Descattering and image restoration with a transformer-based neural network in deep tissue imaging

Imaging biological structures deep inside tissues is crucial but challenging due to common light scattering. This study proposes a multiattention network that directly maps degraded scattering two-photon excitation fluorescence (TPEF) images to high-quality scattering-free images, thereby computationally extending the imaging depth for TPEF without requiring complex optical additions. The model relies solely on simulated data rather than well-registered real data pairs, and is trained to descatter and restore hidden spatial information at greater depths. Quantitative evaluations on simulated fluorescent beads and vasculature show significant performance improvements in peak signal-to-noise ratio (23 to 29 dB) and structural similarity index (23×) compared to the raw data. We also apply the framework to various ex vivo and in vivo experiments, achieving clear visualization of lipid droplets up to a depth of 1,300 μm and of vascular structure and astrocytes up to 950 μm and 500 μm, respectively, in live mouse brains at lower excitation powers.