A Method for Automatic Tracking of Cell Nuclei With Weakly-Supervised Mitosis Detection in 2D Microscopy Image Sequences

Due to a high interest in microscopic cell migration analysis for biological research, numerous cell segmentation and tracking algorithms has emerged. The main tasks of cell tracking methods are to segment each cell and establish individual cell lineages over time, accounting for possible cell disappearance and division events. In some datasets, cells can drastically change their appearance during the mitosis stage, thus making division detection a challenging problem. The most prominent methods exploit different neural network architectures for at least one of these tasks. We propose a method that uses a single UNet topology network to solve both tasks of cell nuclei instance segmentation and detection of mitotic events. For instance segmentation, the network is learned to segment primary masks and object centroids that are used by watershed transform to obtain individual nuclei regions. For mitotic events detection, we first manually mark cells entering and finishing mitosis. Then, previously trained network is used to generate weak nuclei segmentation labels for all data images in sequences with marked mitotic events. We add an additional output to the trained network for segmentation of mitotic events. The training is resumed for both tasks on initial ground truth segmentation, generated weak labels, and crude mitotic events markers. For tracking, we use generalized nearest neighbour method that can greedily search the best 1-to-1 and 1-to-2 instance connections over multiple frames. Segmentation of the mitotic events produced by the trained model is incorporated into the tracking algorithm to improve cell division detection. We evaluate the results of the proposed method and compare it with the previously developed algorithm, achieving better performance on our dataset. We assume it is possible to upgrade other existing segmentation frameworks to also learn the task of segmenting mitotic events and enhance division detection using the proposed pipeline.