Multiclass Gastrointestinal Diseases Classification Based on Hybrid Features and Duo Feature Selection

Abstract

Gastrointestinal Tract (GIT) infections are quite common nowadays. If these abnormalities are left untreated at early stages, they may develop into stomach cancers. Wireless Capsule Endoscopy (WCE) is a method that enables medical professionals to view the internal parts of the GIT and take pictures using a pill camera. Manual detection of abnormalities from the taken images is time-consuming and may lead to misdiagnosis. Several Computer-based methods were developed in this domain, but improving prediction accuracy is still challenging due to the complex textures, colours, irregularities of tissues and quality of images. To address this issue, a novel technique has been introduced in this research based on color, texture, statistical, shape and deep pretrained Densenet features from contrast-enhanced GI images. The extracted features are fused to form a powerful features subset. From the fused features, the minimal-optimal feature subset is selected using the two-stage ReliefF-minimum Redundancy Maximum Relevance (R-mRMR) method and fed to One Against All Support Vector Machine (OAA-SVM) for classification. The proposed work is validated using 8000 images with eight classes of KVASIR V2 and attained the maximum classification accuracy of 99.2% and precision of 99.1%.