Abstract

Mapping brain cell types with high spatial precision is fundamental for understanding brain structure and function. Traditional cell atlases, such as the Allen Brain Cell Atlas, rely on labor-intensive techniques like single-cell sequencing and spatial transcriptomics, often at limited spatial resolution. In this study, we present a deep learning framework—Multi-modal Fusion Network (MFNet)—that predicts neuronal cell types across the whole mouse brain by integrating high-resolution diffusion MRI (dMRI) and multi-modal datasets from the BICCN mouse brain atlas. Using the Allen Brain Cell Atlas as ground truth, we trained MFNet to predict dominant neuronal cell neighborhoods and classes, taking advantage of regional specificity and hierarchical cell-type structures. High-resolution (45 µm isotropic) ex vivo dMRI was acquired and processed with multiple diffusion models (DTI, DKI, NODDI, SANDI), then registered to the Allen Mouse Brain Common Coordinate Framework (CCFv3), along with spatial transcriptomics and epigenomics data. MFNet produces a full-scale cell type atlas at 10 µm isotropic resolution, significantly enhancing the spatial detail of the original atlas. Additionally, we examined correlations between dMRI features and gene expression of marker genes. This study offers a novel, efficient method for generating high-resolution brain cell atlases and demonstrates the potential of advanced imaging combined with deep learning to decode brain cellular architecture and its relevance to development and disease.