Classification of cultured mammalian cells by shape analysis and pattern recognition.

Abstract

We have developed a method for classifying cultured cells on the basis of shape characteristics. High-resolution optical information on thre imensional shape was obtained by anodic oxide interferometry. Each interference order formed in a cell was considered as a closed figure; measurement of 37 mathematical descriptors was carried out for each figure. The individual cells were classified according to the values of their descriptors. We used standard principles of pattern recognition, such as hierarchical cluster analysis and nearest neighbor analysis, as a basis for ordering the cells into groups. Alternatively, linear discriminant functions could be used, but they provided only a slight improvement in correct classification of the cells. We anticipate that the method will be appropriate for classification of cultured cell lines and for determination of the magnitude and direction of cell shape changes implicated in various biological processes. Cell shape is one of the fundamental criteria used for visual classification of cells into histological types. The relationship of cell shape to functional integrity is obvious in some specialized cells (e.g., neurons). Recently, aspects of shape maintenance have also been implicated in mechanisms of growth control in normal cells (1). Other observations have suggested that the shape of cells cultured on solid substrata (2-9), particularly the area of lamellar cytoplasm (2, 3, 7), changes during viral and oncogenic transformation. Research on these problems would be facilitated by methods for quantifying cell shape. Several descriptors of shape have been developed by others for automated cytological diagnosis (10-13). In the present study, we have used many of these descriptors and defined others as the basis of a classification strategy for cultured cells.Besides its application to cultured cells, the present technical approach to the problem of shape analysis differs from that of previous workers in two respects. First, because the threedimensional aspect of cell shape is difficult to evaluate by routine photo-optical methods, we used anodic oxide interference to obtain information on cell height. Because of the multiple reflection characteristics of the interference pattern, this technique also offered better resolving power in the vertical dimension than did traditional light microscopy. The principles of anodic oxide interferometry have been discussed (14, 15) in relation to protein adsorption on model surfaces, including culture substrata. A second area of innovation was in avoiding digitization errors by magnifying the image prior to digitization.We have also employed a large set of shape features for analyzing cells. Frequently, features used for discriminating among biological objects are chosen on the basis of visual discrimination carried out by humans (11-13). We recognized that shape features that might relate to certain aspects of cellular functions would not necessarily be easy to perceive visually (16). Therefore, the approach adopted was to meas...

Quantification of the degree of cell spreading of human fibroblasts by semi‐automated analysis of the cell perimeter

“…In general semiquantitative scores obtained by visual inspection of the cells are used. Recently, the area occupied by cells on a substratum has been measured, often by computer assisted planimetry (8,10,12), while a number of mathematical formulations have been developed to quantify cell shape, orientation or polarization (7,11 14) for automated karyotyping was adapted for our purposes. This system has a high accuracy mainly due to the use of scanned photonegatives and the calculation of a local background for each cell.…”

Effect of Cdc42 domains on filopodia sensing, cell orientation, and haptotaxis

“…Previously, the laboratory developed methods for the unbiased classification of cell features, based on the use of size-invariant measures of shape [8]. Using a standard statistical procedure, we could extract latent factors that correspond to cell features.…”

“…The cells' edges were traced onto a transparency and scanned into an SGI workstation running under IRIX. This workstation hosted software for the derivation of values for geometric variables and model comparisons, as previously described [8]. At least 30 cells from each zone were analyzed, and the set of variable values was reduced to a smaller number of variables by factor analysis [3].…”

Cell morphology classification and clutter mitigation in phase-contrast microscopy images using machine learning

“…In early work on classification of cell shape, Olson et al [20] showed how machine learning techniques known at that time, i.e., hierarchical clustering and nearest neighbor analysis, could be used to classify cells into three classes by interpreting various shape descriptors. More recent classification techniques, in particular, support vector machines, have been applied by Han et al [8] to the problem of detecting cell nuclei and by Ruusuvuori et al [23] for classifying rod, spherical and spiral bacteria.…”