Patch-Wise Hypergraph Contrastive Learning with Dual Normal Distribution Weighting for Multi-Domain Stain Transfer

TL;DR

This work tackles virtual stain transfer by addressing the loss of high-order semantic information caused by cycle-consistency. It introduces STNHCL, a patch-wise hypergraph contrastive learning framework that models higher-order patch relationships across input and generated stains, combined with a dual normal distribution weighting scheme for negative samples based on discriminator heatmaps. The approach constructs hyperedges via soft k-means and propagates information with hypergraph convolution, enforcing robust patch-level topology alignment. Empirical results on H&E-to-special-stain and H&E-to-IHC tasks show state-of-the-art performance, with strong improvements in structural preservation and stain realism, and favorable downstream performance on glomeruli segmentation tasks. The method offers a practical advancement for accurate, multi-domain virtual staining in histopathology.

Abstract

Virtual stain transfer leverages computer-assisted technology to transform the histochemical staining patterns of tissue samples into other staining types. However, existing methods often lose detailed pathological information due to the limitations of the cycle consistency assumption. To address this challenge, we propose STNHCL, a hypergraph-based patch-wise contrastive learning method. STNHCL captures higher-order relationships among patches through hypergraph modeling, ensuring consistent higher-order topology between input and output images. Additionally, we introduce a novel negative sample weighting strategy that leverages discriminator heatmaps to apply different weights based on the Gaussian distribution for tissue and background, thereby enhancing traditional weighting methods. Experiments demonstrate that STNHCL achieves state-of-the-art performance in the two main categories of stain transfer tasks. Furthermore, our model also performs excellently in downstream tasks. Code will be made available.

Figures (4)

• Figure 1: There are implicit higher-order semantic connections between the input and output images. We model these higher-order semantics using a hypergraph and maximize the mutual information between them.
• Figure 2: Overall framework of the proposed method. (a) Our patch-wise hypergraph contrastive learning framework. We promote cross-patch higher-order topological consistency by maximizing the mutual information between the topological features of the input and output images. (b) A dual normal distribution weighting strategy is applied to negative samples to optimize the learning process, leveraging the distinct features of tissue and background regions.
• Figure 3: Illustration of the relationship between the training weights assigned to negative samples and their similarity to the anchor point ($z_i^\top \cdot v_j$). The MoNCE approach places significant emphasis on hard negative samples, while our proposed method adopts a more balanced approach, incorporating both hard and easy negative sample weighting strategies depending on the context of the tissue and background regions.
• Figure 4: The performance comparison of various existing methods and our proposed method for multiple stain transfer of the same H&E-stained image.