Computer-Aided Polyps Classification from Colonoscopy Using Deep Learning Models

Abstract

Abstract Medical imaging has advanced to the extent that conditions including stomach ulcers, bleeding, and polyps can be diagnosed using video endoscopy. It takes a lot of time for doctors to follow up on all the images produced by medical video endoscopy. This complicates the use of labor. Automated diagnosis through computer aided approaches to analyze all the resulting images rapidly and accurately. The proposed methodology is innovative in that it seeks to create a system for diagnosing gastrointestinal disorders. The images that are sent into the deep learning networks have all been improved and have had the noise removed. The 5000 images in the Kvasir dataset are evenly split between five different categories affecting the digestive tract: dye-lifted polyps, dyed resection margins, normal cecum, polyps, and ulcerative coliti. Five finely tuned deep convolutional neural network architectures (Xception, ResNet-101, VGG-19, EfficientNetB2v3, and MobineNetV2) with weights from the ImageNet dataset. EffecientNetV2B3 outperformed and achieved accuracy of 96.0%.