Deep learning for quantitative bi-exponential fluorescence lifetime imaging (Conference Presentation)

Fluorescence lifetime imaging (FLI) has become an invaluable tool in the biomedical field by providing unique, quantitative information about biochemical events and interactions taking place within specimens of interest. Applications of FLI range from superresolution microscopy to whole body imaging using visible and near-infrared fluorophores. However, quantifying lifetime can still be a challenging task especially in the case of bi-exponential applications. In such cases, model based iterative fitting is typically employed but necessitate setting up multiple parameters ad hoc and can be computationally expensive. These facts have limited the universal appeal of the technique and methodologies can be specific to certain applications/technology or laboratory bound. Herein, we propose a novel approach based on Deep Learning (DL) to quantify bi-lifetime Forster Resonance Energy Transfer (FRET). Our deep neural network outputs three images consisting of both lifetimes and fractional amplitude. The network is trained using synthetic data and then validated on experimental FLI microscopic (FLIM) and macroscopic data sets (MFLI). Our results demonstrate that DL is well suited to quantify wide-field bi-exponential fluorescence lifetime accurately and in real time, even when it is difficult to obtain large scale experimental training data.