VM-GAN: a value mapping framework for virtual staining of optical microscopy images

The advent of virtual staining technology has introduced a rapid and efficient paradigm for optical microscopic imaging in tissue pathology. It enables the utilization of microscopic images of unlabeled samples to generate virtual replicas of chemically stained histological slices or facilitate the transformation of one staining type into another, eliminating the need for time-consuming and labor-intensive chemical staining procedures. Generative models, such as CycleGAN, have demonstrated remarkable capabilities in this context, offering unsupervised learning solutions that bypass the requirement for precisely aligned image pairs in supervised approaches. However, applying these models to large-scale, high-resolution whole-slide images in optical microscopy often requires patch-wise processing, which can introduce boundary artifacts and compromise the continuity of virtual staining. Furthermore, transferring between different staining modalities often necessitates extensive customization of loss functions and hyperparameter tuning for each new domain. To address these issues, we introduce a generalizable virtual staining framework tailored for optical microscopy applications in this work. We propose a loss function based on the value mapping constraint to ensure the accuracy of virtual coloring across diverse staining modalities, termed the value mapping generative adversarial network (VM-GAN). Meanwhile, we design a confidence-based tiling method to mitigate boundary inconsistency inherent to patch-wise processing. Experiments across diverse staining protocols and imaging conditions validate that VM-GAN delivers superior accuracy and visual quality, establishing it as a robust, scalable solution for virtual staining in large-scale, high-resolution optical microscopy.