Comparison of computational methods developed to address depth-variant imaging in fluorescence microscopy

摘要

In three-dimensional microscopy, the image formation process is inherently depth variant (DV) due to the refractive index mismatch between the imaging layers. In this study, we present a quantitative comparison among different image restoration techniques developed based on a depth-variant (DV) imaging model for fluorescence microscopy. The imaging models employed by these methods approximate DV imaging by either stratifying the object space (analogous to the discrete Fourier transform (DFT) "overlap-add" method) or image space (analogous to the DFT "overlap-save" method). We compare DV implementations based on maximum likelihood (ML) estimation and a previously developed expectation maximization algorithm to a ML conjugate gradient algorithm, using both these stratification approaches in order to assess their impact on the restoration methods. Simulations show that better restoration results are achieved with iterative methods implemented using the overlap-add method than with their implementation using the overlap-save method. However, the overlap-save method makes it possible to implement a non-iterative DV inverse filter that can trade off accuracy of the achieved result for computational speed. Results from a non-iterative regularized inverse filtering approach are also presented.