Abstract
In vivo low-light fluorescence microscopy is a core technology for resolving neural circuit function and neurotransmitter transmission mechanisms in neuroscience research. However, it faces critical challenges including motion artifact removal in low signal-to-noise ratio (SNR) scenarios, preservation of weak neural fluorescent signals, baseline drift interference with quantification accuracy, and the trade-off between spatial resolution enhancement and noise suppression. Existing methods cannot simultaneously maintain weak signal fidelity and full-process automated processing capabilities. This paper presents a comprehensive neural signal capture and quantification algorithm for in vivo low-light fluorescence microscopy. The proposed method employs phase-correlation registration for sub-pixel motion correction, regularized iterative deconvolution with calibrated double-exponential Gaussian point spread functions to balance resolution enhancement and noise control, and dual-baseline region dynamic selection with pixel-wise double-exponential fitting for temporal baseline stabilization. The core innovation is a biologically-constrained deep learning fusion module that combines convolutional neural networks, sparse representation, and bidirectional gated recurrent networks in a hybrid architecture. Through self-supervised learning constrained by neural signal response priors, the algorithm achieves precise separation of valid signals from background noise. Experimental results demonstrate that the proposed algorithm completely preserves neurotransmitter signal kinetic characteristics in mouse in vivo imaging scenarios, achieves adaptive axial drift compensation, and enables accurate DeltaF/F signal quantification, providing a robust full-process processing solution for low-light long-term in vivo neural imaging.