Image-guided spatial omics enhancement reveals hidden spatial microstructures

The rapid advancement of spatial omics is fundamentally hindered by the resolution gap between physical capture platforms and genuine biological microstructures, a challenge compounded by inherent data sparsity and noise. While current image-guided computational methods attempt to bridge this gap, they often lack the multi-modal flexibility, non-linear modeling, and scalability required for modern, whole-tissue datasets.

To address this, we introduce Bell, a modality-agnostic deep learning framework that reconstructs high-fidelity spatial microstructures by dynamically fusing histological images, spatial coordinates, and low-resolution molecular measurements via an adaptive attention mechanism. The study also presents mmBell, an extension utilizing a unified encoder structure to achieve cross-modal integration for increasingly complex multi-omics data. Systematically validated across over 10 spatial platforms and 20 datasets, Bell and mmBell consistently outperform state-of-the-art methods in resolution enhancement and noise suppression. Ultimately, this framework provides a highly robust, scalable solution for deeply deciphering complex spatial tissue organization.

Bell is available from the GitHub repository.