Automatic ovarian follicle detection using object detection models

Ovaries are of paramount importance in reproduction as they produce female gametes through a complex developmental process known as folliculogenesis. In the prospect of better understanding the mechanisms of folliculogenesis and of developing novel pharmacological approaches to control it, it is important to accurately and quantitatively assess the later stages of ovarian folliculogenesis (i.e. the formation of antral follicles and corpus lutea). Manual counting from histological sections is commonly employed to determine the number of these follicular structures, however it is a laborious and error prone task. In this work, we show the benefits of deep learning models for counting antral follicles and corpus lutea in ovarian histology sections. Here, we use various backbone architectures to build two one-stage object detection models, i.e. YOLO and RetinaNet. We employ transfer learning, early stopping, and data augmentation approaches to improve the generalizability of the object detectors. Furthermore, we use sampling strategy to mitigate the foreground-foreground class imbalance and focal loss to reduce the imbalance between the foreground-background classes. Our models were trained and validated using a dataset containing only 1000 images. With RetinaNet, we achieved a mean average precision of 83% whereas with YOLO of 75% on the testing dataset. Our results demonstrate that deep learning methods are useful to speed up the follicle counting process and improve accuracy by correcting manual counting errors.