Combining independent component analysis and self-organizing maps for cell image classification

Abstract

We consider the task of cell classification in fluorescent micrographs. We combine the use of Independent Component analysis as a preprocessing stepand a Self-organizing Map for the resulting ICA feature space to classify image patches into cell and noncell images and to investigate the features of imagepatches in the vicinity of the classification border. We compare the classification performance of ICA bases of different size, generated from applying the infomax algorithm to image eigenspaces of different dimensionalities. We find an optimal performance for intermediate dimensionalities, characterized by the ICA basis patterns exhibiting salient features of an “idealized” cell shape, and we achieve classification results comparable to a previous approach based on PCA features.