Multi-class single-label classification of histopathological whole-slide images

Digitizing histopathological slides into whole-slide images enables the use of image analysis techniques for a comprehensive study of diseases at tissue level. In this work, we investigate possible configurations of classifiers and features to classify entire tissue slides. Feature candidates are first evaluated individually and then combined to form a strong classifier. For evaluation of this nine-class problem we use one sparsely and one densely extracted data set to obtain a conservative and an optimistic estimate of the performance. To this end, the best results in terms of overall accuracy (86.6%) and F1-score (67.8%) are achieved by a Random Forest classifier with a Color-Histogram feature.