SPATIAL-FREQUENCY FUSION DEBLUR NETWORK FOR SINGLE-SHOT FLUORESCENT NEURON REFOCUSING
Quantitative analyses of labeled neurons are essential for mapping mesoscale connectivity in the macaque brain through neural tracing. However, these analyses are often hindered by out-of-focus blurring during microscopic imaging. In this work, we introduce the Macaque Fluorescent Neuron Defocus (MFND) dataset, comprising 1,800 paired defocused and clear images with neuron count annotations. Building upon this dataset, we propose the Spatial-Frequency Fusion Deblur Network (SFFDNet) for microscopic image refocusing. By incorporating spatial-frequency encoders, wavelet-based pooling, and spatial-frequency attention modules, SFFDNet efficiently recovers high-frequency detail from a single defocused input. Experimental results on the MFND dataset showed that SFFDNet reduces the mean absolute error of neuron counting from 3.36 to 0.81, surpassing several benchmark models.
Macaque Fluorescent Neurons Defocus (MFND) dataset, derived from neuronal tracing experiments on three rhesus macaques. The dataset contains 1,800 pairs of blurred and corresponding sharp images, providing a valuable foundation for data-driven defocus correction methods.
In this project, we have uploaded 360 original-resolution blur-clear image pairs.
If you require access to the full dataset in original resolution, please contact us via email: dongzhenwei2019@ia.ac.cn.