Deep learning-based virtual histological staining and segmentation in label-free quantitative differential phase contrast microscopy

Abstract

Histological staining remains the diagnostic gold standard in renal pathology, but it requires labor-intensive preparation, specialized infrastructure, and skilled personnel. Isotropic quantitative differential phase contrast (iDPC) microscopy provides label-free morphological contrast by recovering the optical phase of transparent specimens under asymmetric illumination. In this work, we extend label-free virtual histological staining into an end-to-end computational pathology pipeline: iDPC phase maps are first translated into virtually stained bright-field images using an unsupervised image-to-image translation framework without paired training data, and the resulting images are then used for downstream automatic glomerulus segmentation to localize glomerular regions and subsequently classify each segmented glomerulus as normal or abnormal, enabling quantitative, region-focused screening on unstained renal tissue sections.