Querying Representative and Informative Super-Pixels for Filament Segmentation in Bioimages

Area of Science:

• Bioimage analysis
• Computational biology
• Image processing

Background:

• Filament segmentation in bioimages is crucial for applications like neuron reconstruction and blood vessel tracing.
• Existing methods often require extensive manual annotation, leading to high costs and time investment.
• There is a need for efficient segmentation techniques that minimize the annotation burden.

Purpose of the Study:

• To develop an interactive segmentation method for bioimage-based filaments that reduces annotation effort.
• To propose a novel active learning strategy for selecting informative super-pixels for annotation.
• To improve the efficiency and cost-effectiveness of filament segmentation in biological imaging.

Main Methods:

• Utilized Simple Linear Iterative Clustering (SLIC) to segment images into super-pixels.
• Developed a batch-mode based active learning method (BMRI) for selecting representative super-pixels.
• Employed a bagging strategy and Laplacian Regularized Gaussian Mixture Models (Lap-GMM) for pixel-level segmentation.
• Implemented a majority voting strategy for classifier ensemble.

Main Results:

• The proposed interactive method significantly reduces the need for manual annotation.
• Achieved comparable segmentation performance to existing methods while saving approximately 40% of expert annotation efforts.
• Demonstrated effectiveness across three public filamentary image datasets.

Conclusions:

• The interactive active learning approach effectively alleviates the high annotation cost in filament segmentation.
• This method offers a practical solution for researchers needing accurate bioimage segmentation with reduced manual input.
• The proposed technique enhances the efficiency of bioimage analysis workflows.