Ranked Features Selection with MSBRG Algorithm and Rules Classifiers for Cervical Cancer

Abstract

In this paper, an automatic three-phase cervical cancer diagnosis system is employed which includes feature extraction, feature selection followed by classification. Firstly, the modified seed-based region growing (MSBRG) algorithm is implemented for automatic segmentation and feature extraction using 500 cervical cancer cells. Processes to obtain the threshold values and the initial seed location are carried out automatically using moving k-mean (MKM) algorithm and invariant moment techniques. Secondly, eight attribute evaluators are applied for selecting and ranking the features, which are Correlation-based Feature Selection, Classifier Attribute Evaluator, Correlation Attribute Evaluator, Gain Ratio, Info Gain, OneR, ReliefF, and Symmetrical Uncertainty. Finally, the classification is compared based on five classifiers: Decision Table, JRip, OneR, PART, and ZeroR. The performance of the classifiers is evaluated using 3 test options: the training percentage splits (50% to 98%), the full training data and the cross validation (2-fold to 10-fold). The experimental results prove the capability of the MSBRG algorithm as an automatic feature extraction method. Furthermore, this paper proves the ability of the ranked feature selection methods to select important features of a cervical cell, and favors the Decision Table as the best classifier for cervical cancer classification.