Numerical dark-field imaging using deep-learning.

Dark-field microscopy is a powerful technique for enhancing the imaging resolution and contrast of small unstained samples. In this study, we report a method based on end-to-end convolutional neural network to reconstruct high-resolution dark-field images from low-resolution bright-field images. The relation between bright- and dark-field which was difficult to deduce theoretically can be obtained by training the corresponding network. The training data, namely the matched bright- and dark-field images of the same object view, are simultaneously obtained by a special designed multiplexed image system. Since the image registration work which is the key step in data preparation is not needed, the manual error can be largely avoided. After training, a high-resolution numerical dark-field image is generated from a conventional bright-field image as the input of this network. We validated the method by the resolution test target and quantitative analysis of the reconstructed numerical dark-field images of biological tissues. The experimental results show that the proposed learning-based method can realize the conversion from bright-field image to dark-field image, so that can efficiently achieve high-resolution numerical dark-field imaging. The proposed network is universal for different kinds of samples. In addition, we also verify that the proposed method has good anti-noise performance and is not affected by the unstable factors caused by experiment setup.