Spectral Graph Matching for Robust Longitudinal Lesion Correspondence on Foundation-Model Segmentations
Abstract
Purpose
Lesion tracking establishes correspondence across imaging time points to assess disease evolution and treatment response. Despite registration-, graph-, and AI-based methods across CT, MR, PET, and PET/MR, robust correspondence in whole-body CT remains challenging. Although recent large models enable accurate lesion segmentation, longitudinal tracking is limited by anatomical deformation and lesion evolution when correspondence is inferred directly from segmentations. We propose a topology-aware spectral graph matching (SGM) framework operating on segmentations for robust longitudinal lesion tracking.
Methods
A longitudinal whole-body CT cohort of melanoma patients (Longitudinal CT database, autoPET VI) with baseline and post-therapy follow-up scans (n=20; median 4 lesions per patient) was analyzed. Baseline and follow-up scans were registered using Elastix to reduce inter-scan variation. Lesion segmentations and centroids were obtained using LesionLocator (Rokuss et al. 2025). For SGM, lesions were modeled as nodes in a fully connected graph, characterized by centroid, volume, and intra-lesion heterogeneity. Edge weights encoded Gaussian affinities in standardized feature space to capture global spatial and morphological relationships. Spectral embeddings from Laplacian eigen-decomposition preserved disease topology. Lesion correspondence was determined by minimizing a joint cost combining spectral similarity and spatial proximity, solved using Hungarian optimization under spatial constraints. Unmatched lesions were labeled as new or disappeared. Tracking performance was evaluated using Connected Lesion (CL) accuracy (TP/(TP+FP+FN)). SGM performance was evaluated on both physician-annotated and foundation model–generated segmentations using identical correspondence criteria.
Results
LesionLocator achieved a mean Dice score of 0.97 and a normalized surface Dice score of 0.91. The SGM method achieved perfect lesion correspondence (mean CL accuracy = 1.0) on physician-segmented lesions (ground truth) and 0.89 on model-segmented lesions.
Conclusion
This work identifies lesion correspondence as a distinct challenge in longitudinal CT imaging not addressed by segmentation advances. The proposed SGM framework enables robust, segmentation-independent longitudinal correspondence by leveraging global lesion relationships.