In this paper, we discuss what is, what is not, and what is only sort of superresolution microscopy. We begin by considering optical resolution, first in terms of diffraction theory, then in terms of linear-systems theory, and finally in terms of techniques that use prior information, nonlinearity, and other tricks to improve resolution. This discussion reveals two classes of superresolution microscopy, “pseudo” and “true.” The former improves images up to the diffraction limit, whereas the latter allows for substantial improvements beyond the diffraction limit. The two classes are distinguished by their scaling of resolution with photon counts. Understanding the limits to imaging resolution involves concepts that pertain to almost any measurement problem, implying a framework with applications beyond optics.
References
“Superresolution” is also sometimes used to describe sub-pixel resolution in an imaging detector. Since pixels are not necessarily related to intrinsic resolution, we do not consider such techniques here.
Superresolution fluorescence microscopy was the 2008 “Method of the Year” for Nature Methods, and its January 2009 issue contains commentary and interviews with scientists playing a principal role in its development. This is a good “cultural” reference.
For example, a STED microscope is sold by the Leica Corporation.
Equation (1) gives the lateral resolution. The resolution along the optical axis is poorer: .
However, the magnification of an objective does not determine its resolution.
A subtle point: The modulation transfer function is zero beyond a finite spatial frequency; yet the response in Fig. 3 is non-zero at all frequencies. The explanation is that an object of finite extent has a Fraunhofer diffraction pattern (Fourier transform) that is analytic, neglecting noise. Analytic functions are determined by any finite interval (analytic continuation), meaning that one can, in principle, extrapolate the bandwidth and deduce the exact behavior beyond the cutoff from that inside the cutoff. In practice, noise cuts off the information (Fig. 3). See Lucy (Ref. 28) for a brief discussion and Goodman's book (Ref. 5) for more detail.
One should set the errors to be the square root of the smooth distribution value deduced from the initial fit and then iterate the fitting process (Ref. 46). The conclusions however, would not change, in this case.
Sparseness can improve resolution in other ways, as well. For example, the new field of compressive sensing also uses a priori knowledge that a sparse representation exists in a clever way to improve resolution (Ref. 58).