Item

Abstract

Mitosis detection poses a major challenge in cell tracking as mitoses are crucial events in the construction of genealogical trees. Making use of typical mitotic patterns that can be seen in phase contrast images of time lapse experiments, we propose a new benchmark data set CeTReS.B-MI consisting of mitotic and non-mitotic cells from the publicly accessible, fully labeled data set CeTReS.B. Using this data, two simple mitosis detectors (based on compactness and intensity) are used exemplarily to train, test and compare their ability to detect mitotic events. As a gold standard, we propose a linear support vector machine (SVM), which is able to separate the classes with a high accuracy (AUC=0.993). To illustrate the potential impact of a robust mitosis detection, the proposed classifiers are combined with two state of the art cell tracking algorithms. For both algorithms, performance does change when adding mitosis detection. Finally, this evaluation also emphasizes how easy implementation and comparison becomes, having suitable benchmark data at hand.