NOISe: Nuclei-Aware Osteoclast Instance Segmentation for Mouse-to-Human Domain Transfer.

Area of Science:

• Biomedical Image Analysis
• Computational Biology
• Osteoporosis Research

Background:

• Osteoclast cell image analysis is crucial for osteoporosis research but relies on laborious manual annotation.
• Existing automated methods lack full instance segmentation capabilities and do not provide accessible code or datasets.

Purpose of the Study:

• To develop a fully automated deep learning method for osteoclast instance segmentation.
• To create and release a large-scale annotated dataset for osteoclast images.
• To enable reproducible and accelerated research in osteoporosis.

Main Methods:

• Developed a deep learning instance segmentation model applicable to both mouse and human osteoclast cells.
• Created a novel nuclei-aware osteoclast instance segmentation (NOISe) training strategy.
• Compiled a dataset of approximately 2 × 10^5 expert-annotated mouse osteoclast masks.

Main Results:

• Achieved 0.82 mAP$_{0.5}$ cross-validation performance for mouse osteoclasts.
• Improved human osteoclast segmentation performance from 0.60 to 0.82 mAP$_{0.5}$ using the NOISe strategy.
• Successfully automated the full osteoclast instance segmentation task.

Conclusions:

• The presented deep learning method and dataset significantly advance automated osteoclast image analysis.
• The NOISe strategy enhances model generalizability and accuracy for osteoclast segmentation.
• Publicly releasing the code, models, and dataset promotes reproducibility and accelerates osteoporosis research.