‐set I consists of 14,320 frames (157 polyps, 111 diminutive). Test‐set II, which is publicly accessible, 3317 video frames (65 polyps, 41 diminutive), which was benchmarked with three expert and three nonexpert endoscopists. Results: Sensitivity for adenoma characterization was 91.6% in test‐set I and 89.7% in test‐set II. Specificity was 91.9% and 88.5%. Sensitivity for diminutive polyps was 89.9% and 87.5%; specificity 90.5% and 88.2%. In NBI‐NF, sensitivity was 89.4% and 89.5%, with a specificity of 94.7% and 83.3%. In NBI, sensitivity was 85.3% and 91.7%, with a specificity of 87.5% and 90.0%, respectively. The CNN achieved preservation and incorporation of valuable endoscopic innovations (PIVI)‐1 and PIVI‐2 thresholds for each test‐set. In the benchmarking of test‐set II, the CNN was significantly more accurate than nonexperts (13.8% difference [95% confidence interval 3.2–23.6], P = 0.01) with no significant difference with experts. Conclusions: A single CNN can differentiate adenomas from SSLs and hyperplastic polyps in both NBI and NBI‐NF. A publicly accessible NBI polyp video database was created and benchmarked.