Nuclei Detection for 3D Microscopy With a Fully Convolutional Regression Network.

Area of Science:

• Biomedical imaging
• Computational biology
• Cell biology

Background:

• Whole-organ imaging with single-cell resolution is advancing due to 3D microscopy and tissue clearing.
• Analyzing large 3D image volumes requires efficient tools for automated cell detection, counting, and analytics.
• Existing deep learning methods struggle with overlapping, non-spherical nuclei and image blurring, leading to detection errors.

Purpose of the Study:

• To develop a novel, highly accurate deep learning approach for nuclei detection in challenging 3D whole-organ images.
• To improve automated cell analysis by addressing limitations in current nuclei detection algorithms.
• To enable precise spatial statistics and cell organization analysis from 3D imaging data.

Main Methods:

• A regression-based fully convolutional network was developed for nuclei centroid localization.
• The network was integrated with the V-net 3D semantic segmentation architecture.
• The method was validated on whole quail embryonic hearts and mouse brain stem datasets.

Main Results:

• The combined approach achieved high-accuracy nuclei detection (F1-scores of 95.3% and 92.5%) in dense, heterogeneous quail embryonic hearts.
• Excellent performance was also observed in mouse brain stem, demonstrating transferability across different nuclei shapes and intensities.
• The method processed thousands of nuclei centroids rapidly, with high accuracy in under a minute.

Conclusions:

• The developed deep learning method effectively detects nuclei in complex 3D whole-organ images, surpassing current limitations.
• This approach provides accurate cell centroids suitable for advanced spatial statistics and cell organization studies.
• The high accuracy and speed make this a valuable tool for analyzing large-scale biological imaging data.