Automatic Feature Selection for Biological Shape Classification in Σ YNERGOS

This work reports the development of a versatile framework allowing the characterization and analysis of computer vision techniques as well as their applications to biological shapes, with attention focused on neural cells. The proposed framework has been implemented within the Σynergos system, a powerful imaging laboratory that includes, among other features, tools for performance assessment of computer vision techniques, image databases, real-time processing by using distributed systems and interface with the Internet. The motivations for the development of such a framework: (i) the importance of biological shape analysis; (ii ) its potential as an effective tool for the systematic assessment of image processing and analysis techniques; and (iii ) the possibil ity of conducting extensive characterizations of biological shapes. The paper describes an experiment to assess multiscale shape features for complexity characterization, which have been adopted for the classification of two types of ganglion neural cells (cat), namely α and β. This experiment involves: (1) a training stage where the k-means clustering algorithm learns the prototypes of each class from the database; (2) the neurons in the database are classified; (3) the classification results are compared to the original classes; and (4) the number of misclassifications is determined. The genetic algorithm is used as a means of effectively investigating the N-dimensional spaces defined by the parameter configurations.