Fast GPU simulations of the FRAP experiment

Abstract

In this thesis we study several models of the Fluorescence Recovery After Photobleaching (FRAP) experiment. FRAP is a technique used to estimate the diffusion coefficient of fluids based on Confocal Laser Scanning Microscopy (CLSM). A fluorescent sample is first photobleached on a user defined region. Then, by studying the recovery of fluorescence in the bleaching region, one can retrieve important parameters of the fluid such as the diffusion coefficient and binding constants by fitting a model to the data. We implemented and compared three models of the FRAP experiment. The first model assumes bleaching and image acquisition is an instantaneous process. The second model, based on the first one, introduces multiple bleach frames. The final model takes into account the scanning movement of the CLSM and is computationally much more complex. For the instantaneous models, two schemes are introduced and compared against each other to ensure correct implementation of the algorithms. The first scheme uses the spectral method to solve the diffusion-reaction equations and the second uses a stochastic formulation of the problem. The last model, due to its complexity, has only been implemented stochasticaly. All three models have been implemented on Graphical Processing Units (GPUs) using the OpenCL API in C++. The GPU has a massively parallel architecture that can be exploited for scientific computing. These schemes are ”embarrassingly parallel” and thus suitable for a GPU implementation. By comparing the different models, we see that a good compromise between precision and computing resource is given by the instantaneous bleaching with multiple bleach frames model. Because of the scanning nature of the CLSM, we would expect the last model to reveal some asymmetry in the results. These were only found for extreme and unrealistic parameters and it is thus not necessary to simulate the FRAP experiment with such complexity.