Physics-guided training-free neural network for suppressing out-of-focus background in fluorescence microscopy

Abstract

Fluorescence microscopy is renowned for its high sensitivity and specificity, making it a vital tool in biomedical research. However, during practical imaging, defocused background signals often introduce strong interference, significantly reducing image contrast and consequently affecting subsequent structural analysis and quantitative assessments. Traditional computational methods for suppressing defocused background, such as deconvolution or filtering algorithms, typically rely on probabilistic models or the assumption that “defocused background is low-frequency signal.” These approaches are prone to noise amplification or loss of valid signals in complex imaging scenarios, compromising processing accuracy and reliability. To address the defocused background interference in fluorescence imaging, we propose a physics-guided training-free neural network for effective background suppression. This method incorporates the forward physical process of fluorescence imaging into the network architecture design, utilizing an optical propagation model as a physical constraint to guide the network's self-adaptive optimization without the need for annotated data. Requiring only a single fluorescence image for network parameter updates, our approach significantly reduces dependency on paired training data and overcomes the bottleneck of data acquisition in supervised learning. Experimental results demonstrate outstanding background suppression capability of the proposed method on both spoke-like sample and real biological tissue images, achieving a contrast improvement of up to approximately 5.5 times. This method establishes a novel imaging paradigm for background removal with neural networks, offering an effective solution to background interference in fluorescence imaging.