Cell Segmentation as Strategic Decision Making

Area of Science:

• Spatial transcriptomics
• Computational biology
• Machine learning in genomics

Background:

• Current cell segmentation in spatial transcriptomics (ST) relies on staining or transcript clustering, often limited by manual annotations and staining quality.
• Existing methods struggle with accurate alignment to true transcriptomic profiles, impacting downstream analysis.

Purpose of the Study:

• To develop a novel, staining-free cell segmentation framework for ST that directly utilizes transcriptomic data.
• To improve the accuracy, robustness, and transcriptome-alignment of cell segmentation in ST.
• To enable whole-transcriptome imputation from sparse ST data for comprehensive spatial biology insights.

Main Methods:

• Introduced RedeFISH, a reinforcement learning-based framework for cell segmentation as a sequential decision-making process.
• Iteratively optimized transcript assignment policy to align segmented cell expression with single-cell expression profiles.
• Developed a method for whole-transcriptome imputation using spatially guided transfer from scRNA-seq data.

Main Results:

• RedeFISH achieved improved cosine similarity (9.7% avg.) and reduced RMSE (13.9%) in segmented cell expression profiles compared to state-of-the-art methods.
• Demonstrated increased agreement (19%) between segmented and ground-truth cell regions, indicating enhanced accuracy and robustness.
• Successfully imputed whole-transcriptome profiles, enabling in situ prioritization of regulators in spatial niches and analysis of developmental processes in cancer.

Conclusions:

• RedeFISH offers a robust, staining-free, and transcriptome-informed approach for cell segmentation in ST, overcoming key limitations of existing methods.
• The framework enhances the accuracy of cell delineation and enables comprehensive molecular profiling of spatial niches and developmental trajectories.
• This work paves the way for unbiased in situ analysis of complex biological processes, including cancer development, using spatial transcriptomics.