Cuda 8

Each frame from an image stack was first box-filtered with the box size of 4 times of the FWHM of a 2D Gaussian PSF. We note that each pixel was weighted by the inverse of its variation during such box-filtering. The low-pass filtered image was then extracted from the raw image, followed by recognition of local maximums. The local maximums from all the frames of the image stack were then submitted for 2D-Gaussian single-PSF fitting.
The 2D-Gaussian single-PSF fitting were performed in GPU (Nvidia GTX 1060, CUDA 8.0) using the Maximum Likelihood Estimation (MLE) algorithm. In brief, the likelihood function at each pixel was built by convolving the Poisson distribution of the shot noise governed by the photons emitted from fluorophores nearby, and the gaussian distribution of the readout noise that characterized by the expectation, variation, and the analog-to-digital conversion factor that pre-calibrated as mentioned above. The fitting accuracy was estimated by Cramér-Rao lower bound (CRLB).

The configuration of the computational station was based in two Intel Xeon E5-2630 v4 2.2 GHz Processors with 10 cores, 64GB DDR4 RAM and operation frequency 2133 MHz. The station was additionally equipped with two nVidia Tesla P100 graphic cards with 12GB memory on each and double precision of 64 bit. The parallel computing library NVidia Cuda 8.0 was used and the computations were operated by 64 bit CentOS 7.4. The coding was performed in Visual Studio 2017 using C# and Python 3.6.