Breaking the speed limit in two-photon microscopy via deep-learning imaging reconstruction from anisotropic sparse sampling

Two-photon microscopy is essential for high-resolution imaging of neural circuits, yet its serial point-scanning architecture imposes a fundamental trade-off among imaging speed, spatial resolution, and field of view. This spatiotemporal constraint limits the ability to capture fast neuronal dynamics. Here, we introduce Imaging Reconstruction from anIsotropic Sparse sampling (IRIS), a deep-learning framework that overcomes this bottleneck by strategically downsampling along the slow scanning axis and reconstructing high-fidelity images using a one-dimensional image reconstruction neural network. Without requiring hardware modifications, IRIS enables imaging at rates spanning several hundred hertz to the kilohertz regime while preserving the field of view. We show that IRIS accurately retains both the spatial structure and temporal dynamics of fluorescent signals in vivo. The framework generalizes across imaging modalities, supporting volumetric cortical imaging and large-field recording of over 1,000 neurons at 60 Hz. Applied across wakefulness-anesthesia-recovery cycles, IRIS successfully captures state-dependent shifts in neuronal synchrony across multiple layers of the motor-sensory cortex. By decoupling acquisition speed from spatial sampling, IRIS provides a robust computational paradigm for high-speed functional neuroimaging.