VITG-VS: Preserving Spatial Heterogeneity in Virtual Staining of Renal Autofluorescence Images

Label-free imaging technique integrated with virtual staining has rapidly advanced in the field of renal pathology. However, most existing convolutional neural network (CNN) based virtual staining methods encounter difficulties in preserving spatial heterogeneity, especially in images characterized by complex tissue variability. To achieve better results in auto-fluorescence virtual staining tasks, we introduce a novel transformer-based generative adversarial network named ViTG-VS. This method integrates the global perceptual advantages of the transformer architecture with convolutional operations that capture local information. Furthermore, we incorporate a content-aware positional encoding to capture the spatial variations of morphological information distribution. Quantitative and qualitative results indicate that ViTG-VS performs well in the virtual staining of renal pathology. The spatially heterogeneous structures of tissue captured in autofluorescence images can be accurately transformed into their PAS-stained counterparts. We believe that ViTG-VS of-fers a new perspective for addressing the challenges of virtual staining in heterogeneous autofluorescence images.