Academic

PhD Oral Presentation

An important analysis in fluorescence spectroscopy is molecular dynamics, in particular, quantification of diffusion coefficient. Imaging Fluorescence Correlation Spectroscopy (FCS) is one of the techniques for analyzing in-plane diffusion. Evaluation of data, however, is complex and remains an active field of investigation. Looking through a different lens, we apply a convolutional neural network (CNN) to determine the diffusion coefficient of freely diffusing molecules. CNN has been at the forefront of deep learning approaches. While it is better known for computer vision, CNN can also be used for the analysis of temporal data. We illustrate this with our case study of using CNN for model identification of autoregression moving average (ARMA) time series. Extending the idea of model-based learning to our fluorescence spectroscopy analysis, our network is trained with purely simulated data and the ground truth diffusion coefficients as training labels. Evaluations by our trained network is an end-to-end process. It processes an input image stack directly. Apart from centering and scaling of the input data, there are no intermediate statistical calculations, such as autocorrelation or curve fitting, involved.

An important analysis in fluorescence spectroscopy is molecular dynamics, in particular, quantification of diffusion coefficient. Imaging Fluorescence Correlation Spectroscopy (FCS) is one of the techniques for analyzing in-plane diffusion. Evaluation of data, however, is complex and remains an active field of investigation. Looking through a different lens, we apply a convolutional neural network (CNN) to determine the diffusion coefficient of freely diffusing molecules. CNN has been at the forefront of deep learning approaches. While it is better known for computer vision, CNN can also be used for the analysis of temporal data. We illustrate this with our case study of using CNN for model identification of autoregression moving average (ARMA) time series. Extending the idea of model-based learning to our fluorescence spectroscopy analysis, our network is trained with purely simulated data and the ground truth diffusion coefficients as training labels. Evaluations by our trained network is an end-to-end process. It processes an input image stack directly. Apart from centering and scaling of the input data, there are no intermediate statistical calculations, such as autocorrelation or curve fitting, involved.