Spatio-Temporal Lesion Graphs: A Graph-Based Framework for Tracking Topology Changes in Longitudinal Oncologic Imaging

Abstract

Longitudinal oncologic imaging poses a tracking challenge because lesions can split, merge, or transiently disappear, violating one-to-one correspondence assumptions. We propose spatio-temporal lesion graphs constructed via nearest-centroid matching across all previous time points for lineage assignment, with bidirectional matching between consecutive scans to classify splits, merges, and continuations without learning. Modality-agnostic matching naturally captures phenotypic switching across imaging different modalities. Evaluated on longitudinal PSMA-/FDG-PET/CT during 177Lu-PSMA therapy (4 patients, 11-13 scans each), the graphs recovered 50 topology changes (22 splits, 28 merges) missed by forward-only matching and revealed persistently high intra-patient SUVpeak heterogeneity (median CV=0.74), demonstrating that aggregate metrics obscure substantial inter-lesion variation. The resulting graph provides a compact lesion-level representation for response assessment and potentially lesion-specific diagnostics and local therapy. The code and data are available at: https://github.com/lukasf98/spatio-temporal-graph.