Segmentation Matters: Recognizing the Cell Segmentation Challenge in Spatial Transcriptomics

Area of Science:

• Neuroscience
• Genomics
• Computational Biology

Background:

• Spatial transcriptomics, particularly probe-based in situ hybridization, is a powerful tool in neuroscience.
• Accurate segmentation of individual cells (neurons and non-neuronal cells) is a critical prerequisite for downstream spatial transcriptomics analysis.
• Current automated segmentation methods face challenges in achieving high accuracy.

Purpose of the Study:

• To systematically evaluate automatic segmentation approaches for cells in human sensory ganglia using spatial transcriptomics.
• To identify the strengths, weaknesses, and characteristic error patterns of different segmentation models.
• To propose strategies for improving the accuracy and reliability of cell segmentation in spatial transcriptomics data.

Main Methods:

• Evaluation of multiple automated cell segmentation algorithms.
• Quantitative performance assessment using established metrics.
• Analysis of downstream analysis results to gauge segmentation impact.
• Systematic exploration using human sensory ganglia neuron data.
• Proposal of a manual quality control step for refining automated segmentation.

Main Results:

• Careful parameter tuning is essential for optimizing automated segmentation performance.
• Even with optimized parameters, different automated methods produce distinct types of segmentation errors.
• Automated segmentation methods exhibit unique strengths, weaknesses, and characteristic error patterns.
• A manual quality check is effective in validating and refining automated segmentation results.

Conclusions:

• No single automated segmentation method is universally superior; each has specific limitations.
• Manual review and quality control are necessary to ensure accurate cell segmentation in spatial transcriptomics.
• Future research directions include integrating multi-modal imaging data and developing tailored neural networks to enhance segmentation accuracy.