To Improve In-Vivo Bio Images of Fast Temporal Focusing Multiphoton Microscopy by Multi-Stage U-Net Image Restoration

The imaging speed of Temporal focusing multiphoton excitation microscopy (TFMPEM) is up to hundreds frames rate. However, the plane illumination manner suffers from the sever scattering of biotissue and signal crosstalk that blurs the image. And the deeper the worse. Nevertheless, the high acquisition rate decreases the effective excited fluorescent, which reduces the signal-to-noise ratio (SNR) of the image. In order to solve the scattering and low SNR issues, the deep learning method is proposed to restore the TFMPEM image. In this work, we construct a powerful neuron network which called multi-stage 3D U-Net. Different from the cascade method, it becomes more connection between each U-Net. The previous stage information can share with the next stage, instead of seeing as independent. Thus, we try to restore the TFMPEM via this network with Point scanning multiphoton excitation microscopy (PSMPEM) image as the ground truth. But before that way, our two systems are not sharing the same optical path architecture, it needs to do the registration first. For cross modality registration, we utilize Voxelmorph which is also a 3D U-Net architecture. And it can do the not only global but also local deformation, is flexible than classical algorithm. Hence, we do the registration and restoration via all deep learning method. Therefore, the peak signal-to-noise ratio (PSNR) of the image can be improved around 20 to 30 dB and, and structural similarity (SSIM) is close to 0.9