PartVessel: Vessel Synthesis & Partial Annotation

Updated 5 July 2026

PartVessel is a term covering both hierarchical generative modeling for 3D vessel synthesis and streamlined annotation for coronary artery segmentation.
It separates global tree-like topology from local segment geometry to capture anatomical realism and improve computational efficiency.
By decomposing vessel complexity into meaningful parts, PartVessel advances diagnostic analysis and weakly supervised deep learning methods.

PartVessel is used in recent vessel research in two distinct senses. In one sense, it denotes a hierarchical part-based generative model for realistic 3D blood-vessel synthesis that separates global binary-tree topology from local segment geometry and then assembles the vessel hierarchically (Chen et al., 21 Jul 2025). In another sense, it denotes partial vessels annotation, or PVA, a clinically motivated annotation strategy for coronary artery segmentation in CCTA in which only contiguous vessel portions are labeled and a progressive weakly supervised pipeline reconstructs the full coronary tree (Zhang et al., 2023). The shared term therefore refers not to a single universal formalism, but to two structurally related efforts to reduce vessel complexity by decomposing it into meaningful parts.

1. Terminological scope and disambiguation

In current usage, “PartVessel” refers to two separate research constructs.

Usage	Domain	Core idea
PartVessel (Chen et al., 21 Jul 2025)	3D blood-vessel generation	Generate a vessel from global topology, local segments, and hierarchical assembly
PartVessel / PVA (Zhang et al., 2023)	Coronary artery segmentation in CCTA	Annotate contiguous vessel portions and recover the full tree with weak supervision

The first usage is explicitly a generative model. It addresses 3D blood vessel generation for diagnostic assessment, treatment planning, preoperative simulation, medical image analysis, and synthetic vascular data generation. Its core premise is that vessels should not be generated as monolithic objects, because their tree-like branching topology and local tubular geometry are qualitatively different structures (Chen et al., 21 Jul 2025).

The second usage is explicitly an annotation and learning paradigm. It arises from the cost and expertise required for full coronary artery annotation in CCTA, and from the observation that clinicians prioritize main trunks more than exhaustive distal labeling. PVA therefore labels vessels continuously from proximal to distal ends rather than marking random disconnected regions, preserving local topology and continuity while reducing annotation burden (Zhang et al., 2023).

This dual usage is important because it prevents a common category error: PartVessel is not exclusively a generative model name, nor exclusively a weak-supervision label format. The term spans both vessel synthesis and vessel segmentation, with “parts” meaning local vessel segments in the former and partially annotated contiguous vessel paths in the latter.

2. PartVessel as a hierarchical part-based generative model

PartVessel, in the generative sense, is a hierarchical model for 3D blood vessel generation that decomposes vascular structure into two levels: global topology and local geometry. The overall vessel is represented as a tree-like hierarchical structure derived from a skeleton-based binary-tree key graph, while each edge of that key graph corresponds to a distinct vessel segment represented as an ordered 3D sequence of points $\mathbf x=[x,y,z,r]$ , where $r$ is vessel radius (Chen et al., 21 Jul 2025).

At the global level, each key-graph node is associated with 3D spatial coordinates $[x,y,z]$ , direction $[n_x,n_y,n_z]$ , and a geometric descriptor $C=[\ell, \delta, \kappa, \rho]$ , where $\ell$ is segment length, $\delta$ is straight-line distance between endpoints, $\kappa$ is curvature, and $\rho$ is depth in the binary tree. This representation makes topology explicit and keeps local segment generation conditionable by geometric context. The model thus treats vascular anatomy as a composition of reusable tubular parts governed by a tree scaffold rather than as an undifferentiated point cloud or monolithic mesh (Chen et al., 21 Jul 2025).

The model is motivated by limitations of three prior families of methods. Point-cloud generation methods are reported to struggle with tubular continuity, interior-versus-exterior distinction, and hole-free meshing. Implicit or tree diffusion methods can model tree-like structure but are described as limited in fidelity for complex vascular geometries and can produce irregular, block-like, or disconnected outputs qualitatively. Recursive vessel VAE methods such as VesselVAE generate vessels from skeletal graphs, but are characterized as generating the vessel more monolithically and showing degraded performance when bifurcation complexity increases (Chen et al., 21 Jul 2025, Feldman et al., 2023).

The significance of this representation is methodological rather than merely architectural. By making topology a first-class object and relegating geometry to segment generation, PartVessel encodes a strong anatomical prior: vessels are branching systems composed of locally smooth tubular units whose arrangement is constrained by a tree. This suggests a decomposition strategy more aligned with vascular structure than generic 3D object synthesis.

3. Three-stage generation pipeline and training objectives

The generative PartVessel pipeline proceeds in three stages: key graph generation, vessel segment generation, and hierarchical vessel assembly (Chen et al., 21 Jul 2025).

In the first stage, key graph generation models the global binary tree-like vessel structure using a Recursive Variational Autoencoder. Encoding proceeds bottom-up. If $v_{parent}$ denotes a parent node attribute and $r$ 0 denote hidden states of the left and right children, the parent hidden state is computed as

$r$ 1

Recursive merging continues until the root node $r$ 2, which serves as the global latent embedding of the vessel key graph. Decoding proceeds top-down: child existence is predicted by

$r$ 3

and child attributes are reconstructed by side-specific multilayer perceptrons. The stage-1 loss combines node-attribute reconstruction, node-existence classification, and KL regularization:

$r$ 4

In the second stage, vessel segment generation models local branch geometry using a Transformer-based variational framework. Each segment is an ordered sequence $r$ 5 conditioned on the descriptor

$r$ 6

Here the Transformer treats points along the segment as tokens and generates segment geometry conditioned on length, endpoint distance, curvature, and tree depth. The stage-2 loss is given as

$r$ 7

The formulation couples geometric reconstruction with conditional latent regularization.

In the third stage, hierarchical vessel assembly integrates the outputs of the previous stages. A latent sample from stage 1 is decoded into a key graph, generated segments are selected and placed according to that graph, and depth-first search traversal is used to assemble the vessel skeleton. Each segment is aligned by scaling, translation, and rotation so that its orientation matches the key-graph direction $r$ 8. After skeleton assembly, predicted segment radii are used to reconstruct the final 3D vessel surface (Chen et al., 21 Jul 2025).

This pipeline is explicitly part-based in a strong sense. It does not merely segment a latent vector into subcodes; it factorizes vascular synthesis into structure generation, geometry generation, and assembly. That separation is the defining trait of the method.

4. Data, evaluation, and reported performance of the generative model

The generative PartVessel model is evaluated on three public datasets with a 90% training and 10% testing split: ImageCAS, consisting of 1,000 real 3D CCTA scans of coronary arteries; VascuSynth, consisting of 120 synthetic 3D vascular trees; and Processed CoW, consisting of 300 3D vascular meshes of intracranial arteries (Chen et al., 21 Jul 2025).

Preprocessing constructs the key graph from binary vessel labels by applying morphological operations to obtain skeleton and surface, using adaptive mapping to construct a key graph, building a maximum spanning tree, and identifying root, leaf, and branch nodes as graph vertices. The model is implemented in PyTorch and trained on an NVIDIA A800. Stage 1 uses latent dimension 512, learning rate 0.001, reduction by 0.2 every 100 epochs, 20k epochs, and batch size 128; stage 2 uses latent dimension 64, learning rate 0.0002, 2k epochs, and batch size 512. The Transformer has 4 layers, 4 multi-head self-attention heads, and maximum sequence length 200 (Chen et al., 21 Jul 2025).

Evaluation combines point-based and graph-based metrics. Point metrics are Jensen-Shannon Divergence and Chamfer Distance on sampled point clouds. Graph metrics are degree-distribution maximum mean discrepancy, Laplacian spectrum discrepancy, and Graph Wasserstein Distance. The paper emphasizes that graph metrics are necessary because point metrics can miss topology errors (Chen et al., 21 Jul 2025).

The reported results for PartVessel are as follows.

Dataset	JSD	CD	Deg.	Spec.	GWD
ImageCAS	50.1	24.4	0.601	0.079	0.029
VascuSynth	40.1	34.8	0.190	0.159	0.049
CoW	44.1	15.1	1.475	0.182	0.017

On these datasets, the model is reported to outperform PointDiffusion, TreeDiffusion / TrIND, and VesselVAE in structural fidelity and visual realism. Qualitatively, it is described as producing more robust and anatomically consistent reconstructions on ImageCAS, preserving branch connectivity, curvature, local morphology, and full-tree consistency, while avoiding the holes, block-like artifacts, and disconnected components attributed to competing approaches (Chen et al., 21 Jul 2025).

The reported limitations are also structural. Topology metrics are not uniformly best across all graph measures, especially on CoW; the method depends on preprocessing quality in skeleton and key-graph extraction; the representation is explicitly binary-tree-like; the segment generator imposes a maximum sequence length of 200; and broader generalization beyond coronary and cerebral vessels remains to be established (Chen et al., 21 Jul 2025). These constraints indicate that the model’s strength comes from a specific anatomical prior rather than unrestricted shape expressivity.

5. PartVessel as partial vessels annotation in coronary artery segmentation

In coronary CCTA segmentation, PartVessel denotes partial vessels annotation, a label-efficient strategy in which vessels are labeled continuously from proximal to distal ends rather than through random disconnected partial regions (Zhang et al., 2023). The rationale is clinical and operational: coronary arteries are small, thin, branching, and hard to delineate slice by slice; annotation is expensive and expert-dependent; and clinicians care more about trunk continuity than exhaustive distal-branch annotation.

The method built on PVA has two stages. The first stage, Local Feature Extraction, trains a lightweight segmentation model $r$ 9 on small patches centered around labeled regions so that vessel appearance learned from annotated fragments can be propagated to unlabeled regions. This produces whole-volume logits $[x,y,z]$ 0, which are fused with the original PVA labels $[x,y,z]$ 1 to form the initial pseudo-label volume:

$[x,y,z]$ 2

The second stage, Global Structure Reconstruction, trains a second model $[x,y,z]$ 3 on pseudo labels to recover the global coronary tree structure. Pseudo labels are updated iteratively through an Adaptive Label Propagation Unit. The confidence degree at iteration $[x,y,z]$ 4 is

$[x,y,z]$ 5

where $[x,y,z]$ 6 is the current global-model output. Pseudo labels are updated only if the current confidence is the best so far:

$[x,y,z]$ 7

and then revised through an EWMA rule,

$[x,y,z]$ 8

This mechanism is intended to improve pseudo labels while preventing degradation from low-quality updates (Zhang et al., 2023).

A further refinement is the Feature Prototype Analysis block, used at test time to improve continuity, especially in major trunks. For class $[x,y,z]$ 9, the prototype is

$[n_x,n_y,n_z]$ 0

and from penultimate-layer feature map $[n_x,n_y,n_z]$ 1, the similarity map is

$[n_x,n_y,n_z]$ 2

The FPA module is trained while freezing the rest of the network, using only labeled voxels:

$[n_x,n_y,n_z]$ 3

The stated role of FPA is to enhance final predictions for vessel continuity, trunk completeness, and robustness of long thin structures (Zhang et al., 2023).

This framework uses a 3D U-Net backbone in PyTorch, Adam optimizer with initial learning rate $[n_x,n_y,n_z]$ 4, local patch size $[n_x,n_y,n_z]$ 5 for $[n_x,n_y,n_z]$ 6, and global patch size $[n_x,n_y,n_z]$ 7 for $[n_x,n_y,n_z]$ 8. It is trained on 108 clinical CCTA subjects split 2:1 into train and test, with only 24.29% of vessels labeled in training under PVA (Zhang et al., 2023).

6. Clinical results, misconceptions, and relation to adjacent vessel research

Under PVA with only 24.29% labeled vessels, the coronary segmentation framework reports Dice $[n_x,n_y,n_z]$ 9, RDice $C=[\ell, \delta, \kappa, \rho]$ 0, OV $C=[\ell, \delta, \kappa, \rho]$ 1, and OF values of $C=[\ell, \delta, \kappa, \rho]$ 2 for LAD, $C=[\ell, \delta, \kappa, \rho]$ 3 for LCX, and $C=[\ell, \delta, \kappa, \rho]$ 4 for RCA (Zhang et al., 2023). The fully supervised 3D U-Net baseline with 100% annotation reports Dice $C=[\ell, \delta, \kappa, \rho]$ 5, RDice $C=[\ell, \delta, \kappa, \rho]$ 6, OV $C=[\ell, \delta, \kappa, \rho]$ 7, and OF values $C=[\ell, \delta, \kappa, \rho]$ 8, $C=[\ell, \delta, \kappa, \rho]$ 9, and $\ell$ 0 for LAD, LCX, and RCA respectively. The principal claim is therefore not full parity in volumetric overlap, but comparable trunk continuity under substantial label reduction (Zhang et al., 2023).

The ablation study clarifies the contribution of each component. Using only $\ell$ 1 yields Dice 60.60, RDice 69.45, and OV 62.24. Adding pseudo-label initialization raises these to 64.23, 73.81, and 66.19. Adding self-training update yields 71.43, 81.70, and 72.13. Adding FPA yields the final best result, with the strongest continuity scores on LAD, LCX, and RCA (Zhang et al., 2023). The dominant improvement is attributed to pseudo-label refinement, while FPA provides the final gain in continuity.

This segmentation-oriented PartVessel usage sits within a broader line of vessel-structure-aware learning. PC-Net improves local-global interaction for thin-vessel segmentation by combining pyramid squeeze-and-excitation with coarse-to-fine refinement (Wu et al., 2020). PASC-Net replaces standard convolutions with shape self-learning strip convolutions and adds hierarchical topological constraints to preserve connectivity and reduce breakpoints (Zhang et al., 5 Jul 2025). VPBSD constructs a vessel-pattern codebook from unlabeled 3D microscopic cerebrovascular data and distills that structural knowledge into a lightweight student model (Lin et al., 2024). In generative and geometric modeling, VesselVAE learns a recursive latent manifold of binary vascular trees (Feldman et al., 2023), differentiable voxelization methods fit spline-based centerline-and-radius models from segmentations (Dima et al., 3 Jul 2025), and the Vessel Coordinate System provides a centerline-plus-frame coordinate system for correspondence, shape analysis, and CFD sampling in single-branch vessels (Romero et al., 2023).

Taken together, these works suggest two recurring design principles. First, vessel methods increasingly separate topology from local geometry rather than representing vasculature as an undifferentiated voxel or point set. Second, clinically meaningful vessel analysis often privileges continuity, branching structure, and tractable annotation over raw voxel overlap alone. In that sense, both meanings of PartVessel—hierarchical part-based generation and partial vessels annotation—belong to the same broader shift toward structurally explicit vascular modeling.