Image classification and recognition from sub-cell to tissue scale

Diagnostic images at all biological scales benefit from quantitative morphological analysis which can lead to their classification and recognition. With reference to an image from an optical microscope, quantitative morphology consists of extracting from the raw image, by a suitable algorithm, some numerical indicators or "descriptors". The latter replace the image and represent it for subsequent analysis by algorithms capable of: classification, recognition, feature extraction and other higher level functions. Herewith the extraction of morphological descriptors relies on the "spectrum enhancement" algorithm: the discrete Fourier transform of the image undergoes some nonlinear operatons and comparison with a model power spectral density. As a result the graph of a function of wavenumber, the enhanced spectrum, is obtained, which carries information about the structure and texture of the image. Enhanced spectra are submitted to a classifier based on multivariate statistics, which can be trained, validated and applied to recognize new images. The whole procedure will be illustrated by applications at different scales: a) (sub-cell) cytoskeletal structures altered by xenobiotics; b) (colony) cell colonies ("foci") from in vitro neoplastic transformation assays; c) (tissue) cross sections of reconstituted human corneal epithelia. The results of diagnostic and pathological interest, which will be presented, include the quantitative estimation of morphological alterations, dose-response relationships and recovery from damage. The same paradigm can apply to images acquired by different (optical) instruments and is intended for integration with other methods, the eventual goal being the high content analysis of biological data.