Motion Artifact Correction in Deep-Tissue Three-Photon Fluorescence Microscopy Using Adaptive Optical Flow Learning With Transformer.

Abstract

Three-photon fluorescence microscopy (3PFM) enables high-resolution volumetric imaging in deep tissues but is often hindered by motion artifacts in dynamic physiological environments. Existing solutions, including surgical fixation and conventional image registration algorithms, frequently fail under intense and nonuniform motions, particularly in low-texture or highly deformed regions. To overcome these problems, we propose StabiFormer, a transformer-based optical flow learning network designed for robust motion correction. Central to StabiFormer is the stable-dynamic feature extractor, which captures interlayer dynamics to facilitate accurate image registration. Our validation across cerebrovascular and intestinal 3PFM datasets demonstrates that StabiFormer achieves near-zero displacement error relative to ground truth in brain vasculature. Furthermore, it enables artifact-free 3D visualization of intestinal macrophages and vasculature at 300 μm depth, a physiologically relevant depth for studying intestinal immune microvasculature. These results establish a noninvasive computational solution for motion-artifact-free volumetric imaging, paving the way for quantitative investigations in previously inaccessible dynamic organ systems.