Semi-Synthetic Test Data Generation and Content-aware Image Restoration of Fluorescent Microscopy Images

Anonymous

Abstract:

Although fluorescence microscopes theoretically allow imaging with resolutions at the diffraction limit, in practice the resolution is often drastically limited by several factors: Image quality of fluorescence microscopes is always limited by the interaction of the light with the tissue, often resulting in a low image intensity and weak contrast of the biological structure. Excitation of fluorophores and molecules outside the imaged area leads to photobleaching or phototoxic effects in the sample. In addition, diffraction effects and the low intensity of fluorescent light reduce the quality of the images.
In this paper, we present an application of using deep learning to improve the image quality of a fluorescence microscope in biological imaging. We trained a predefined architecture of a deep convolutional neural network provided by the CSBDeep project with semi-synthetic training data of mitochondria and cardiomyocytes. The results show good improvement of the image quality.