NISNet3D: three-dimensional nuclear synthesis and instance segmentation for fluorescence microscopy images

Area of Science:

• Computational Biology
• Biomedical Imaging
• Machine Learning

Background:

• Accurate cell segmentation is crucial for tissue cytometry.
• 3D nuclei segmentation is a significant challenge, hindering advanced tissue analysis.
• Deep learning methods require substantial annotated data, limiting their application.

Purpose of the Study:

• To introduce NISNet3D, a novel deep learning network for 3D nuclei instance segmentation.
• To address the bottleneck of 3D nuclei segmentation in tissue cytometry.
• To enable accurate organ-level characterization using 3D imaging techniques.

Main Methods:

• Developed NISNet3D, integrating a modified 3D U-Net, 3D marker-controlled watershed transform, and instance segmentation.
• Trained the network on large-scale synthetic nuclei data, reducing reliance on manual annotations.
• Compared NISNet3D performance against existing 3D nuclei segmentation techniques.

Main Results:

• NISNet3D achieved accurate segmentation of challenging 3D image volumes.
• The network demonstrated robust performance when trained on synthetic data, even without annotated volumes.
• Quantitative comparisons showed superior or comparable results to existing methods.

Conclusions:

• NISNet3D effectively overcomes limitations in 3D nuclei segmentation for tissue cytometry.
• The use of synthetic data significantly reduces the annotation burden for deep learning models.
• This method advances the potential of high-throughput tissue and organ analysis.