Adaptive Vectorial Restoration from Dynamic Speckle Patterns Through Biological Scattering Media Based on Deep Learning.
Journal:
Sensors (Basel, Switzerland)
PMID:
40292951
Abstract
Imaging technologies based on vector optical fields hold significant potential in the biomedical field, particularly for non-invasive scattering imaging of anisotropic biological tissues. However, the dynamic and anisotropic nature of biological tissues poses severe challenges to the propagation and reconstruction of vector optical fields due to light scattering. To address this, we propose a deep learning-based polarization-resolved restoration method aimed at achieving the efficient and accurate imaging reconstruction from speckle patterns generated after passing through anisotropic and dynamic time-varying biological scattering media. By innovatively leveraging the two orthogonal polarization components of vector optical fields, our approach significantly enhances the robustness of imaging reconstruction in dynamic and anisotropic biological scattering media, benefiting from the additional information dimension of vectorial optical fields and the powerful learning capacity of a deep neural network. For the first time, a hybrid network model is designed that integrates convolutional neural networks (CNN) with a Transformer architecture for capturing local and global features of a speckle image, enabling adaptive vectorial restoration of dynamically time-varying speckle patterns. The experimental results demonstrate that the model exhibits excellent robustness and generalization capabilities in reconstructing the two orthogonal polarization components from dynamic speckle patterns behind anisotropic biological media. This study not only provides an efficient solution for scattering imaging of dynamic anisotropic biological tissues but also advances the application of vector optical fields in dynamic scattering environments through the integration of deep learning and optical technologies.