Trexplorer Super CT Centerline Tracking

• Trexplorer Super is a recurrent model for tracking tree-like tubular structures in CT volumes with topological correctness using graph representations.
• It leverages Super Trajectory Training to maintain memory across patches, Focal Cross Attention for fine local feature fusion, and Target Augmentation for robust bifurcation detection.
• Evaluations on synthetic and real datasets demonstrate superior precision, recall, and zero topological errors, addressing issues like duplicate predictions and premature termination.

Trexplorer Super is a recurrent model for topologically correct centerline tree tracking of tubular objects in CT volumes, introduced as an enhanced version of Trexplorer to address duplicate branch predictions and premature termination during breadth-first tracking of structures such as airways and blood vessels. It reconstructs a centerline tree as a graph $G=(V,E)$ whose nodes are 4D vectors $v=[x,y,z,r]$, encoding 3D position and local radius, and it is evaluated on one synthetic and two real datasets of increasing difficulty. In the reported experiments, Trexplorer Super outperforms previous state-of-the-art models on every dataset while preserving topology, with zero mean absolute error in the Betti-number metrics used to measure connected components and cycles (Naeem et al., 15 Jul 2025).

1. Problem setting and representation

Trexplorer Super addresses centerline tracking for tubular tree structures in 3D medical images. The target anatomy includes bronchial trees and vascular trees, which are branching, elongated networks that are important for diagnosis, treatment planning, and quantitative analysis. In the formulation used by the model, the centerline tree is represented as a graph $G=(V,E)$, where each node $v \in V$ is encoded as $v=[x,y,z,r]$, with $x,y,z$ denoting spatial position and $r$ denoting local radius (Naeem et al., 15 Jul 2025).

Topologically correct tracking is defined by three properties: correct connectivity, no spurious disconnected components, and no cycles where anatomy is acyclic, together with accurate bifurcation structure. These requirements are not only geometric but graph-theoretic. They determine whether the reconstructed centerline can support downstream tasks such as quantification, navigation, and simulation. The paper identifies four central difficulties: branching and long-range dependencies, thin structures requiring high spatial fidelity, noise and weak CT signals in distal segments, and evaluation difficulty caused by the scarcity of public 3D centerline datasets.

The paper’s framing makes clear that the core challenge is not merely local centerline continuation. It is global tree reconstruction under topological constraints, in settings where local ambiguity can induce duplicate tracking of the same branch or early stopping before leaf nodes are reached. This suggests that accuracy at the point level alone is insufficient; the graph structure itself must be preserved.

2. From Trexplorer to Trexplorer Super

Trexplorer, the predecessor model, is described as a DETR-based transformer that performs breadth-first tracking. Starting at the root, it predicts children nodes one level at a time. Each predicted child is classified as end, intermediate, or bifurcation. At a bifurcation, a fixed pool of object queries is allocated to start tracking new branches. Training uses short sub-trajectories and creates new patches at endpoints; a “past trajectory token” encodes up to 10 previous node positions. Object queries carry tracking history within a patch, but this history is lost when moving to the next patch (Naeem et al., 15 Jul 2025).

The original Trexplorer exhibits two reported failure modes. First, duplicate branch predictions arise because past tracking memory is not carried across patches, allowing multiple queries to attach to the same branch in the next patch. Hungarian matching and self-attention help, but the paper states that they are insufficient under ambiguous local features. Second, premature termination occurs because the learned trajectory embedding stored in object queries is lost across patches, reducing confidence in continuing branches and increasing false “end” predictions. The stated consequence is low recall and incomplete trees on real data.

Trexplorer Super is designed specifically to remove these two failure modes while retaining breadth-first traversal. Its tracking logic remains based on predicting children nodes approximately one voxel away from each parent node at the next level. Branches terminate when a node is classified as end, continue when classified as intermediate, and split when classified as bifurcation, at which point a new set of 26 object queries is allocated for newly formed branches and unused queries are discarded. Because expansion is breadth-first and governed by end/bifurcation rules, the paper states that topological correctness is preserved by construction.

3. Core architectural components

Trexplorer Super introduces three named components: Super Trajectory Training (STT), Focal Cross Attention (FCA), and Target Augmentation (TA) (Naeem et al., 15 Jul 2025).

Super Trajectory Training (STT) is introduced to preserve tracking memory across patches. Each training sample forms a “super trajectory” of length 54, divided into six sub-trajectories of length 10 that share link nodes. Each sub-trajectory is paired with a volume patch centered at its start node. During training, object query outputs from sub-trajectory $i$ are fed as inputs to sub-trajectory $i+1$ together with the new patch. The first sub-trajectory still uses the past trajectory token, but subsequent sub-trajectories inherit memory through the chained object queries. Inference mirrors this mechanism: tracking begins at the root and carries forward object query outputs between successive patches. The reported effect is improved recall and reduced premature termination.

Focal Cross Attention (FCA) is designed to combine high-resolution features with controlled computation. The model extracts high-resolution features over a large region but restricts cross-attention to a small focal region around where branches are being tracked. According to the paper, this lets the encoder or feature extractor absorb long-range dependencies and fine spatial patterns into the focal region, while decoder object queries remain focused on tracking history rather than sparse-attention decisions. The intended result is improved localization of thin branches and improved robustness.

Target Augmentation (TA) addresses positional ambiguity at bifurcations. The paper argues that ambiguity scales with radius, so a single bifurcation target is too restrictive and can lead both to missed bifurcations and duplicate tracking. For each primary-branch bifurcation point, an offset is sampled along the primary branch from a Laplace distribution,

$\Delta \sim \mathrm{Laplace}(\mu=0,b),$

with $v=[x,y,z,r]$0, and secondary branches are shifted up or down along the primary centerline to produce augmented targets. Branches around new bifurcation points are smoothed to preserve natural trajectories. Small Gaussian noise is added elsewhere,

$v=[x,y,z,r]$1

The paper states that this enlarges the set of valid bifurcation targets, improves new-branch detection and recall, strengthens Hungarian loss matching and inter-query self-attention, and removes the need for post-processing duplicate cleanup.

Taken together, these components define the enhanced model’s operating logic. STT preserves longitudinal memory, FCA improves local feature selection under large-context requirements, and TA broadens the positive target manifold at branch points. A plausible implication is that Trexplorer Super’s gains arise from coupling memory continuity with more permissive supervision at bifurcations, rather than from any single change in local prediction alone.

4. Datasets and evaluation protocol

To support evaluation, the paper develops three centerline datasets: one synthetic and two real, with increasing difficulty (Naeem et al., 15 Jul 2025).

Dataset	Modality / anatomy	Size
Synthetic dataset (SVT-based)	Synthetic vascular trees	train 368, val 32, test 100
ATM’22	Chest CT, bronchial tree	train 220, val 16, test 60
Parse 2022	Chest CT, pulmonary artery tree	train 72, val 8, test 20

The synthetic dataset uses vessel trees generated with collision avoidance using the Synthetic Vascular Toolkit, with images created following the approach in DeepVesselNet. Its reported graph statistics are max node degree 4, mean points 2763, mean depth 343, mean width 20, and radius $v=[x,y,z,r]$2. ATM’22 is an airway dataset derived from chest CT, resampled to 0.5 mm isotropic resolution, with root points extracted using VMTK and centerlines traced with Kimimaro. Its graph statistics are max node degree 4, mean points 7398, mean depth 504, mean width 55, and radius $v=[x,y,z,r]$3. Parse 2022 is a pulmonary artery dataset, also resampled to 0.5 mm isotropic, with VMTK roots and Kimimaro centerlines; it has max node degree 4, mean points 19644, mean depth 498, mean width 126, and radius $v=[x,y,z,r]$4.

The paper emphasizes the domain gap between synthetic and real data. Synthetic data has cleaner signals and simpler morphology, whereas real datasets, especially Parse 2022, contain denser branching and weaker signals. The authors explicitly note that strong synthetic performance does not necessarily translate to real datasets.

Evaluation is performed at three levels. At the point level, a predicted node is a true positive if there exists a ground-truth node within a radius of 1.5 voxels that has not been matched to another prediction; otherwise it is a false positive, and unmatched ground-truth nodes are false negatives. Precision, recall, and F1 are then computed as usual:

$v=[x,y,z,r]$5

Radius error is measured by

$v=[x,y,z,r]$6

At the branch level, a predicted branch is counted as a true positive if it matches at least 80% of the points of a ground-truth branch within 1.5 voxels and that ground-truth branch has not already been matched. Tree-level topology is evaluated by the MAE of Betti numbers,

$v=[x,y,z,r]$7

where $v=[x,y,z,r]$8 counts connected components and $v=[x,y,z,r]$9 counts cycles.

5. Quantitative performance and ablations

Trexplorer Super is compared against Vesselformer and the original Trexplorer, with results reported as mean $G=(V,E)$0 standard deviation across five runs (Naeem et al., 15 Jul 2025).

On the synthetic dataset, Trexplorer Super achieves point-level precision $G=(V,E)$1, recall $G=(V,E)$2, F1 $G=(V,E)$3, and radius MAE $G=(V,E)$4, which the paper identifies as the best F1 and MAE. Its branch-level F1 is $G=(V,E)$5. On topology metrics, both Trexplorer and Trexplorer Super record $G=(V,E)$6 and $G=(V,E)$7, whereas Vesselformer records $G=(V,E)$8 and $G=(V,E)$9.

On ATM’22, Trexplorer Super reports precision $v \in V$0, recall $v \in V$1, F1 $v \in V$2, and radius MAE $v \in V$3, described as dominant across all metrics. Its branch-level F1 is $v \in V$4. Trexplorer, by contrast, records precision $v \in V$5, recall $v \in V$6, F1 $v \in V$7, and radius MAE $v \in V$8. All Trexplorer variants maintain $v \in V$9 and $v=[x,y,z,r]$0, while Vesselformer records $v=[x,y,z,r]$1 and $v=[x,y,z,r]$2.

On Parse 2022, the most difficult dataset, Trexplorer Super achieves precision $v=[x,y,z,r]$3, recall $v=[x,y,z,r]$4, F1 $v=[x,y,z,r]$5, and radius MAE $v=[x,y,z,r]$6, again the best among the compared models, with branch-level F1 $v=[x,y,z,r]$7. Trexplorer obtains precision $v=[x,y,z,r]$8, recall $v=[x,y,z,r]$9, F1 $x,y,z$0, and radius MAE $x,y,z$1. As on the other datasets, Trexplorer and Trexplorer Super both achieve $x,y,z$2 and $x,y,z$3.

The ablation study on ATM’22 isolates the contributions of the three new modules. The baseline without STT, FCA, or TA reaches precision $x,y,z$4, recall $x,y,z$5, and F1 $x,y,z$6. Adding FCA only gives precision $x,y,z$7, recall $x,y,z$8, and F1 $x,y,z$9. Adding STT only yields precision $r$0, recall $r$1, and F1 $r$2. Adding STT and FCA together yields precision $r$3, recall $r$4, and F1 $r$5. The full model, with STT, FCA, and TA, reaches precision $r$6, recall $r$7, and F1 $r$8.

The authors’ explicit finding is that STT is the most critical component, while FCA and TA further improve completeness and duplicate suppression. This reading is consistent with the earlier diagnosis that loss of query memory across patches is the dominant source of Trexplorer’s premature termination and duplicate branch failures.

6. Implementation details, limitations, and nomenclature

Trexplorer Super is trained on a single node with 4× A100 GPUs and uses mixed precision. For fairness, it shares key hyperparameters with Trexplorer, including 26 tokens allocated at bifurcation and a maximum of 196 concurrently tracked tokens. Trexplorer and Trexplorer Super are each trained for approximately 2 million iterations, while Vesselformer is trained for approximately 12 million iterations with 80 object queries per patch. Inference proceeds by breadth-first tracking from the root, chaining object query outputs across patches as in STT inference and using FCA to focus cross-attention on focal regions; the paper states that duplicates are sufficiently suppressed to avoid post-processing. Runtime and memory footprint are not reported, and the code and datasets are released at the project repository (Naeem et al., 15 Jul 2025).

The paper also identifies several limitations. Some failures occur on a few test samples without obvious input anomalies, which the authors suggest may reflect sensitivity to training set size or feature extractor capacity. Performance on Parse 2022, while better than the baselines, remains substantially lower than on ATM’22 or the synthetic dataset, indicating that denser trees and weaker signals remain challenging. Proposed future directions include stronger pretrained feature extractors, advanced DETR variants or recurrent architectures such as LSTMs, more real training data, and explicit topology-aware regularizers.

The name Trexplorer Super refers specifically to the CT centerline tracking model described above. It is unrelated to Trust Region Entropy (TRE) for exploration in LLM reinforcement learning (Huang et al., 3 Feb 2026) and to TravExplorer, a cross-floor embodied exploration framework for zero-shot object navigation (Zheng et al., 19 May 2026). The similarity in naming can create ambiguity, but the underlying domains, data modalities, and algorithmic objectives are distinct.