Graph-radiomic learning (GrRAiL) descriptor to characterize imaging heterogeneity in confounding tumor pathologies

Abstract

Abstract Distinguishing benign confounding pathologies from malignant neoplasms on routine imaging remains a significant challenge in solid tumors. Radiomics approaches attempt to capture lesion heterogeneity on clinical imaging (CT,MRI), but typically aggregate feature values across the region of interest (ROI), losing the complex spatial relationships among intensity compositions within the lesion. We present Graph-Radiomic Learning (GrRAiL), a new descriptor for characterizing intralesional heterogeneity (ILH) on clinical MRI. GrRAiL (1) identifies clusters of sub-regions from per-voxel radiomic measurements and (2) computes graph-theoretic measurements that quantify the spatial associations of these radiomic clusters. The resulting weighted graphs capture higher-order spatial relationships within the ROI to reliably distinguish confounding pathologies from malignancies. GrRAiL was evaluated on n = 947 subjects across three multi-institutional use-cases: distinguishing radiation effects from tumor recurrence in glioblastoma ( n = 106) and brain metastasis ( n = 233), and differentiating no+low- versus high-risk pancreatic intraductal papillary mucinous neoplasms (IPMNs) ( n = 608). GrRAiL consistently outperformed Graph Neural Networks, textural radiomics, and intensity graph analysis, achieving cross-validation and test accuracies of 89% and 78% (glioblastoma), 84% and 74% (metastasis), and 85% and 79% (IPMN), representing >10%, >13%, and >10% test-accuracy improvements over comparative approaches, respectively.