Generating bright-field images of stained tissue slices from Mueller matrix polarimetric images with CycleGAN using unpaired dataset

Abstract

Recently, Mueller matrix (MM) polarimetric imaging-assisted pathology detection methods are showing great potential in clinical diagnosis. However, since our human eyes cannot observe polarized light directly, it raises a notable challenge for interpreting the measurement results by pathologists who have limited familiarity with polarization images. One feasible approach is to combine MM polarimetric imaging with virtual staining techniques to generate standardized stained images, inheriting the advantages of information-abundant MM polarimetric imaging. In this study, we develop a model using unpaired MM polarimetric images and bright-field images for generating standard hematoxylin and eosin (H&E) stained tissue images. Compared with the existing polarization virtual staining techniques primarily based on the model training with paired images, the proposed Cycle-Consistent Generative Adversarial Networks (CycleGAN)-based model simplifies data acquisition and data preprocessing to a great extent. The outcomes demonstrate the feasibility of training CycleGAN with unpaired polarization images and their corresponding bright-field images as a viable approach, which provides an intuitive manner for pathologists for future polarization-assisted digital pathology.