Virtual Cell Type Atlas of Mouse Brain from MRI Signatures using Attention Res-UNet

Motivation: Magnetic resonance imaging (MRI) has shown promise in illustrating tissues' microstructure. However, its potential on predicting cell distribution remains unexplored. Goal(s): Our objective is to predict three-dimensional representations of different cell types across the entire mouse brain via deep learning on multi-contrast MRI. Approach: We trained an Attention Res-UNet model using a template-based MRI dataset and an atlas-based cell distribution dataset. Results: Our findings demonstrated that multi-contrast MRI with Attention Res-UNet model could predict composition of various cell types across whole mouse brain. Our predictions aligned well with typical cell distribution patterns and regional characteristic. Impact: We demonstrated MRI-based deep learning could predict three-dimensional representations of different cell types at whole-brain level, which were consistent with typical cell distribution patterns and regional characters. Our findings highlight the potential of MRI for predicting three-dimensional cell atlas.