Spectral-augmented learning for viable bio-speckle reconstruction

Speckle correlation imaging (SCI) shows extensive application potential in imaging through biological media, and deep learning methods have achieved high performance in imaging scenarios through homogeneous media. However, the insufficient volume of real biological speckle data, coupled with the non-ideal intensity noise inherent in these speckles, has impeded the practical implementation and further advancement of SCI technology. To address these challenges, a dataset of speckles from viable biological media was constructed and made publicly available. The properties of biological speckles were characterized, and their autocorrelation constraints as well as the characteristics of the Fourier magnitude spectrum were modeled and validated. Leveraging a dual-path architecture with a dedicated frequency-domain input branch, the spatial-frequency hybrid reconstruction neural network (SFHRnet) was proposed. Benefiting from its rational architecture, it exhibits superior performance compared to previous approaches in reconstruction accuracy, notably in recovering complex vessel targets. This method is expected to advance the practical biological applications of speckle imaging and provide valuable references for scattering imaging scenarios.