Abstract
Feature subset selection, applied as a pre- processing step to machine learning, is valuable in dimensionality reduction, eliminating irrelevant data and improving classifier performance. In the classic formulation of the feature selection problem, it is assumed that all the features are available at the beginning. However, in many real world problems, there are scenarios where not all features are present initially and must be integrated as they become available. In such scenarios, online feature selection provides an efficient way to sort through a large space of features. It is in this context that we introduce online feature selection for the classification of emphysema, a smoking related disease that appears as low attenuation regions in High Resolution Computer Tomography (HRCT) images. The technique was successfully evaluated on 61 HRCT scans and compared with different online feature selection approaches, including hill climbing, best first search, grafting, and correlation-based feature selection. The results were also compared against “density maskâ€, a standard approach used for emphysema detection in medical image analysis.