1. Introduction
In digital pathology, slides of stained histopathological tissue are scanned using whole-slide imaging systems, resulting in a digital multi-scale representation of the slide. Computer-aided analysis of these whole-slide images (WSI) allows to calculate objective measures facilitating diagnosis and medical research. Recent work on immunotherapies evaluates the influence of certain characteristics in the tumor microenvironment (TME) as an indicator for tumor progression and therapeutic success [1]–[3]. A first step towards an automated analysis of the TME is the exhaustive classification of the composing tissue. An initial approach of tissue classification [4] analyzes the tissue composition by solving a segmentation problem with three classes using morphology and texture features. However, information contained in the stain is not considered. More recent proposals to classify histopathological images can be found in [5], [6], where Bag-of-(Visual-)Features is applied for categorization. These algorithms extract visual patterns at key points of an image and bin them into a histogram representation. An SVM is used to map the histograms to their respective class. Furthermore, [7] introduces a multi-class solution for automated Gleason-grading of prostate cancer. A cascaded classification pipeline, based on graph-features of the nuclei positions, labels six classes which denote the different tumor grades and three additional tissue types. Finally, a review on texture-based system is given in [8] showing cases of successful applications of color and texture representations in Gleason-grading of prostate cancer. The paper covers various subcategories of features and reviews methods very similar to this application. Image categorization and prostate cancer grading are use-cases for a region-of-interest classification approach and work with relatively large selected image patches. However, the required region size for a comprehensive description in their respective representation imposes limitations on the use of these approaches in a spatially dense classification scenario.