Unsupervised deep learning enables blur-free super-resolution in two-photon microscopy

Abstract

We developed an unsupervised deep learning method to simultaneously perform deblurring, super-resolution, and segmentation of two-photon microscopy images. Two-photon microscopy is an excellent technique for non-invasively observing deep biological tissues, but blurring during deep imaging has been a challenge. Conventional deblurring methods have limited performance and are not suitable for deblurring two-photon microscopy images. Moreover, methods that simultaneously perform segmentation, which is usually required in downstream analysis, have not been developed. Therefore, in this method (TENET), we precisely modeled the blur of two-photon microscopy and simultaneously achieved deblurring, super-resolution, and segmentation through unsupervised deep learning. In simulation and experimental data, we achieved deblurring, resolution improvement, and segmentation accuracy surpassing conventional methods. Furthermore, we applied the method to live imaging of microglia, enabling quantitative 3D morphological analysis that was previously difficult. This method allows non-invasive visualization of detailed structures in deep biological tissues, and is expected to lead to a more high-definition understanding of biological phenomena. Future applications to time-series morphological analysis of microglia are anticipated.