Automated classification of cells in electron microscopic images of lymphoreticular tissue.

In classifying cells in tissue sections, one must consider the fact that only random projections of cells and of subcellular structures are available in the two-dimensional image. Therefore, measurement values that solely reflect the size of such projections cannot be taken on their own as a basis for cell classification. More complex morphologic features such as shape, texture and distribution pattern of cells and their components should be analyzed. Using cell nuclei as an example, the relationship between such features and geometric measurement values is evaluated. It can be shown that a well balanced combination of geometric parameters provides a suitable basis for reproducing the visual preclassification of lymphocytes in tissue sections. Moreover, using a cluster algorithm, which allows different levels of similarity to be defined, a hierarchical sequence of subclusters turns out, indicating the heterogeneity of the visually determined cell classes. Whether or not these subclusters can be correlated to functionally defined subpopulations of lymphocytes remains a matter for further investigation.