Self-Supervised Segmentation of 3D Fluorescence Microscopy Images Using CycleGAN.

Area of Science:

• Computational biology
• Bioimage analysis
• Machine learning for microscopy

Background:

• Deep learning excels at segmenting microscopy images but requires extensive manual annotation.
• Manual annotation is time-consuming and challenging, particularly for 3D datasets.
• Self-supervised learning offers a potential solution to reduce annotation burden.

Purpose of the Study:

• To develop a self-supervised segmentation method for 3D microscopy images.
• To leverage CycleGAN for image-to-image translation without paired data.
• To enable accurate segmentation of nuclei and Golgi organelles without manual masks.

Main Methods:

• Utilized CycleGAN, an image-to-image translation model, for self-supervised segmentation.
• Trained the model using automatically generated synthetic masks instead of manual annotations.
• Compared performance against supervised models like 3D U-Net and Vox2Vox.

Main Results:

• Achieved Dice coefficients of 78.07% for nuclei and 67.73% for Golgi.
• Performance was comparable to supervised methods, with minimal differences (1.4% and 0.61%).
• CycleGAN model training and testing were 5.78 times faster than 3D U-Net.

Conclusions:

• Self-supervised segmentation using CycleGAN is a viable alternative to supervised methods for 3D microscopy.
• The proposed method significantly reduces manual annotation effort while maintaining high segmentation accuracy.
• This approach facilitates quantitative analysis of organelles for disease diagnosis and biological research.