Performance comparison of machine learning methods for prognosis of hormone receptor status in breast cancer tissue samples

Abstract

We examined the classification and prognostic scoring performances of several computer methods on different feature sets to obtain objective and reproducible analysis of estrogen receptor status in breast cancer tissue samples. Radial basis function network, k-nearest neighborhood search, support vector machines, naive bayes, functional trees, and k-means clustering algorithm were applied to the test datasets. Several features were employed and the classification accuracies of each method for these features were examined. The assessment results of the methods on test images were also experimentally compared with those of two experts. According to the results of our experimental work, a combination of functional trees and the naive bayes classifier gave the best prognostic scores indicating very good kappa agreement values (κ=0.899 and κ=0.949, p<0.001) with the experts. This combination also gave the best dichotomization rate (96.3%) for assessment of estrogen receptor status. Wavelet color features provided better classification accuracy than Laws texture energy and co-occurrence matrix features.