PRompt-aware Interactive Stain Model for Unpaired H&E-to-IHC Stain Translation

Virtual staining technology has emerged as a promising solution for generating immunohistochemistry (IHC) images from hematoxylin and eosin (H&E) stained samples, addressing the resource-intensive nature of traditional IHC staining. In this paper, we present PRompt-aware Interactive Stain Model (PRISM), a novel framework for unpaired H&E-to-IHC stain translation. Our approach introduces two key innovations: the Style-Prompt Integrated Residual Image Translator Generator (SPIRIT-G), which employs a Prompt-Aware Fusion Layer to effectively control stain generation, and the Dual Auxiliary Classifier Attention Discriminator (DA-CAD), which incorporates an auxiliary classifier for stain type identification. Unlike conventional methods that use one-hot encoding for task information, PRISM leverages learned priors to capture intricate relationships among tasks and modalities. Comprehensive experiments conducted on two public datasets show that our approach delivers leading results across five distinct translation tasks, markedly enhancing both structure preservation and staining pattern fidelity. These results validate PRISM's effectiveness as a universal translation model for histopathological image analysis.