Automatic ovarian follicle detection using object detection models.

Area of Science:

• Reproductive biology
• Computational pathology
• Biomedical imaging

Background:

• Ovarian folliculogenesis is crucial for female reproduction, involving complex follicular development.
• Accurate quantification of later-stage ovarian structures (antral follicles, corpora lutea) is vital for research and drug development.
• Manual counting of these structures in histology is time-consuming and prone to errors.

Purpose of the Study:

• To evaluate the efficacy of deep learning models for automated counting of antral follicles and corpora lutea.
• To compare the performance of different deep learning architectures (YOLO, RetinaNet) for this task.
• To demonstrate the potential of AI in improving accuracy and efficiency in reproductive biology research.

Main Methods:

• Development of two one-stage object detection models: YOLO and RetinaNet, utilizing various backbone architectures.
• Implementation of transfer learning, early stopping, and data augmentation to enhance model generalizability.
• Application of sampling strategies and focal loss to address class imbalance issues in the dataset.

Main Results:

• RetinaNet achieved a mean average precision of 83%, while YOLO achieved 75% on the testing dataset.
• The models were trained and validated on a dataset of 1000 images.
• Deep learning models demonstrated significant improvements in speed and accuracy compared to manual counting.

Conclusions:

• Deep learning offers a powerful tool for automating the quantification of ovarian follicular structures.
• These AI-driven methods can accelerate research in reproductive biology and drug discovery.
• The developed models provide a more accurate and efficient alternative to traditional manual counting techniques.