Classification of Colorectal Polyps from Narrow-Band Imaging Using Deep Learning

Colorectal cancer, a leading cause of cancer-related deaths worldwide. It can often be prevented through early detection and removal of precancerous polyps. Artificial intelligence (AI) can improve the accuracy of colorectal polyp evaluation, aiding in decisions about removal and pathology examination. This study is aimed to compare the performance of colorectal polyp assessments in narrow-band imaging (NBI) mode images between AI and experienced endoscopist. In this deep learning model development, a retrospective dataset of 1,648 polyp images were collected and divided into training (1,198 images), validation (100 images), and testing sets (350 images). Polyps were categorized as adenoma (requiring removal) or hyperplastic (non-dangerous), confirmed by pathology results. Seven deep learning models, including EfficientNetB5, Xception, InceptionV3, ResNet50, VGG16, ResNet34 and ResNeSt50, were applied. Only the best-performing model was selected to classify the images in the testing sets. Two endoscopist (who have over 5 years of experience) independently classified the testing set using the NBI International Colorectal Endoscopic (NICE) classification system. McNemar's test was used to compare the classification results between model and endoscopist. The Xception model had the highest accuracy of 73.43% and a negative predictive value (NPV) of 81.09%. The endoscopist had an accuracy of 69.21% and an NPV of 76.84%. The differences were not statistically significant, with p-value = 0.3149, between AI and endoscopist. This result shows the capability of integrating AI for automated polyp classification into clinical practice.