A feature selection analysis in breast cancer diagnosis

In this paper, it is aimed to design a computer aided diagnosis system for breast cancer diagnosis and a mamogram dataset prepared during the Image Retrieval in Medical Applications (IRMA) project is used for the verification of the system. In accordance with this purpose, feature extraction is realized using Local Configuration Pattern algorithm on the pre-processed mamogram images by histogram equalization followed by Non-Local Means Filtering. In addition to these features, vector space is extended by some statistical and frequency-domain features. Besides, feature selection is performed by applying Sequential Forward Feature Selection (SFS) algorithm on the obtained features. Finally, selected features are classified in a 2-stage scheme into 3 different categories (normal, benign, malignant) using linear discriminant classifier, Fisher's linear discriminant analysis, logistic linear classifier, k-nearest neighbor classifier, Naïve Bayes and decision tree classifiers. The results attained at the cases, in which feature selection is performed and not, are compared and it is concluded with approximately 88 % maximum success rate is accomplished in both cases. This success rate is achieved using logistic linear classifier when 15 features are selected among 108 features via SFS algorithm. Analyzing the success performance of all classifiers in both cases, appropriateness of feature selection is decided by means of data storage, memory occupation, and computational time.