Development of a SCGP-CGAN framework for bidirectional H&E–Masson’s Trichrome Virtual Staining

In virtual staining of histopathological sections, deep learning-based methods continue to face challenges arising from the irreversible nature of chemical staining, morphological variations between consecutive tissue sections, and the limited availability of multi-staining paired data from the same tissue region. While existing unsupervised approaches partially alleviate the reliance on paired datasets, they often struggle to maintain consistent color reproduction and preserve fine tissue structures. To address this issue, we develop a task-oriented framework for unsupervised virtual staining, referred to as SCGP-CGAN, where “S,” “C,” “G,” and “P” denote skip connections, CBAM-based attention, WGAN-GP-based adversarial optimization, and perceptual constraints, respectively, while “CGAN” refers to the underlying CycleGAN framework. Here, perceptual constraints refer to the complementary SSIM, LPIPS, style, and color consistency constraints used to promote structural, perceptual, stylistic, and color consistency in the generated images. This framework systematically integrates and evaluates existing components to enhance structural fidelity and perceptual quality in virtual staining. Experimental results on real histopathological datasets indicate that SCGP-CGAN exhibits favorable performance compared with CycleGAN and conventional staining normalization approaches under the evaluated dataset setting, showing improvements in perceptual quality and multiple quantitative evaluation metrics. All experiments are conducted on a publicly available patch-level dataset, and no independent WSI-level or multi-center validation is included. Within this experimental setting, the study demonstrates the effectiveness of the proposed framework for bidirectional H&E–Masson’s trichrome virtual staining and provides a methodological reference for further research in digital pathology image analysis.