Effective and efficient classification of gastrointestinal lesions: combining data preprocessing, feature weighting, and improved ant lion optimization

Abstract

This paper presents an approach that combines the data preprocessing, feature weighting, and the improved antlion optimization algorithm for effective and efficient classification of gastrointestinal lesions. A high-dimensional gastrointestinal lesion dataset that consists of extracted texture, color, and shape features from the colonoscopy videos is obtained from the UCI repository. The data has certain imperfections such as the presence of zero-valued features, outliers, and dominant features. So, it is preprocessed to cope with these problems. Then, feature weighting is used to boost the classification performance by assigning the weights to the features according to their relevance in classification. The improved antlion optimization algorithm is used to search for feature weights and the parameters of the Support Vector Machines simultaneously. The experiments are performed using different combinations of features and endoscopic images to analyse the performance. The outcomes show that the combination of texture and color features from NBI images is the best. The accuracy of 97.37% and 98.68% for multi-class and binary classification problems respectively is attained using only $$\sim$$ 31% features. Moreover, feature reduction helps to lower the runtime of the classifier by approx. 60%. In conclusion, a better approach is presented for colorectal lesions classification that competes with well-experienced colonoscopists and outperforms the existing methods.