Approximate Bilevel Graph Structure Learning for Histopathology Image Classification
Year: 2026
Type: article
Abstract: The structural and spatial arrangements of cells within tissues represent their functional states, making graph-based learning highly suitable for histopathology image analysis. Existing methods often rely on fixed graphs that may not capture complex tissue interactions. In this work, we propose ABiG-Net (Approximate Bilevel Optimization for Graph Structure Learning via Neural Networks), a novel framework that learns patch interactions within whole slide images (WSI) or large regions of interest (ROI) while simultaneously learning discriminative node embeddings for the downstream classification task. ABiG-Net hierarchically models tissue architecture by constructing local patch-level graphs from cellular organization and learning a global image-level graph that captures sparse, biologically meaningful patch connections through first-order approximate bilevel optimization. Here, ‘‘bilevel" denotes a nested optimization strategy where classifier parameters are optimized at a lower level using training data, while graph adjacency parameters are optimized at an upper level using validation loss. The first-order approximation removes costly second-order hypergradient terms, reducing memory and runtime while keeping image-level graph learning tractable for large WSIs. Clinically, the learned long-range patch interactions link spatially separated but morphologically related regions, producing interpretable, diagnostically relevant region-to-region evidence. Experiments on two histopathology datasets demonstrate its effectiveness: on the Extended CRC dataset, ABiG-Net achieves 97.33±1.15% accuracy for three-class colorectal cancer grading and 98.33±0.58% for binary classification; on the melanoma dataset, it attains 96.27 ± 0.74% for tumor–lymphocyte ROI classification. These results show that first-order approximate bilevel graph structure learning provides an adaptive, interpretable, and tractable alternative for large-scale computational pathology.
Source: IEEE Access
Institutions: Rensselaer Polytechnic Institute, New York University
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W4415274654