Feature Selection for Shape-Based Classification of Biological Objects

Abstract. In this paper, feature selection methodology from the machine learning literature is applied to the problem of shape-based classification. This methodology discards statistical features that are least relevant for classification and often improves the generalization ability of classifiers. In context of biological shape classification, feature selection can pinpoint, in a robust manner, the regions of objects where interclass differences are most pronounced. A new feature selection algorithm is developed by extending an existing support vector machine based algorithm to take advantage of locality properties of shape features. The performance of new algorithm is tested on synthetic and clinical data. The clinical data comes from a study of hippocampal shape in schizophrenia. The results on this data indicate that the head of the right hippocampus is significant for understanding the effects of schizophrenia.

• Type: Conference
• Year: 2003
• Where: IPMI
• Authors: Paul A. Yushkevich, Sarang C. Joshi, Stephen M. Pizer, John G. Csernansky, Lei E. Wang
• Keywords: Biological Shape Classification, Feature Selection Algorithm, Feature Selection Methodology, IPMI 2003, Medical Imaging