Evaluation of cell segmentation methods without reference segmentations.

Area of Science:

• Bioimage Informatics
• Computational Biology
• Microscopy Image Analysis

Background:

• Accurate cell segmentation is vital for reliable bioimage analysis, yet current evaluation methods rely on human comparisons, which are costly and potentially biased.
• Developing robust cell segmentation techniques requires objective and reproducible evaluation metrics.

Purpose of the Study:

• To establish an objective framework for evaluating cell segmentation methods using quantitative metrics.
• To compare the performance of various segmentation algorithms across diverse biological samples and imaging modalities.
• To identify superior cell segmentation strategies and assess the impact of postprocessing on performance.

Main Methods:

• Definition of novel segmentation quality metrics applicable to multichannel fluorescence images.
• Application of metrics to 14 segmentation methods across datasets from four multiplexed microscopy modalities and five tissues.
• Utilizing principal component analysis (PCA) to aggregate metrics into an overall quality score for method ranking.

Main Results:

• Two deep learning-based methods demonstrated superior overall performance in cell segmentation.
• Significant improvements in segmentation accuracy were observed across all tested methods after implementing postprocessing for cell and nuclear mask matching.
• The developed evaluation tool and associated data are publicly available as open-source resources.

Conclusions:

• An objective, human-free evaluation approach for cell segmentation methods has been successfully developed and validated.
• Deep learning approaches show promise for high-performance cell segmentation, with postprocessing enhancing overall accuracy.
• The open-source evaluation tool facilitates reproducible research and aids in selecting optimal segmentation methods for bioimage analysis.