Hierarchical classifier of cytoskeletal organization

Quantitative morphology has already been shown by these authors to describe the processes of cytoskeletal disorganisation and re-organisation related to the exposure of cultured cells to xenobiotics [G F Crosta, C Urani, L Fumarola, Estimating Structural Damage of the Cytoskeleton by means of Morphological Descriptors, Proc. SPIE 5461 Biophotonics New Frontier: From Genome to Proteome, M D Faupel, P P Meyrueis, Eds., SPIE: Bellingham, WA, 2004, to appear]. Namely, an automatic classifier was developed, which relies on morphological descriptors. Epifluorescence images of cytoskeletal microtubules were processed by Fourier analysis (“spectrum enhancement”), fractal analysis (mass and contour) and spatial differentiation aimed at extracting morphological descriptors. The classifier was successfully applied to rank the structural damage and re-organisation of microtubules. This work focusses on the automatic training of the classifier by a two-layer hierarchical scheme. On the first layer the parameters which control the extraction of the Fourier descriptors are chosen according to some optimality condition (either by solving a saddle point problem or maximising specificity and selectivity); reduction of the Fourier descriptors is obtained by selecting the first few principal components. The latter are passed on to the second layer and merged with descriptors from fractal analysis and spatial differentiation. As a consequence, classification improves. Internal and external (by linear discriminant analysis) validation results will also be presented. The scheme is being applied to images of microtubules and microfilaments from different toxicological experiments.