A Deep Feature Learning Scheme for Counting the Cells in Microscopy Data

Cell counting remains a challenging task especially because of the extreme variation of the size and shape of the microscopy data. Conventional counting methods are mostly based on the utilization of the segmentation masks as the prior information for estimating the number of the cells in the image. We propose in this work a novel cell counting scheme that uses the features provided by a deep convolutional autoencoder (DCAE) as the inputs of a shallow regressor network, instead of using the segmentation masks. First, the cellular image is given to the DCAE whose task is to reconstruct the original input image with an encodingdecoding scheme. The latent representations located in the middle of the DCAE are extracted and used as the final feature representation of the images. The second step consists of using these features as the inputs of a neural network based regressor whose outputs will represent the number of the cells in the image. The results demonstrate how the proposed scheme really manages to recognize the exact number of the cells even in the case of significant overlapping situations, exactly where most of the state-of-the-art cell counting methods fail.