Deep Learning Based Cell Segmentation Using Cascaded U-Net Models

In this study, a deep learning-based cell segmentation method is proposed. Since existing methods have difficulty in separating touching cells, problem is solved in two parts. First, the position of each cell on the image is determined using a convolutional neural network in U-Net architecture. Then, a second U-Net network is trained to determine the cell boundaries by using first network output as prior information. Finally, complete segmentation is achieved by applying the Watershed algorithm to the output of the second network. The proposed method achived 0.958 detection and 0.85 segmentation accuracies in the tests performed in the DIC-C2DH-HeLa data set. These scores are very close to the state-of-art methods in this data set.