3D Domain Adaptive Instance Segmentation via Cyclic Segmentation GANs.

Area of Science:

• Neuroscience
• Computer Vision
• Biomedical Imaging

Background:

• 3D instance segmentation is crucial for unlabeled imaging modalities.
• Expert annotation is costly and time-consuming.
• Existing methods use separate image translation and segmentation networks.

Purpose of the Study:

• To develop a unified network for simultaneous image translation and 3D instance segmentation.
• To improve segmentation performance on unlabeled target domains.
• To reduce computational cost compared to sequential methods.

Main Methods:

• Proposed Cyclic Segmentation Generative Adversarial Network (CySGAN) with weight sharing.
• Integrated CycleGAN losses, supervised losses, self-supervised, and adversarial objectives.
• Leveraged unlabeled target domain data for enhanced performance.

Main Results:

• CySGAN achieved superior performance in 3D neuronal nuclei segmentation.
• Outperformed pre-trained models and sequential translation-segmentation approaches.
• Demonstrated effectiveness on electron microscopy (EM) and expansion microscopy (ExM) data.

Conclusions:

• CySGAN offers an efficient and effective solution for 3D instance segmentation across unlabeled modalities.
• The unified network design reduces computational overhead.
• Publicly released dataset (NucExM) and implementation facilitate further research.