Deep learning-enabled pixel super-resolution in fluorescence lifetime imaging microscopy (FLIM)

Abstract

Fluorescence lifetime imaging microscopy (FLIM) is fundamentally constrained by long pixel dwell times and low photon budgets, which limit its deployment in many clinically relevant settings. To mitigate this, we present FLIMPSR, a deep learning-based pixel super-resolution framework for multi-channel FLIM that enhances resolution without requiring hardware modifications. Using a conditional generative adversarial network, the method reconstructs high-resolution lifetime and intensity maps directly from undersampled data. Reliable fivefold pixel super-resolution is demonstrated on human tissue FLIM data, with quantitative improvements over interpolation approaches and successful recovery of mid- and high-spatial-frequency content. Compared to diffusion-based super-resolution, FLIMPSR provides more stable reconstructions and substantially faster inference. This approach enables accelerated FLIM imaging and supports translation toward compact and clinically relevant lifetime imaging systems.