Intelligent SEEQPI enables reconstruction of inhibitory neuronal nuclei and dry-mass mapping deep in the living mouse brain

Abstract

Multiphoton fluorescence microscopy enables deep brain imaging with subcellular resolution and molecular specificity, but limited photon budgets constrained by tissue heating and nonlinear damage restrict temporal resolution of structural imaging in deeper regions. Moreover, imaging multiple molecularly defined cell populations requires additional fluorescent labels, while spectral bandwidth and crosstalk constrain multiplexed detection. Here we develop an intelligent imaging framework that uses label-free Scattering-Enabled Epi-Quantitative Phase Imaging (SEEQPI) to infer molecularly specific information deep in vivo. Using paired SEEQPI and three-photon fluorescence images, we developed a Feature-wise Linear Modulation U-Net with generative adversarial learning to reconstruct inhibitory-neuron nuclei. The network accurately predicts nuclear distributions and maintains performance when training and validation volumes are separated axially by up to 6 m, enabling sparse fluorescence supervision. Cross-animal training further improves predictions in sparsely sampled target animals. Inferred nuclear maps enable dry-mass density estimation, demonstrating molecularly specific, label-free quantitative imaging deep within the living brain.