Artificial intelligence for early gastric cancer boundary recognition in NBI and nF-NBI endoscopic images.

Abstract

Objectives: Precise delineation of early gastric cancer (EGC) margins is essential for complete resection during endoscopic submucosal dissection. This study aimed to develop deep learning-based models for EGC boundary detection in narrow-band imaging (NBI) and near-focus NBI (NF-NBI) images.

Methods: A total of 1215 NBI and 1646 NF-NBI images from EGC patients were used to train three convolutional neural networks (CNN1-CNN3), generating six deep learning models (Model1-Model6). Segmentation performance was compared among models and endoscopists of varying seniority.

Results: On NBI images, Model3 achieved an accuracy of 0.9348, compared to 0.7272, 0.7277, and 0.9435 for junior, intermediate, and senior endoscopists, respectively. The corresponding Dice coefficients were 0.8310 (95% CI, 0.8120-0.8500), 0.6153 (95% CI, 0.5827-0.6480), 0.6528 (95% CI, 0.6237-0.6819), and 0.8360 (95% CI, 0.8169-0.8550), with recall values of 0.9773, 0.6845, 0.7596, and 0.9784, respectively. On NF-NBI images, Model6 showed an accuracy of 0.9483, compared to 0.6885 (junior), 0.7826 (intermediate), and 0.9621 (senior endoscopists). Dice coefficients were 0.8526 (95% CI, 0.8410-0.8642), 0.6757 (95% CI, 0.6569-0.6944), 0.7161 (95% CI, 0.6941-0.7382), and 0.8618 (95% CI, 0.8512-0.8725), with recall values of 0.9831, 0.8095, 0.8317, and 0.9889, respectively.

Conclusions: The proposed deep learning models accurately delineated EGC boundaries in NBI and NF-NBI images, achieving diagnostic performance comparable to that of senior endoscopists.