Non-invasive estrous cycle classification in mice using convolutional neural networks

Abstract

Accurately identifying the estrous cycle phases in laboratory mice is essential for neurological research, reproductive studies, and breeding programs. Here, we introduce a non-invasive approach using Convolutional Neural Networks (CNNs) to classify estrous stages from external genital images, with cytological smears serving as the reference standard for labeling. Images were curated, preprocessed, and augmented to ensure consistency and generalization. Seven pre-trained CNNs and a lightweight custom architecture, Repro Cycle Net (RCN), were evaluated. All models achieved accuracies above 75%, with RCN exceeding 83% and showing the lowest test loss, highlighting its efficiency despite a simple four-layer design. Saliency map analyses revealed that classification relied on perivaginal features, while irrelevant regions such as the fur area were largely ignored. Importantly, binary classification of estrus versus non-estrus directly informs mating feasibility, underscoring the immediate utility of this method for reproductive studies and colony management. This work demonstrates that combining deep learning with external genital observation enables efficient and reproducible estrous monitoring, supporting both experimental reliability and animal welfare.