EstrousBank: rodent vaginal cytology database

EstrousBank is the first generalized database of rodent vaginal cytology images. The current 13,781 images in EstrousBank were contributed from the Goard lab, Ostroff lab, Shansky lab, Galea lab, and Sutoh lab. These labs provided cytology images from a diverse set of histological stains, magnifications, species, and strains. Initial classifications were made based on traditional cell type proportionality, as determined by the source lab. For cross-group consistency, benchmark classifications were made between the experimenters who provided the cytology images and those compiling EstrousBank. Images were classified into a given stage when 2 or more expert coders agreed on a stage classification, including those from transition stages. Five total examiners were involved in generating benchmark classifications, each with > 2 years of experience in classifying cytology images. Images containing excessive debris, n<10 cells, or <300 pixels were excluded (4.6%).

The primary purpose of EstrousBank was to create a deep learning algorithm capable of classifying estrous stage at an expert level. This deep learning network, "EstrousNet", is available at https://github.com/ucsb-goard-lab/EstrousNet. Due to the heterogeneity of the input dataset, EstrousNet classifications are not significantly different than expert human examiners in any stage surveyed, and accuracy does not differ across rodent species, stains, or subjects. Additionally, we hope that EstrousBank will be a training resource for inexperienced classifiers, with the bechmark classifications in the bank providing an intuitive guide for estrous staging.