Quantitative profiling of whole-brain projectomes at single-axon resolution using deep learning and high-resolution light sheet microscopy

Revealing how individual axons contribute to a brain-wide projectome is indispensable for understanding brain function and behavior, yet remains technically challenging, with contemporary models failing to generalize. We introduce MAPL3, an end-to-end pipeline that integrates self-supervised learning with an innovative deep architecture to segment axons and reconstruct brain-wide projections in tera-voxel-scale light-sheet fluorescence microscopy. Its SPECTRE network integrates local axonal features with broader information shared across neighboring 3D patches. MAPL3 generalizes across out-of-distribution datasets with different tracers, microscopes, resolutions, and protocols, outperforming state-of-the-art methods. It also provides model uncertainty estimates and improves segmentation precision on unseen data. MAPL3 tractography module reconstructs long-range projections and quantifies passing vs. terminating streamlines. We showcase its ability to map the projectome of the orbitofrontal cortex, reconstructing streamlines that innervate established projection regions, which were missed by current pipelines. MAPL3 enables subject- and population-level quantitative profiling of whole-brain projectomes in tera-voxel-scale light sheet microscopy.