Roman Naeem, Signal Processing and Biomedical Engineering

Tubular tree structures such as blood vessels and airways are central to medical imaging from diagnosis to follow-up. A useful representation is the centerline graph, which captures both the medial course and the branching topology of the structure. Classical tracking and segmentation-based pipelines often produce graphs with topological errors such as disconnected components or cycles, along with missing or duplicate branches, reducing downstream usefulness. Recent learning-based image-to-graph methods address some of these issues but remain limited in topology preservation, 3D scalability, and clinical applicability.

This thesis develops recurrent Transformer-based image-to-graph methods that produce topologically valid centerline trees by construction. Trexplorer (Paper A) formulates centerline tracking as recurrent structured prediction, ensuring topological validity through sequential traversal. Trexplorer Super (Paper B) improves robustness through expanded trajectory training and focused higher-resolution features, with broader evaluation across synthetic and real CT data. RefTr (Paper C) advances this line with recurrent refinement of branch trajectories, duplicate suppression, and radius-aware evaluation.

Beyond extraction, centerline graphs also serve as effective inputs for localized clinical prediction. In follow-up after endovascular aneurysm repair, CEVAR (Paper D) uses point-wise centerline embeddings to predict protocol-driven measurements such as vessel diameters and seal lengths, replacing a post-hoc geometric step. The resulting automated pipeline outperforms a commercial semi-automatic workflow on a clinical cohort. Finally, ARTA (Paper E) is a mixed-resolution token allocation method that improves the trade-off between spatial detail and computational cost in dense feature extraction, with direct relevance to the sparse fine-structure analysis central to this thesis.