AnyStar: Domain randomized universal star-convex 3D instance segmentation.

Area of Science:

• Biomedical image analysis
• Computer vision
• Medical imaging

Background:

• Star-convex shapes like nuclei and nodules are common in bio-microscopy and radiology.
• Current instance segmentation methods require extensive, dataset-specific annotations, hindering broad applicability.
• Adapting models to new datasets or imaging modalities necessitates significant reengineering due to variations in imaging properties.

Purpose of the Study:

• To develop a general-purpose instance segmentation network for star-convex shapes.
• To overcome the limitations of manual annotation and domain-specific model adaptation.
• To create a robust method applicable across diverse biological and medical imaging datasets.

Main Methods:

• Introduced AnyStar, a domain-randomized generative model to synthesize realistic training data.
• Simulated blob-like objects with randomized appearance, environments, and imaging physics.
• Trained a single instance segmentation network on the generated synthetic data.

Main Results:

• Networks trained with AnyStar generalize to unseen datasets without retraining or finetuning.
• Achieved accurate 3D segmentation of nuclei (C. elegans, P. dumerilii, mouse cortex, zebrafish brain) and placental cotyledons (human fetal MRI).
• Demonstrated robust performance across fluorescence microscopy, micro-CT, EM, and MRI modalities.

Conclusions:

• AnyStar's synthetic data approach enables the development of versatile instance segmentation networks.
• The method significantly reduces the need for manual annotation and domain adaptation.
• This approach holds promise for advancing automated analysis in bio-microscopy and radiology.