Evaluation of Spatial Relations in the Segmentation of Histopathological Images

Abstract

In this work, improvement of final segmentation results is aimed by evaluating spatial relations in the segmentation of histopathological images. In the first step features are extracted using Haralick texture descriptor in the La*b* color space for pre-segmentation of histopathological images. Some training sets with different number of samples are obtained by cellular and extra-cellular structures in images and classifier models are formed by these training sets using support vector machine (SVM) and random forest methods. To improve the accuracies of pre-segmentation results obtained by supervised learning methods, spatial information must also be considered. In this purpose, hidden Markov random fields methods is used to ensure the regularization of pre-segmentation results.