Three-Dimensional Virtual Optical Clearing With Cycle-Consistent Generative Adversarial Network

Abstract

High-throughput deep tissue imaging and chemical tissue clearing protocols have brought out great promotion in biological research. However, due to uneven transparency introduced by tissue anisotropy in imperfectly cleared tissues, fluorescence imaging based on direct chemical tissue clearing still encounters great challenges, such as image blurring, low contrast, artifacts and so on. Here we reported a three-dimensional virtual optical clearing method based on unsupervised cycle-consistent generative adversarial network, termed 3D-VoCycleGAN, to digitally improve image quality and tissue transparency of biological samples. We demonstrated the good image deblurring and denoising capability of our method on imperfectly cleared mouse brain and kidney tissues. With 3D-VoCycleGAN prediction, the signal-to-background ratio (SBR) of images in imperfectly cleared brain tissue areas also showed above 40% improvement. Compared to other deconvolution methods, our method could evidently eliminate the tissue opaqueness and restore the image quality of the larger 3D images deep inside the imperfect cleared biological tissues with higher efficiency. And after virtually cleared, the transparency and clearing depth of mouse kidney tissues were increased by up to 30%. To our knowledge, it is the first interdisciplinary application of the CycleGAN deep learning model in the 3D fluorescence imaging and tissue clearing fields, promoting the development of high-throughput volumetric fluorescence imaging and deep learning techniques.