Enhancing Microscopy through Deep Learning

Extending the field of view, depth of-field and resolution of images acquired using a microscope is the end goal of various techniques involving different combinations of hardware and/or software improvements. Here, we demonstrate that a convolutional deep neural network can enhance optical microscopy images, without any hardware modification to the microscope. For this aim, a deep network is trained using experimentally acquired high-resolution and low-resolution microscopic images of different samples. Following its training, the deep network remains fixed and rapidly outputs an image with enhanced resolution, matching the performance of a high numerical aperture objective lens, while also significantly advancing its limited field-of-view and depth-of-field. This deep learning based framework can be broadly applied to imagers at different parts of the electromagnetic spectrum and it demonstrates the potential of convolutional neural networks for solving inverse problems in imaging, which is especially important when accurate modeling of the light-matter interaction is a challenging task.