Fast In Vivo Two-Photon Fluorescence Imaging via Lateral and Axial Resolution Restoration With Self-Supervised Learning

Abstract

Two-photon fluorescence (TPF) imaging opens a new avenue to achieve high resolution at extended penetration depths. However, it is difficult for conventional TPF imaging systems to simultaneously achieve high resolution and speed. In this work, we develop an innovative deep learning framework of Lateral and Axial Resolution Restoration (LARR) to break the contradiction between imaging resolution and speed. LARR employs a self-supervised training scheme to computationally restore the sparsely sampled TPF images to resolution isotropic images by 4-fold axial and 16-fold lateral resolution enhancement. The simulation studies and experimental results demonstrate the excellent performance of LARR to preserve fine structural features with improved signal-to-noise ratio and structure similarity index. Moreover, the TPF imaging system with the LARR is able to achieve 60-fold improved imaging speed and comparable resolution as compared with the conventional TPF system. The outstanding performance makes LARR a potential tool for fast TPF imaging with high resolution.