Automated classification of estrous stage in rodents using deep learning

Abstract

The rodent estrous cycle modulates a range of biological functions, from gene expression to behavior. The cycle is typically divided into four stages, each characterized by distinct hormone concentration profiles. Given the difficulty of repeatedly sampling plasma steroid hormones from rodents, the primary method for classifying estrous stage is by identifying vaginal epithelial cell types. However, manual classification of epithelial cell samples is time-intensive and variable, even amongst expert investigators. Here, we use a deep learning approach to achieve classification accuracy at expert level. Due to the heterogeneity and breadth of our input dataset, our deep learning approach (“EstrousNet”) is highly generalizable across rodent species, stains, and subjects. The EstrousNet algorithm exploits the temporal dimension of the hormonal cycle by fitting classifications to an archetypal cycle, highlighting possible misclassifications and flagging anestrus phases (e.g., pseudopregnancy). EstrousNet allows for rapid estrous cycle staging, improving the ability of investigators to consider endocrine state in their rodent studies.

Improved accuracy for estrous cycle staging using supervised object detection

“…The higher performance of ODES in classifying estrous stages may be attributed to its novel approach that integrates object detection with subsequent classification based on detected cell types. This methodology provides a contrast to traditional image classification models that rely only on supervised learning paradigms [10,11]. Previous models typically label entire images according to the estrous stage, without the ability to identify individual cell types [10,11].…”

“…This methodology provides a contrast to traditional image classification models that rely only on supervised learning paradigms [10,11]. Previous models typically label entire images according to the estrous stage, without the ability to identify individual cell types [10,11]. This may have led to pattern detections within the image that do not capture the features essential for accurate stage classification.…”

“…The SECREIT model [10] was developed to identify images representing three stages of the estrous cycle—diestrus, estrus, and proestrus—using solely Giemsa stain, achieving a reported accuracy of 93%. The EstrousNet Model [11] was trained on thousands of images from various labs, encompassing different animal species (mice and rats), stains, and magnifications, with a reported accuracy of 83%. However, it is yet unknown whether the accuracies of these models are reproducible on new data sets.…”

The Track‐by‐Day Method for Monitoring the Rodent Estrous Cycle

“…Stages other than estrus are traditionally defined by subtle, nuanced features of cytology that require a trained eye regardless of whether qualitative (Cora et al 2015;OECD, 2008;Sugiyama et al, 2021) or semi-quantitative (Hubscher et al, 2005;Paccola et al, 2013) criteria are used. Recently, machine learning tools have been developed to assist in staging cytology smears and, interestingly, these tools most readily detect the estrus stage (Sano et al, 2020;Wolcott et al, 2022). Estrus is not only easily identifiable in smears, but is also the only stage known to follow immediately after a change in serum hormones (Barker & Walker, 1966;Buchanan et al, 1998).…”

“…Recently, machine learning tools have been developed to assist in staging cytology smears and, interestingly, these tools most readily detect the estrus stage (Sano et al., 2020; Wolcott et al., 2022). Estrus is not only easily identifiable in smears, but is also the only stage known to follow immediately after a change in serum hormones (Barker & Walker, 1966; Buchanan et al., 1998).…”