SpaCut: Transcript-aware Morphology Fusion for Robust Cell Segmentation in Subcellular Spatial Transcriptomics

Subcellular Spatial Transcriptomics (SST) offers unprecedented opportunities to decipher tissue architecture at subcellular resolution. Consequently, accurate cell segmentation is a fundamental prerequisite for analyzing SST data, as precise transcript assignment is essential for downstream single-cell interpretation. Existing methods either ignore the spatial organization of transcripts, rely on simplified geometric priors, or perform only shallow fusion between modalities. This limitation restricts their adaptability to irregular cell shapes, often leading to transcript misassignment. To address this, we propose SpaCut, a unified deep learning framework designed to synergize transcriptomic feature learning with dynamic morphological guidance.SpaCut incorporates two modules: a Gene-Wise Attention (GWA) module and a Nuclei-Guided Attentive Fusion (NGAF) module. The GWA module encodes high-dimensional gene expression into informative gene features. Coupled with FiLM, NGAF adaptively modulates nuclear soft masks using these gene features, guiding the expansion from nuclear anchors to precise cellular boundaries via a progressive training strategy.Extensive evaluations across seven datasets from diverse platforms (including Xenium, CosMx, MERSCOPE, and Stereo-seq) demonstrate that SpaCut significantly outperforms state-of-the-art methods in both segmentation accuracy and consistency with single-cell reference data.