Diffusion Priors Improve Low-Light Fluorescence Microscopy Images

Fluorescence microscopy images suffer from diffraction blur, which standard Richardson-Lucy (RL) deconvolution tries to reverse by modeling photon counts under Poisson noise. However, RL is ill-posed and often amplifies noise, especially in low-light conditions. Traditional regularizers like total variation (TV) reduce instability but can oversmooth fine structures such as filaments and punctae. The problem is fundamental: the measurement alone provides insufficient evidence to distinguish signal from noise.

Chen and Howard address this by incorporating a score-based diffusion model as a learned generative prior within a decoupled inverse-problem framework. The diffusion prior guides RL iterations toward realistic image structures while RL enforces consistency with the raw photon data. Tested on diverse cell morphologies, the hybrid method significantly reduces noise amplification and better preserves weak, fine details compared to TV-regularized RL or plain RL. This allows scientists to achieve high-resolution images under lower photon counts, reducing phototoxicity and enabling longer live-cell imaging sessions.