Feature Selection in Classification of Biomedical High-Resolution Micro-CT Images
Benjamin Auffarth · 2007
This thesis presents a work whose general objective is classification based on characteristics of voxel from biomedical imagery. Because classification performance depends on the selection of features extracted from images, we study several features selection methods based on two kinds of filters, one for measuring relevance of features for target prediction, the other for measuring redundancy between features. We propose the application of different measures of distributional similarity, develop the value difference metric (VDM, [101]) as a measure for relevance and redundancy, and introduce a new measure, called “fit criterion “ (FC). We also propose a novel selection method, a Hopfield Network. We benchmark selection methods using unitary redundancy and relevance filters, a greedy algorithm with redundancy thresholds[31], the min-redundancy max-relevance integration[29; 85], and the Hopfield Network selection. We conclude that in our setting, min-redundancy max-relevance is the best selection method for few variables, but the Hopfield network feature selection method outperforms when more features are taken into account. FC was a good relevance measure, but a bad redundancy measure. VDM was very good in our experiments as both redundancy and relevance measure.