Comparison Study of Deep Learning Models for Colorectal Lesions Classification

In this paper, we performed a comparison study between GoogLeNet, AlexNet, and InceptionV3 deep learning models to recognize and classify colorectal cancer tumors. Colorectal tumors are one of the very common cancer types and early detection could result in a significantly higher survival rate of 95% as opposed to 12%. In this work, we aim to investigate the deep learning models to automatically detect the tumor types from polyp images. We, therefore, used actual images taken from the colorectal surgery or colonoscopy using Narrow-band imaging (NBI). The images are classified based on NBI International Colorectal Endoscopic (NICE) classification. We used NICE 1 and NICE 2 types with a total of 2604 images in the size of 64x64. Our results show that the InceptionV3 model has the most accurate results by average 92.39% where AlexNet is 88.19% and GoogLeNet is 85.73%.