Work

Fine–Coarse Dual Multi-Scale Graph Convolutional Network with Adaptive Edge Construction for Classification and Survival Prediction of Whole-Slide Images

Year: 2026
Type: article
Abstract: Whole-slide images (WSIs) provide important morphological evidence for cancer diagnosis and prognostic assessment. However, their gigapixel size, substantial regional heterogeneity, complex spatial organization, and limited fine-grained annotations make it difficult to model the topological relationships among pathological regions, which in turn limits the use of WSI analysis in automated diagnosis and prognosis. We propose AECGCN, a fine–coarse dual multi-scale graph convolutional network with adaptive edge construction, for WSI classification and survival prediction. Patch-level features are first extracted using the pretrained CONCH (CONtrastive learning from Captions for Histopathology) model and compressed by MiniBatch K-Means clustering to construct fine- and coarse-grained graphs. Graph edges are then constructed adaptively by combining spatial distance with feature-based semantic similarity, where the relative weighting is calibrated by Moran’s I spatial autocorrelation statistic, so that tissue topology and pathological relationships are captured jointly. Finally, two task-specific subnets, AECGCN-C for classification and AECGCN-S for survival prediction, integrate local pathological patterns and global tissue structure through dual-branch graph convolutions and cross-scale attention. Evaluated on datasets from The Cancer Genome Atlas (TCGA), AECGCN achieves classification accuracy (ACC) values of 0.944±0.024 on the combined kidney cohort (TCGA-Kidney) and 0.914±0.027 on the breast invasive carcinoma cohort (TCGA-BRCA). For survival prediction, it achieves concordance index (C-index) values of 0.709±0.041 on the kidney renal clear cell carcinoma cohort (TCGA-KIRC) and 0.644±0.043 on TCGA-BRCA. The study provides a research-stage framework for image-based subtype classification and prognostic risk stratification with exploratory attention visualization. External validation, probability calibration, prospective evaluation, and human-in-the-loop review remain necessary before clinical use. Because the framework relies only on slide-level labels and generic patch-level features, the same fine–coarse dual graph design with adaptive edge construction can be extended to other clinical imaging applications and to other multi-scale or gigapixel image analysis tasks.
Language: English
Open Access status: gold
Estimated APC paid: $2,172
OpenAlex ID: W7236317347
FWCI:
Cites:
Cited by:
Related to: 10