RESUMO
Multiphoton fluorescence microscopy excited with femtosecond pulses at high repetition rates, particularly in the range of 100's MHz to GHz, offers an alternative solution to suppress photoinduced damage to biological samples, for example, photobleaching. Here, we demonstrate the use of a U-Net-based deep-learning algorithm for suppressing the inherent shot noise of the two-photon fluorescence images excited with GHz femtosecond pulses. With the trained denoising neural network, the image quality of the representative two-photon fluorescence images of the biological samples is shown to be significantly improved. Moreover, for input raw images with even SNR reduced to -4.76 dB, the trained denoising network can recover the main image structure from noise floor with acceptable fidelity and spatial resolution. It is anticipated that the combination of GHz femtosecond pulses and deep-learning denoising algorithm can be a promising solution for eliminating the trade-off between photoinduced damage and image quality in nonlinear optical imaging platforms.