Improved accuracy for estrous cycle staging using supervised object detection

Abstract

The estrous cycle regulates reproductive events and hormone changes in female mammals and is analogous to the menstrual cycle in humans. Monitoring this cycle is necessary as it serves as a biomarker for overall health and is crucial for interpreting study results. The estrous cycle comprises four stages influenced by fluctuating levels of hormones, mainly estradiol and progesterone. Tracking the cycle traditionally relies on vaginal cytology, which categorizes stages based on three epithelial cell concentrations. However, this method has limitations, including time-consuming training and variable accuracy among researchers. To address these challenges, this study assessed the feasibility and reliability of two machine learning methods. An object detection-based machine learning model, Object Detection Estrous Staging (ODES), was employed to identify cell types throughout the estrous cycle in mice. A dataset of 555 vaginal cytology images with four different stains was annotated, with 335 images for training, 45 for validation, and 175 for testing. A novel, accurate set of rules for classification was derived by analyzing training images. ODES achieved an average accuracy of 87% in classifying cycle stages and took only 3.9 minutes to analyze 175 test images. The use of object detection machine learning significantly improved accuracy and efficiency compared to previously derived supervised image classification models (33-45% accuracy) and human accuracy (66% accuracy), refining research practices for female studies. These findings facilitate the integration of the estrous cycle into research, enhancing the quality of scientific results by allowing for efficient and accurate identification of the cycle stage.