• Accuracy;
  • cytology;
  • image analysis;
  • intraobserver reproducibility;
  • interobserver reproducibility;
  • Monte Carlo methods;
  • morphometry;
  • segmentation;
  • variance component models.

The segmentation of nuclear images is a crucial step in the development of procedures using image analysis for the cytological diagnosis of cancer. The purpose of this study is to evaluate the reproducibility and accuracy of several interactive segmentation methods which can be used in this context. Four methods were studied: a thresholding-based method enabling selection of intensity histogram contrast and brightness, manual tracing with a stylus, and arc- and ellipse-fitting routines. Features of nuclear size and shape were derived from nuclei segmented on repeated occasions by several individuals. Variance component models provided a statistical framework for evaluating the intraobserver and interobserver variability of these measurements in terms of their intraclass correlation coefficients. Of the methods tested, the arc-fitting segmentation method gave the most reproducible results, and thresholding the least. Reproducibility was generally very high both between individuals and for repeated segmentations by a single individual. Accuracies of area measurements for the various methods, as determined with respect to point counting, paralleled the reproducibilities of the methods. Sample size requirements were observed to be more dependent on the biological variability of the tissue sampled than on the particular segmentation method or on the number of individuals performing segmentation.