Self-supervised deep-learning two-photon microscopy
Yuezhi He, Jing Yao, Lina Liu, Yufeng Gao, Jia Yu, Shiwei Ye, Hui Li, Wei Zheng
Abstract
Artificial neural networks have shown great proficiency in transforming low-resolution microscopic images into high-resolution images. However, training data remains a challenge, as large-scale open-source databases of microscopic images are rare, particularly 3D data. Moreover, the long training times and the need for expensive computational resources have become a burden to the research community. We introduced a deep-learning-based self-supervised volumetric imaging approach, which we termed “Self-Vision.” The self-supervised approach requires no training data, apart from the input image itself. The lightweight network takes just minutes to train and has demonstrated resolution-enhancing power on par with or better than that of a number of recent microscopy-based models. Moreover, the high throughput power of the network enables large image inference with less postprocessing, facilitating a large field-of-view ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" id="m1"> <mml:mrow> <mml:mn>2.45</mml:mn> <mml:mtext> </mml:mtext> <mml:mi>mm</mml:mi> <mml:mo>×</mml:mo> <mml:mn>2.45</mml:mn> <mml:mtext> </mml:mtext> <mml:mi>mm</mml:mi> </mml:mrow> </mml:math> ) using a home-built two-photon microscopy system. Self-Vision can recover images from fourfold undersampled inputs in the lateral and axial dimensions, dramatically reducing the acquisition time. Self-Vision facilitates the use of a deep neural network for 3D microscopy imaging, easing the demanding process of image acquisition and network training for current resolution-enhancing networks.