EmbedSeg: Embedding-based Instance Segmentation for Biomedical Microscopy Data.

Area of Science:

• Biomedical Image Analysis
• Computational Biology
• Machine Learning

Background:

• Accurate segmentation of biological objects in 2D/3D images is crucial for biomedical research.
• Existing automated methods require further quality improvements.
• Spatial embedding-based instance segmentation shows promise but is underexplored in biomedical imaging.

Purpose of the Study:

• Introduce EmbedSeg, a novel embedding-based instance segmentation method for 2D/3D biomedical images.
• Evaluate EmbedSeg's performance against state-of-the-art methods.
• Enhance the usability and accessibility of advanced segmentation tools.

Main Methods:

• Developed EmbedSeg, an instance segmentation method utilizing spatial embeddings.
• Applied EmbedSeg to eleven benchmark datasets (4 x 2D, 7 x 3D).
• Created and released training data for four new 3D datasets.

Main Results:

• EmbedSeg matched or outperformed existing state-of-the-art methods on benchmark datasets.
• Demonstrated robust performance across diverse 2D and 3D biomedical image data.
• Validated the method's effectiveness on both established and novel datasets.

Conclusions:

• EmbedSeg provides high-quality instance segmentation for 2D/3D biomedical images.
• The method is fully open-source with accessible training tutorials and a napari plugin.
• EmbedSeg lowers the barrier for researchers needing advanced segmentation capabilities without programming expertise.