Graph-Theoretic Post-Processing of Segmentation With Application to Dense Biofilms.

Area of Science:

• Microscopy and Image Analysis
• Computational Biology
• Biofilm Research

Background:

• Deep learning excels at general cell segmentation but struggles with dense, small cells in biofilms, leading to segmentation errors.
• Existing post-processing methods face challenges in balancing cell counting accuracy with morphological fidelity.
• Accurate segmentation of 3D bacterial cells in biofilms is crucial for understanding microbial community dynamics.

Purpose of the Study:

• To introduce m-LCuts, a novel graph-based recursive clustering approach for post-processing 3D bacterial biofilm segmentation.
• To address over- and under-segmentation errors common in deep learning models for dense cell populations.
• To develop a method that automatically detects and segments collinearly structured cell clusters.

Main Methods:

• A graph-based recursive clustering algorithm (m-LCuts) was developed for post-processing segmentation.
• Outlier-removed graphs were constructed to identify and leverage collinearity features within the data.
• The method was applied to segment unsolved cells in 3D bacterial biofilm images.

Main Results:

• m-LCuts achieved over 90% accuracy in cell counting.
• Maintained a lower bound of 0.8 for average single-cell segmentation accuracy.
• Demonstrated effective detection of collinearly structured cell clusters.

Conclusions:

• m-LCuts significantly improves 3D bacterial biofilm segmentation accuracy, particularly for dense cell arrangements.
• The method offers a robust solution for cell counting and segmentation without requiring manual cell number specification.
• m-LCuts shows broad applicability to various datasets exhibiting data collinearity.