IHC-DualNet: a dual-branch graph-based architecture for interpretable and precise immunohistochemistry tissue segmentation

Objective. Immunohistochemistry (IHC) is a cornerstone technique in oncology, where accurate tissue region segmentation is critical for diagnosis and prognosis. However, current clinical workflows rely heavily on manual annotation, which is time-consuming, subjective, and poorly reproducible. Existing automated methods often struggle to simultaneously capture global tissue context and local structural details, resulting in fragmented boundaries and unreliable segmentation of differentially expressed regions. Furthermore, their "black-box" nature lacks interpretability, failing to align with the hierarchical reasoning processes of pathologists. Approach. To address these challenges, we propose IHC-DualNet, a novel dual-branch graph-based framework that emulates pathological reasoning. The Global Transformer Reasoning (GTR) branch leverages a newly designed IHC-GTR layer to model comprehensive tissue-wide context, while the Local Graph Attention branch focuses on capturing fine-grained local features. An Adaptive Feature Fusion (AdaFuse) module dynamically integrates these complementary representations, followed by a decoder that generates high-precision segmentation maps. Main results. IHC-DualNet demonstrates superior accuracy and robustness across three challenging IHC datasets—breast cancer (HER2), esophageal cancer (PCNA), and lung cancer (IL-24)—achieving, for example, 93.88% Dice, 90.26% IoU, 92.17% Recall, and 95.59% Accuracy on the breast cancer dataset. Significance. IHC-DualNet advances the state of the art in automated IHC analysis by providing a more accurate, robust, and interpretable segmentation framework, thereby enabling more reliable tumor grading and patient stratification.