PASB: Pathology-aware Schrödinger bridge for virtual immunohistochemical staining.

Abstract

Virtual immunohistochemistry (IHC) staining automatically translates Hematoxylin and Eosin (H&E) images into IHC images using deep generative models, enabling automation of IHC staining. Weakly supervised methods for virtual IHC staining leverage guidance signals from adjacent tissue sections, eliminating the need for precise alignment, and becoming one of the mainstream paradigms. However, these methods still face two main challenges: (i) they often fail to extract clinically meaningful pathological semantics, relying on low-level features that cannot guarantee pathological consistency; (ii) existing generative frameworks struggle to capture the complex and heterogeneous nature of histopathological data, leading to problems like mode collapse and the loss of diagnostically critical structures. To address these challenges, we propose a new weakly supervised method, namely Pathology-Aware Schrödinger Bridge (PASB). Specifically, we employ the Schrödinger Bridge as a generative backbone to avoid fixed prior assumptions as well as preserving generative diversity and reducing mode collapse. Additionally, we investigate the Constraint-Driven Alignment Learning (CDAL) module to provide high-level semantic supervision and design a Similarity-based Dynamic Path Refinement (SDPR) module to enhance pathological consistency during the generation process. Extensive experiments demonstrate that the proposed PASB outperforms existing methods in terms of both generative quality and pathological consistency, producing IHC images with clinical potential comparable to real IHC in downstream diagnostic tasks.