IMAGING FLUORESCENCE CORRELATION SPECTROSCOPY ANALYSIS USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Fluorescence Correlation Spectroscopy (FCS) utilizes the fluctuations of intensities to measure the dynamics of fluorescent particles of interest through spatial and temporal statistical analysis. An intensity trace is recorded over long time scales to obtain sufficiently robust statistics, and correlated at a semi-logarithmic timescale to obtain autocorrelation functions (ACF) corresponding to the observed region. In this work, we train a CNN to predict diffusion coefficients using simulations of freely diffusing particles. We observe that CNNs can be trained with simulated data using a suitable noise model, while retaining generalizability to experimental lipid bilayer data. We show that a simple CNN architecture manages to outperform conventional nonlinear least-squares (NLS) fitting in terms of precision and accuracy on simulated training data. We then apply our CNN fitting procedures on ACFs derived from shorter measurement times, and show that the CNN approach achieves better precision than NLS fitting.