Photonic Data Science

Fluorescence lifetime imaging microscopy (FLIM) is a powerful tool for studying molecular interactions and cellular microenvironments. Conventional lifetime estimation relies on curve fitting, a computationally intensive and noise-sensitive process. To address these challenges, we developed a dual-branch deep learning-based framework for fit-free estimation of FLIM parameters. It integrates multitasking and knowledge-informed learning to improve accuracy and robustness. The dual-branch model consists of an autoencoder (AE) and a convolutional neural network. The AE processes decay traces to generate a reconstruction and a latent representation, which is subsequently fed into the CNN to predict fluorescence lifetimes and abundances. To overcome the limitation of scarce experimental data, we constructed a training dataset by combining simulated and experimental traces. The method was first validated with a leave-one-IRF-out cross-validation on simulated data to confirm its capacity to generalize across instrument response functions (IRFs). Next, we evaluated the method on experimental data, with the training data from different proportions of simulated and experimental decays. Compared to baseline CNN and the FLIMview software (curve-fitting based), the proposed model demonstrated superior performance even with limited experimental data. These findings underscore the potential of deep learning to enable scalable, real-time FLIM analysis.