Cell Segmentation as Strategic Decision Making

Cell segmentation in imaging-based spatial transcriptomics (ST) relies on stained image segmentation or transcript clustering, but current methods are limited by dependence on manual annotations, sensitivity to staining quality and cell density, and lack of alignment with true transcriptomic profiles. We introduce RedeFISH, a reinforcement learning-based framework that formulates cell segmentation as a sequential strategic decision-making process. Through iterative optimization of the transcript assignment policy, RedeFISH aligns segmented cell expression profiles with single-cell expressions to achieve optimal segmentation. This approach enables staining-free, transcriptome-informed delineation of individual cells, thereby overcoming key limitations of existing methods. Benchmarking across multiple ST platforms shows that RedeFISH improves cosine similarity of segmented cell expression profiles by 9.7% on average and reduces root mean squared error (RMSE) by 13.9% compared with state-of-the-art methods. It also increases agreement between segmented and ground-truth cell regions by 19%, demonstrating improved accuracy and robustness. RedeFISH further imputes whole-transcriptome profiles from sparse ST measurements via spatially guided transfer from scRNA-seq data. The whole-transcriptome coverage enables unbiased and in situ prioritization of the critical regulators of spatial niches, e.g., mouse intestinal stem cell niches, and unlocks the possibility to delineate, with only one sample, the panoramic molecular and cellular changes along the whole developmental process from initiation to invasion for human cancer.