Physics-informed multi-encoder adaptive optics enables rapid aberration correction for intravital microscopy of deep complex tissue

Tissue-induced optical aberrations fundamentally constrain intravital microscopy in deep or complex biological specimens. While adaptive optics (AO) can compensate for these aberrations, conventional AO methods are limited by either guide-star dependency or slow correction speeds. Here, we develop MeNet-AO, a multi-encoder network-based AO method that enables rapid, guide-star-free aberration correction. By integrating a noise-resilient, structure-independent feature extraction model with a physics-informed multi-encoder architecture, MeNet-AO jointly decodes multiple large-amplitude aberration modes from wavefront-modulated image pairs, achieving an effective balance between prediction accuracy and temporal efficiency. Validated in living organisms, MeNet-AO improves fluorescence imaging in zebrafish brain and eye, enhances neuronal calcium transients and direction selectivity in mouse visual cortex, and enables subcellular-resolution microglial calcium imaging through thinned-skull windows – revealing spatiotemporally heterogeneous signaling patterns previously obscured by skull aberration. The speed and robustness of MeNet-AO in low-signal and scattering conditions establish it as a versatile platform for dynamic subcellular imaging deep within native tissue environments. This study presents MeNet-AO, a deep learning method that enables rapid, guide-star-free aberration correction in living tissue, revealing high-fidelity neuronal and microglial dynamics under low-signal, highly scattering conditions.