Between Stages: A Deep Learning Approach to Understanding Biological Ambiguity in Estrous Cycle Data

Abstract

In this thesis, I developed and evaluated a deep learning approach for classifying estrous stage from vaginal cytology images in a lab-specific rodent dataset. Using a VGG16-based convolutional neural network trained with class balancing, augmentation, and cross-validation, the final model achieved stable three-class performance (~65% accuracy), substantially improving over earlier baselines. However, the most important findings were not about overall accuracy alone. Model errors were concentrated in biologically adjacent stages, particularly between diestrus and proestrus, and these same regions showed the highest levels of annotator disagreement. This suggests that classification difficulty is not just technical and that it reflects underlying biological ambiguity and limitations in how continuous endocrine states are represented with discrete labels. I also tested whether adding short-term temporal context through adjacent-day image pairs could improve classification, but this approach did not resolve ambiguity and instead introduced a strong bias toward diestrus. The results show that estrous-stage classification is shaped by an interaction between model design, annotation structure, and the dynamic nature of the biology itself. Rather than building a perfect classifier, this thesis demonstrates that classifier performance can be used as a tool to identify where biological boundaries are most difficult to define and where current labeling frameworks may fall short.