Accurate lesion tracking across longitudinal CT scans is essential for evaluating treatment response and disease progression in oncology. Yet, matching lesions between consecutive examinations remains largely manual, time-consuming and subject to variability.

This study introduces GraphTrack, a two-stage deep learning framework designed to automate lesion correspondence by combining local lesion similarity with global spatial, temporal and morphological information.

GraphTrack first uses a self-supervised 3D Masked Autoencoder (3DMAE) to learn transferable lesion representations, which are refined through a 3D Siamese Neural Network (3DSNN) for pairwise lesion similarity. These features are then integrated into an Edge Graph Attention Network (EdgeGAT), allowing the model to reason over relationships between multiple lesions and timepoints rather than evaluating candidate matches independently.

Evaluated on a large multicenter thoraco-abdominal CT dataset, GraphTrack achieved 88.9% precision, 80.5% recall and 97.8% AUC-ROC. On the independent DeepLesion dataset, it maintained strong performance, with 90.2% precision, 97.7% recall and 89.6% AUC-ROC.

By unifying self-supervised learning, pairwise similarity and graph-based reasoning, GraphTrack provides a scalable approach to automated longitudinal lesion tracking, with potential to streamline oncology imaging workflows and support more consistent assessment of treatment response and disease evolution.