The project is generically related to advanced optical microscopy and more particularly to fluorescence superresolution microscopy. For many years, optical microscopy has been troubled by an apparently insurmountable obstacle: the diffraction limit imposed by the wave nature of light, which restricts its capacity to resolve sample details below 200 nm, an increasingly inadequate size limit. It is then of little surprise that the knocking out of the diffraction resolution barrier by new techniques such as Stimulated Emission Depletion (STED) and Stochastic Optical Reconstruction Microscopy (PALM-STORM) has so quickly deserved worldwide recognition and a Nobel Prize (Chemistry 2014).
Unfortunately, the amazingly resolved pictures produced by the STED and STORM instruments (down to a few tens of nm) come at a hefty price: a sophisticated and expensive technology in the case of STED (Leica’s sells for €1M, for example) or a very slow image acquisition in the case of PALM-STORM (up to 30 min. to obtain a single image, in some cases). Therefore, the microscopy market eagerly seeks super-resolution alternatives with more balanced trade-offs. A good example is Photon Reassignment Microscopy that, despite the modest resolution improvement it offers, around 150 nm, has been very quickly adopted and turned into commercial products by several companies (such as the Yokowaga CSU-W1 SoRa module or the Zeiss AiryScan-enabled confocal).
This project is a proof-of-concept of a technology capable of satisfying this market need through a combination of approaches. Precise over-sampling of an object can produce super-resolution images, as demonstrated in the well-known multi-frame super-resolution scheme. The application of this principle to microscopy, under the name of Translation Microscopy (TRAM), produces images with lateral resolution better than 50 nm. However, in TRAM, image acquisition is cumbersome and image reconstruction extremely slow (hours) so as to render it ineffective. On the other hand, deep learning neural networks have been proven capable of effortlessly estimating the high-resolution details of a low-resolution image, including optical micrographs, based on prior training. However, the network “hallucinates” the added details so that these may not faithfully represent real features present in a particular scene, which is unacceptable in quantitative microscopy.
Here, we intend to distil and merge these two ideas. We have developed and patented a laser confocal microscope that is able to project on a sample arbitrary laser excitation patterns, positioned with nanometre accuracy and at hundreds of kHz rates. Now, we intend to develop the means to hand over control of this laser-scanning device to a neural-network intelligence that will learn how to probe a sample so as to extract and display actual superresolution information, with the goal to be competitive in quality with STED and STORM, but at a much more reasonable speed and cost.