Selection of geometrical features of nuclei оn fluorescent images of cancer cells

Abstract

The methods of geometric informative features selection of nuclei on fluorescent images of cancer cells are considered. During the survey, a review of existing geometric features was carried out, including both the signs of rotation resisted shape and displacement of the image, as well as signs of location in space. For the selection of characteristics, the methods were used: median, correlation with calculation of the Pearson correlation coefficient, correlation with calculation of the Spearman correlation coefficient, logistic regression model, random forest with CART trees and Gini criterion, random forest with CART trees and error minimization criterion. As a result of the investigation 11 characteristics were selected from 59 features, the quality of classification and time costs were calculated depending on the number of features for describing the objects. The use of 11 features is sufficient for the accuracy of classification as it allows to reduce time costs in 2,3 times.

Keywords

• correlation
• random forest
• logistic regression
• median
• classification