Zero‐shot learning enabled image enhancement in Bessel beam light sheet fluorescence microscopy

Bessel beams offer superior depth of field and self‐healing properties compared to conventional Gaussian beams in light sheet fluorescence microscopy (LSFM), enabling enhanced axial resolution and optical sectioning. However, their side lobes generate significant out‐of‐focus fluorescence, which degrades image contrast and signal‐to‐noise ratio (SNR). In this context, we demonstrate the use of a physics‐guided, unsupervised deep learning model‐based image restoration pipeline for Bessel beam LSFM images to suppress the side‐lobe‐induced artifacts that does not require ground‐truth data. A synthetic point spread function (PSF) of the Bessel beam LSFM is simulated using the Debye‒Wolf model and wave‐optical simulations based on the optical parameters of the microscope. The modeled PSF is then used with a hybrid zero‐shot deconvolution and U‐Net model for denoising and contrast enhancement of fluorescent biological sample images without the requirement of paired ground‐truth data. The parameters of the deep learning model are optimized to obtain substantial image quality improvements in terms of SNR and contrast across different biological specimens, namely, mouse kidney tissues (in vitro) and three‐dimensional image stack of trabecular meshwork of a porcine eye sample (ex vivo). The results showed up to 2.2‐fold enhancement in SNR and 89% enhancement in Michelson contrast. The robustness of the method under moderate PSF mismatch is also investigated. The proposed approach holds significant potential in a variety of biomedical imaging applications and monitoring of sub‐cellular dynamics utilizing LSFM.