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

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 part of obtaining structural feature information of virus. Under noisy conditions, however, manually identifying viral particles in PIMI images is a very time-consuming process with a high error rate. The systematic noise degrading the image resolution and contrast are mainly due to the mechanical or electrical disturbance from the modulation of the illumination when taking raw images. To achieve efficient noise suppressing and accurate virus identification in PIMI images, we developed a neural network-based framework of algorithms. Firstly, a fairly effective denoising method particularly for PIMI imaging was proposed based on a generative network. Both the numerical and experimental results show that the developed method has the best capability of noise removal for PIMI images compared with the traditional denoising algorithms. Secondly, we use a convolutional neural network to detect and recognize viral particles in PIMI images. The experimental results show that viral particles can be identified in PIMI images with high accuracy.

© 2021. This manuscript version is made available under the CC-BY-NC-ND 4.0 license https://creativecommons.org/licenses/by-nc-nd/4.0/