Convolutional Neural Networks for Structured Illumination Microscopy

Abstract

Optical super-resolution, the ability to exceed the Abbe diffraction limit, has revolutionized microscopy during the last decades. At the same time, machine learning models, and in particular neural networks, are being created for performing more and more complex tasks. In this work, we set out to developing a convolutional neural network for the reconstruction of images from structured illumination microscopy, one of the leading super-resolution techniques. To achieve this, we designed the model’s architecture and implemented it with the Python library PyTorch, we prepared an artificial dataset with a preexisting microscope simulator, we trained the model with these images, and we tested it with new ones to assess its generalizability. We find that, under the conditions simulated, our model recreates the original images with a high level of detail, surpassing the resolution of a classical reconstruction method, the joint Richardson-Lucy algorithm, by approximately 35 %, and obtaining better scores in different similarity indices