Application Of Artificial Neural Networks On Biomedical Microscopy Image Super Resolution

We combine deep learning with microscopy imaging to achieve super resolution under a large field of view(FOV).Without introducing any extra hardware setups,this approach is able to increase the imaging throughput of the conventional wide-field microscopy and light-sheet fluorescent microscopy on a large scale.A state-of-the-art generative adversarial network(GAN)is adapted to learn the mapping from the low resolution images to their high resolution counterparts,where the low resolution ones used during the training stage is generated by an image degrading model that accurately simulates the transfer function of the optical system,thus the inefficient and error-prone image registration can be avoided in the data preparation stage.While the training is done,the GAN takes as input a single low resolution measurement of a new specimen,and reconstructs a high-resolution one with large FOV.Its capacity has been broadly demonstrated via imaging various types of samples,such as the bright filed images of the USAF resolution target,the color images of human pathological slides,the fluorescent images of fibroblast cells,and the light sheet images of deep tissues in transgenic mouse brain.Our approach outperforms the conventional multi-frame methods in temporal resolution,and achieves an equal spatial resolution to them.The fidelity of our reconstructions is evaluated by calculating the peak signal-to-noise ratio(PSNR)and the structural similarity(SSIM)index.When applied to the biomedical researches(including but not limited to cell counting tasks and histopathological diagnoses),this GAN-based approach is capable of greatly increasing their accuracy.Though only two dimensional image super resolution is illustrated currently,we have faith to extend its application to three dimensional mapping such as the light sheet volumetric imaging and the light field imaging.