H-EMD: A Hierarchical Earth Mover's Distance Method for Instance Segmentation.

Area of Science:

• Biomedical imaging
• Computer vision
• Machine learning

Background:

• Deep learning (DL) excels at semantic segmentation in biomedical imaging, generating probability maps for instance extraction.
• Effective utilization of these probability maps for optimal instance segmentation remains a challenge.

Purpose of the Study:

• To develop a novel framework, Hierarchical Earth Mover's Distance (H-EMD), for improved instance segmentation in biomedical 2D+time videos and 3D images.
• To effectively explore DL semantic segmentation probability maps for enhanced instance segmentation.

Main Methods:

• H-EMD framework involves two stages: instance candidate generation and instance candidate selection.
• Instance candidates are generated in a forest structure from probability maps.
• Instance selection is formulated as an optimization problem using Earth Mover's Distance (EMD) and solved via integer linear programming.

Main Results:

• H-EMD consistently enhances the performance of DL semantic segmentation models.
• The framework demonstrates high competitiveness against state-of-the-art instance segmentation methods.
• Experiments were conducted on eight diverse biomedical video and 3D datasets.

Conclusions:

• H-EMD offers an effective approach to instance segmentation by optimizing the selection of candidates from DL probability maps.
• The proposed method provides a significant advancement in biomedical image analysis.
• H-EMD is a valuable tool for extracting rich instance information from complex biomedical data.