Deep learning–based identification and counting of early-stage follicles in rat ovaries

Counting ovarian early follicles helps assess the impact of pharmaceuticals and endocrine-disrupting chemicals on female fertility. Traditional manual counting methods are labor-intensive, time-consuming, and prone to intra- and interobserver variability, leading to inconsistent results. This study addresses these limitations by developing a fully automated solution for the identification and quantification of early-stage follicles in hematoxylin and eosin-stained rat ovarian sections, with the aim of enhancing efficiency, accuracy, and reproducibility. We developed a deep learning model designed to identify primordial type 1 and type 2 follicles, and a second model focusing on detecting early growing follicles. The automated system exhibited high performance on an independent test data set of 200 whole-slide images, whose predictions were validated by pathologists. The models achieved high precision and recall, with 98.49% and 90.92%, respectively, for type 1; 95.97% and 97.54% for type 2; and 99.45% and 80.09% for early growing follicles. Further comparisons with manual annotations performed by 4 independent pathologists revealed significant interobserver variability. While human performance was hindered by factors such as fatigue and inconsistent criteria application, resulting in variable precision and recall and high rates of missed follicles, the algorithm maintained superior consistency and overall higher F1-scores. Furthermore, a comparison between manual work and the algorithm's productivity showed that the algorithm was 11 times faster than manual quantification. These results demonstrate that the automated approach not only matches expert accuracy but also effectively eliminates subjectivity and diagnostic drift inherent in manual counting, providing a reproducible method for follicle counting, and significantly enhancing workflow efficiency.