ABSTRACT Three‐photon fluorescence microscopy (3PFM) enables high‐resolution volumetric imaging in deep tissues but is fundamentally constrained by a trade‐off among imaging speed and spatial resolution, due to low photon flux and prolonged laser exposure. We present DeepR‐SXYZ, a deep learning framework that achieves pixel dwell time equivalent to hundreds of nanoseconds (0.168–0.38 µs) through sparse X‐Y‐Z reconstruction for 3PFM. Trained on paired datasets consisting of sparsely acquired low‐resolution volumetric scans (e.g., 256 × 256 pixels, 1.2 µs per pixel, Z step: 16 µm) and their corresponding densely sampled high‐resolution counterparts (e.g., 512 × 512 pixels, 3.2 µs per pixel, Z step: 2 µm), DeepR‐SXYZ integrates convolutional neural networks (CNNs) with a structure‐dynamic attention (SDA)‐enhanced transformer, synergistically capturing intra‐layer morphological features (for X‐Y plane reconstruction) and inter‐layer dynamic variations (for Z‐axis interpolation). This design enables accurate 3D volume reconstruction from sparsely sampled data. Experimental validation on cerebral vasculature and muscle macrophages demonstrates that DeepR‐SXYZ achieves 8.8× acceleration in X‐Y plane imaging over conventional dense sampling and >60% Z‐axis layer recovery, substantially improving imaging throughput and reducing single‐pixel dwell time significantly. Crucially, the framework enables large‐field 3D cerebral vasculature imaging and dynamic volumetric tracking of cellular behaviors, revealing previously inaccessible spatiotemporal biological processes. This work establishes a computational paradigm for high‐speed, low‐phototoxicity three‐photon fluorescence microscopic imaging via sparse X‐Y‐Z reconstruction, effectively balancing imaging speed and spatial resolution.
Li et al. (Fri,) studied this question.