G2*-Net learns optimal tissue graphs for cancer diagnosis from histopathology images

Histopathology images contain rich spatial organization cues crucial for cancer diagnosis, but traditional graph-based methods rely on fixed or heuristic graph structures that may misrepresent true tissue connectivity. To address this, researchers from Rensselaer Polytechnic Institute and NYU Grossman School of Medicine introduce G2*-Net – a hierarchical framework that learns the graph structure end-to-end. The method first divides whole-slide images (WSIs) or large regions of interest into patches, constructs cell-level graphs within each patch to capture local tissue architecture, and then models each patch as a node in a learnable image-level graph. The key innovation is formulating image-level graph structure learning as a second-order bilevel optimization problem: the outer loop learns the graph connectivity while the inner loop optimizes the classifier, coupled through validation-driven feedback. To make this computationally tractable, the authors adopt a DARTS-inspired one-step unrolled approximation for efficient hypergradient estimation – enabling practical training on large-scale WSIs.

Experimental validation on three distinct histopathology datasets (including lung, breast, and colon cancer) demonstrates that G2*-Net consistently outperforms methods using fixed or handcrafted graphs, achieving higher classification accuracy and better generalization. The learned graphs reveal meaningful tissue structures that align with known histological patterns, providing interpretable insights for pathologists. This work, accepted at ICMLA 2026, represents a step toward more reliable AI-assisted pathology by eliminating the guesswork in graph construction. For practitioners, it means diagnostic models that can adapt to different tissue types and staining protocols without manual graph engineering – potentially reducing false negatives in cancer screening and enabling automated analysis of large biopsy cohorts.