Topology-aware Pathological Consistency Matching for Weakly-Paired IHC Virtual Staining

TL;DR

This work tackles the challenge of weakly-paired H&E-to-IHC virtual staining, where spatial misalignment between adjacent tissue sections undermines supervision. It introduces TA-GAN, a topology-aware framework comprising TACM for topology-driven structural alignment and TCPM for pathology-aware region enhancement, implemented via graph neural networks and InfoNCE-based losses. The method leverages adjacency-based patch graphs, topology perturbations, PageRank-based node importance, and correlation matching to enforce both structural and pathological consistency, achieving state-of-the-art results on two breast tissue datasets across four staining tasks. The approach yields superior generation quality in terms of distributional similarity (FID/KID) and stronger pathological consistency (ICC) while maintaining competitive structural fidelity, suggesting significant clinical relevance for reducing reliance on costly IHC staining in practice.

Abstract

Immunohistochemical (IHC) staining provides crucial molecular characterization of tissue samples and plays an indispensable role in the clinical examination and diagnosis of cancers. However, compared with the commonly used Hematoxylin and Eosin (H&E) staining, IHC staining involves complex procedures and is both time-consuming and expensive, which limits its widespread clinical use. Virtual staining converts H&E images to IHC images, offering a cost-effective alternative to clinical IHC staining. Nevertheless, using adjacent slides as ground truth often results in weakly-paired data with spatial misalignment and local deformations, hindering effective supervised learning. To address these challenges, we propose a novel topology-aware framework for H&E-to-IHC virtual staining. Specifically, we introduce a Topology-aware Consistency Matching (TACM) mechanism that employs graph contrastive learning and topological perturbations to learn robust matching patterns despite spatial misalignments, ensuring structural consistency. Furthermore, we propose a Topology-constrained Pathological Matching (TCPM) mechanism that aligns pathological positive regions based on node importance to enhance pathological consistency. Extensive experiments on two benchmarks across four staining tasks demonstrate that our method outperforms state-of-the-art approaches, achieving superior generation quality with higher clinical relevance.

Figures (7)

• Figure 1: Spatial misalignment or local deformations between adjacent tissue sections can lead to misleading supervision signals. Previous methods, such as CUTpark2020contrastive, rely only on local semantic correspondence, resulting in inconsistent staining of regions with identical structures. In contrast, our method incorporates spatial topological matching to ensure consistent and coherent staining across corresponding regions.
• Figure 2: (a) The overall architecture of TA-GAN. Following the general GAN paradigm, a generator takes an H&E image as input and outputs an IHC stained image. The encoder of the generator is used to extract features for the H&E image, the generated virtual IHC image, and the real IHC image from adjacent slices. (b) The adjacency matrix is computed via feature cosine similarity. To ensure structural consistency, two GNN branches are employed: one aggregates information using their respective topologies, while the other applies H&E topology with perturbations to the virtual IHC graph. Finally, a mutual information loss is calculated to align both branches. (c) Pathology Matching based on node importance. We compute node importance scores $p_{\text{gen}}$ and $p_{\text{real}}$ for the virtual and real IHC images, respectively, based on their connectivity relationships. These scores are then used to enhance the node features. Subsequently, we obtain correlation matrics $C^{\text{VIHC}}$ and $C^{\text{IHC}}$ via cosine dot product. Finally, $C^{\text{VIHC}}$ and $C^{\text{IHC}}$ are used to compute the correlation matching loss $\mathcal{L_{\text{cm}}}$.
• Figure 3: The quantitative comparison of different state-of-the-art methods on the MIST dataset. The first column is the H&E-stained images and the last columns is the adjacent slices real IHC images. Columns 2 to 11 are the images virtually stained by different methods.
• Figure 4: The Intraclass Correlation Coefficient (ICC) between the positive area ratios of the virtually stained IHC images and the real IHC images from adjacent tissue sections.
• Figure 5: The trend equation derived from the scatter plot of positive area ratios uses the slope and intercept as its horizontal and vertical axes, respectively. The closer the position is to the ideal point (1, 0) (i.e., the line $y = x$), the stronger the agreement with the real IHC images.