A hybrid deconvolution algorithm in three-dimensional fluorescence microscopy

Three-dimensional fluorescence microscopy is often affected by out-of-focus blur and noise, which limits accurate visualization of biological structures. Traditional deconvolution methods face a trade-off: non-iterative approaches often leave residual background contamination, whereas iterative methods such as Richardson–Lucy (RL) may amplify noise under strong background conditions. To address these limitations, we propose a training-free hybrid deconvolution framework that combines nearest-neighbor-based background suppression with RL iterative reconstruction. The proposed strategy separates out-of-focus background removal from structural refinement, enabling improved restoration without requiring dataset-specific model training. The framework was evaluated on synthetic structures, fluorescent beads, C. elegans embryos, and three-dimensional neuronal cell datasets. Compared with RL reconstruction, the proposed method improved the PSNR from 20.61 dB to 25.92 dB and the SSIM from 0.780 to 0.944 on the Synthetic Hollow Bars dataset. For biological datasets, the proposed method improved the SBR by 22.28 dB for fluorescent bead images and by 5.68 dB for neuronal cell images compared with the original blurred images, enabling clearer visualization of fine structures under strong background contamination. Moreover, the proposed method achieved approximately 2.5–8-fold faster execution than the representative SecMRA framework under identical hardware conditions. These results demonstrate that the proposed hybrid framework provides an effective and computationally practical solution for three-dimensional fluorescence microscopy restoration.