Abstract
Fluorescence microscopy performance can be significantly enhanced with image post-processing algorithms, particularly deconvolution techniques. These methods aim to revert optical aberrations by deconvolving the image with the point spread function (PSF) of the microscope. However, analytical deconvolution algorithms are computationally demanding, time-consuming, and require precise PSF estimation and careful parameter selection for optimal results. This paper introduces NeuroDecon, a neural network-based method for volumetric deconvolution of confocal images with residual blocks and U-net based architecture. NeuroDecon employs a training strategy that implicitly incorporates the experimental PSF, which acts as a "fingerprint" of system aberrations. This open-source approach allows for personalized training dataset generation, enabling its wide usage for various applications, reduces imaging artifacts and improves computational efficiency. NeuroDecon network outperforms analytical deconvolution methods in image restoration, resolution, and signal-to-noise ratio enhancement and facilitates further data analysis with methods based on automatic segmentation, including protein cluster detection, endoplasmic reticulum network, and dendritic spine 3D-morphology analysis.