SIRP: A Self-Supervised Fluorescence Denoising method with Implicit Representation Priors

Abstract

Fluorescence microscopy plays a critical role in live-cell imaging, yet the captured images often suffer from significant noise due to limited excitation intensity and constrained exposure times. Deep learning has emerged as a powerful solution for image denoising, especially the self-supervised methods can achieve good restoration performance without requiring clean target data. However, their performance in preserving fine image details remains challenging under low signal-to-noise ratio (SNR) conditions. In this work, we propose a novel self-supervised denoising framework named Self-Supervised Implicit Representation Prior Image Restoration Network (SIRP). It is based on the neighbor-sampling strategy and leverages image representation priors to guide the network training procedure. The neighbor-sampling strategy takes advantage of the inherent statistical consistency within local image regions to denoise. The image representation priors are learned through the Implicit Neural Representations (INR) that map spatial coordinates directly to their corresponding pixel intensities to suppress noise and effectively preserve details. Results of SIRP on simulated data and experimental-acquired data showed superior denoising performance and detail-preserving ability compared to filter-based approaches and self-supervised approaches, highlighting the potential of INR-guided architectures for fluorescence microscopy image restoration.