Automated Quantification of Human Osteoclasts Using Object Detection.

Area of Science:

• Cell Biology
• Biotechnology
• Artificial Intelligence in Medicine

Background:

• Skeletal remodeling, regulated by osteoclasts, is vital for bone health.
• Osteoclasts, derived from mononuclear cells, are key bone-resorbing cells.
• Manual quantification of osteoclasts in vitro is labor-intensive and prone to bias.

Purpose of the Study:

• To develop an efficient and reliable machine learning method for quantifying human osteoclasts.
• To automate osteoclast identification and counting from microscopic images.

Main Methods:

• Utilized a deep learning object detection framework (Darknet, YOLOv4) trained on 307 well images.
• Trained models on 94,974 marked osteoclasts from seven human peripheral blood mononuclear cell (PBMC) donors.
• Evaluated model performance using mean average precision and correlation with human annotators.

Main Results:

• The trained model achieved 85.26% mean average precision.
• Demonstrated a high correlation (0.99) with human annotators.
• The AI model counted 2.1% more osteoclasts on average and showed higher inter-annotator agreement than humans.

Conclusions:

• The deep learning method provides a reliable, less biased, and time-saving approach for osteoclast quantification.
• This AI-driven tool can significantly enhance research efficiency in skeletal biology and disease.
• Encourages further validation and testing of the models with diverse datasets.