RefTr: Recurrent Refinement of Confluent Trajectories for 3D Vascular Tree Centerline Graphs

TL;DR

RefTr introduces a recurrently refined, confluent-trajectory representation for 3D vascular centerline graphs, addressing the need for high recall while preserving topological validity. The model uses a Producer–Refiner Transformer architecture with memory-augmented decoding, enabling many-to-one trajectory assignments and a Tree Non-Max Suppression step to reduce duplicates. Radius-aware evaluation thresholds are proposed to account for anatomical variability, and experiments across synthetic, ATM'22, and Parse 2022 datasets show improved recall with competitive precision and reduced parameter count. The approach offers a practical, efficient framework for vascular tree analysis in 3D medical imaging and sets a new state-of-the-art in centerline graph generation.

Abstract

Tubular trees, such as blood vessels and lung airways, are essential for material transport within the human body. Accurately detecting their centerlines with correct tree topology is critical for clinical tasks such as diagnosis, treatment planning, and surgical navigation. In these applications, maintaining high recall is crucial, as missing small branches can result in fatal mistakes caused by incomplete assessments or undetected abnormalities. We present RefTr, a 3D image-to-graph model for centerline generation of vascular trees via recurrent refinement of confluent trajectories. RefTr uses a Producer-Refiner architecture based on a Transformer decoder, where the Producer proposes a set of initial confluent trajectories that are recurrently refined by the Refiner to produce final trajectories, which forms the centerline graph. The confluent trajectory representation enables refinement of complete trajectories while explicitly enforcing a valid tree topology. The recurrent refinement scheme improves precision and reuses the same Refiner block across multiple steps, yielding a 2.4x reduction in decoder parameters compared to previous SOTA. We also introduce an efficient non-maximum suppression algorithm for spatial tree graphs to merge duplicate branches and boost precision. Across multiple public centerline datasets, RefTr achieves superior recall and comparable precision to previous SOTA, while offering faster inference and substantially fewer parameters, demonstrating its potential as a new state-of-the-art framework for vascular tree analysis in 3D medical imaging.

Figures (11)

• Figure 1: Illustration of confluent trajectory representation: two trajectories are confluent for the initial five points before diverging, modeling the branching structure of vessel centerlines.
• Figure 2: Overview of the RefTr architecture. The top illustrates the model, which includes an Image Encoder, the Producer, the Refiner, and prediction heads. The Producer uses learned token sets and image features to propose multiple trajectory embeddings, which are refined by the Refiner over multiple steps. Loss is computed at every step, including the Producer step (omitted for clarity). Both the Producer and Refiner also receive input from a memory bank containing refined token sets from previous image patches (also omitted for clarity). The bottom shows recurrent refinement of multiple confluent trajectories toward two target trajectories over $S$ steps.
• Figure 3: Detailed structure of the Producer and Refiner blocks.
• Figure 4: Visual comparison of ground truth and predictions from Vesselformer, Trexplorer, Trexplorer Super, and RefTr on one sample from each dataset. Centerline marker sizes are scaled according to the radius at each node. The vessel mask (green) is shown only for visualization and is not used during training or evaluation.
• Figure 5: Point-level performance metrics versus the number of refinement steps for ATM'22 Dataset.