Researchers at the California NanoSystems Institute (CNSI) at UCLA have created a novel, data-driven, deep learning framework that allows for the generation of super-resolution images directly from images acquired on conventional, diffraction-limited microscopes. This is completed without prior knowledge about the sample and/or the image formation process to super-resolve microscopic images beyond the diffraction limit. The deep network output is extremely fast, without any iterations or parameter searches. In another demonstration, the researchers have used a deep neural network to perform virtual histological staining of a label free tissue sections, using a single autofluorescence image. This transformation, which uses only the endogenous contrast of the tissue section, was applied to multiple types of tissues and stains and blindly validated by board a panel of pathologists.

 

These results represent an important step towards computational microscopy and illustrate some of the potential machine learning to the field.

 

Learning Objectives: