Djupinlärningsbaserad FCS

Fluorescence correlation spectroscopy (FCS) is a method for statisticallyanalyzing the time correlation of fluctuations in a sample’s fluorescenceintensity. The data obtained from experiments are often noisy, and thecurrent procedure—where the final result is derived through curve fitting ofthe experimental data’s autocorrelation—does not necessarily yield the highestprecision for diffusion time and concentration, which are the key parameters ofinterest. In recent years, machine learning (ML) has been successfully appliedto a wide range of problems, particularly those involving noisy data. Thisthesis explores the potential of improving the accuracy of FCS using deepsupervised learning by training a neural network on raw data with knownparameter values. To generate sufficiently large datasets for this purpose,a large number of experiments have been simulated. The results from thetraining were then evaluated by comparing them to the traditional approachbased on autocorrelation curve fitting. Deep learning demonstrated higherprecision than curve fitting for the same measurement duration. In fact,for measurement durations around one second, curve fitting approximatelyrequired twice as long measurement time to achieve the same accuracy indetermining diffusion time as deep learning.