180 / 2024-06-14 22:30:30

Deep learning-based De-scattering in Two-photon Fluorescence Microscopy

Deep learning, descattering, two-photon-excited fluorescence microscopy

Two-photon-excited fluorescence (TPEF) microscopy excels in in vivo deep tissue imaging, making it a vital tool for exploring biological tissues. However, deep layer imaging faces challenges from biological scattering, impacting image quality. To address this, we develop a deep learning-based de-scattering method for two-photon excitation fluorescence imaging (DeS-TPEF). This method employs a multi-attention model that focuses on the reconstruction target. It also utilizes a multi-component optimization strategy, guided by the minimum-cross-entropy threshold segmentation with Dice similarity coefficient (MCE_DSC) loss function, to minimize the false positives in the reconstruction process. To train a network capable of de-scattering images affected by real-world high scattering, we designed a simulated scattering model. This model can degrade images from conditions of shallow depth, no/low scattering, and high signal-to-noise ratio to those representing deep depth, high scattering, and low signal-to-noise ratio. Our quantitative validation involved static scattering fluorescent beads and vascular systems under biological dynamic scattering. The results showed significant improvements in NRMSE, PSNR, and SSIM compared to the original data. We demonstrate our method in real-world experimental imaging studies, includinging in vivo imaging of the cerebrovascular system within mice at depths of up to 850 μm.