Signal denoising of viral particle in wide-field photon scattering parametric images using deep learning

Abstract Polarization parametric indirect microscopic imaging (PIMI) can obtain anisotropic nanoscale structural information of the sample by utilizing a polarization modulated illumination scheme and fitting the far-field variation of polarization states of the scattered photons. The rich scattering information of PIMI images can be exploited for identification of viral particles, aiming for early infection screening of viruses. Accurate processing and analysis of PIMI results is an important step for obtaining structural feature information of virus. However, the systematic noise, mainly caused by the mechanical or electrical disturbance from the modulation of the illumination when taking raw images, fairly degrades the image resolution and contrast, making the analysis of results more time-consuming with higher error rate. To achieve efficient noise suppressing in the obtained PIMI images, we developed a neural network-based framework of algorithms. A fairly effective denoising method particularly for PIMI imaging was proposed based on a U-Net. From both the numerical and experimental results, the developed method significantly improves the quality of PIMI images with the best capability of noise removal compared with the traditional denoising algorithms and other neural network architecture. This method can be employed as a fixed preprocessing procedure of raw PIMI images, which is greatly helpful to realize rapid, accurate and programmed analysis of results in PIMI applications.