TMAC-GAN: Topology and Marker-Aware Chromatic Learning for Virtual Stain

Generating virtual immunohistochemistry (IHC) images from standard hematoxylin and eosin (H&E) histopathology may support more efficient pathological assessment. However, this cross-modal image translation task is challenged by spatial misalignment in consecutive tissue sections and visual heterogeneity across biomarkers. To address these issues, we propose the Topology and Marker-Aware Chromatic Generative Adversarial Network (TMAC-GAN) for virtual staining under adjacent-section misalignment. The topology-based bidirectional contrastive learning (TBCL) module reduces dependence on strict pixel-level correspondence by aligning cross-modal representations at the region level through feature graphs. The marker-aware chromatic-textural (MACT) module addresses marker-specific visual variation through 3D color angular and gradient texture constraints that encourage chromatic consistency and subcellular boundary similarity. Evaluations on the BCI and MIST datasets show competitive perceptual and structural consistency compared with representative virtual staining methods.