Accurate Classification of Diminutive Colorectal Polyps Using Computer-Aided Analysis

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Narrow-band imaging (NBI) enhances endoscopic visualization of colorectal microstructures.
• Accurate differentiation of hyperplastic from neoplastic polyps using NBI requires specialized expertise.
• Computer-aided diagnosis (CAD) systems can aid in polyp classification.

Purpose of the Study:

• To develop and evaluate a deep neural network computer-aided diagnosis (DNN-CAD) system for analyzing NBI images of diminutive colorectal polyps.
• To compare the diagnostic performance of DNN-CAD with expert and novice endoscopists.

Main Methods:

• A deep neural network (DNN) was trained on 1476 neoplastic and 681 hyperplastic colorectal polyp images.
• The DNN-CAD system was tested on a validation set of 96 hyperplastic and 188 neoplastic polyps (<5 mm).
• Diagnostic accuracy, sensitivity, specificity, PPV, NPV, and time were compared between DNN-CAD and 6 endoscopists (2 expert, 4 novice).

Main Results:

• DNN-CAD achieved 96.3% sensitivity, 78.1% specificity, 89.6% PPV, and 91.5% NPV.
• DNN-CAD classified polyps significantly faster (0.45s) than expert (1.54s) and novice (1.77s) endoscopists (P < .001).
• DNN-CAD demonstrated perfect intra-observer agreement (kappa=1), while endoscopists showed low agreement.

Conclusions:

• The developed DNN-CAD system effectively identifies neoplastic and hyperplastic colorectal polyps with high accuracy and speed.
• DNN-CAD shows potential for improving endoscopic image analysis and other medical imaging applications.
• This AI-driven approach can assist endoscopists in real-time polyp diagnosis.