Printed in U.S.A. OBJECTIVE CELL IMAGE ANALYSIS’-3

A system of computer programs processes digitized images of cells. The gray value arrays are considered as a two-dimensionally extended stochastic process. Super-vised learning is employed to estimate the parameters of this process in different cell types, to derive classification algorithms and to compute a likelihood for cor-rect classification for each cell. These were found to range from 10:1 to 10:1 for tumor cells from the female genital tract. A data bank for tumor cells has been established. Nonsupervised learning algorithms are used to examine sets of cells for homogeneity. Synthesized cell images of known stochastic properties can be gen-erated to test the completeness of the derived classification rules. Under develop-ment are programs defining the changes observable in images of a given cell type during a disease process, by the deformation of the covariance matrix of image properties. The exacting requirements of quantitative cytochemistry have led, during the l)ast 30 years, to the development of microinstrumentation o