FPGA implementation of sequence-to-sequence encoder-decoder deep learning model for real-time fluorescence parameter estimation through SwissSPAD2 camera

Fluorescence Lifetime Imaging (FLI) is widely used in biomedical research for its high sensitivity and detailed cellular insights. It is effective for multiplexed studies and provides information about metabolic states, protein interactions, and ligand-target engagements, independent of signal intensity and fluorophore concentration. Traditional FLI techniques are slow and computationally demanding. This study addresses these limitations by implementing a deep learning-based sequence-to-sequence model directly into the SwissSPAD2 (SS2), a photon-counting camera coupled to an FPGA. The integration, enhanced by efficient resource management and parallel computation, allows for near-real-time output of FLI parameters. We tested the performance using in-silico and in-vitro data, demonstrating that this approach advances the feasibility of real-time FLI analysis, with promising implications for both preclinical and clinical applications.