SemanticVessel: Structured Vessel Semantics

• SemanticVessel is a framework that treats vessels as structured objects with explicit anatomy, topology, context, and task semantics to enhance interpretability.
• It integrates anatomy-aware pretraining, topology-preserving segmentation, and structured modeling to achieve robust vessel analysis in both medical and maritime domains.
• The approach leverages foundation-model adaptation, prompt-guided methods, and graph reconstructions to overcome annotation scarcity and domain shifts.

SemanticVessel, as reflected across recent vessel-focused research, denotes a mode of representation in which a vessel is treated not merely as a binary mask or a stream of coordinates, but as a structured object with anatomy, topology, context, and task semantics. In medical imaging, this orientation appears in anatomy-aware pretraining, topology-preserving segmentation, prompt-guided foundation-model adaptation, graph reconstruction, and parametric centerline-radius models (Huang et al., 14 Aug 2025, Song et al., 2024, Palaniappan et al., 1 May 2026, Dima et al., 3 Jul 2025). In maritime computing, it appears in semantic trajectory abstraction from AIS and in candidate-constrained vision-language fallback maneuver selection for autonomous vessels (Patroumpas et al., 8 Mar 2026, Christensen et al., 30 Dec 2025). This suggests that SemanticVessel is best understood as an umbrella for vessel analysis methods that preserve meaning across representation, modality, and downstream decision-making.

1. Semantic scope and representational invariance

A useful formal lens for SemanticVessel comes from semantic equivalence. SEAM defines two inputs as semantically equivalent when they encode the same task-relevant content despite different surface forms, and evaluates both cross-modal correctness and cross-modal agreement (Tang et al., 25 Aug 2025). For a semantically equivalent pair $(x^{(img)}, x^{(txt)})$, the ideal behavior is

$$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$

SEAM shows that contemporary vision-LLMs exhibit systematic modality imbalance, with vision frequently lagging language, and identifies two main drivers: textual perception failures from tokenization in domain notation and visual perception failures that induce hallucinations (Tang et al., 25 Aug 2025).

Applied to vessel research, this establishes a general criterion: a SemanticVessel system should preserve vessel meaning when the same structure is presented in different encodings, such as images, masks, graphs, prompts, or trajectory episodes. This is not a claim that all vessel papers explicitly optimize cross-modal agreement; rather, it is a plausible unifying interpretation. The supplied literature repeatedly replaces raw appearance with representations intended to be more semantically stable: anatomy-guided masking in angiograms, morphology-aware tensor fields in retinal images, voxel-wise direction vectors for graph recovery, and structured episode sequences for AIS data (Huang et al., 14 Aug 2025, Hu et al., 2023, Palaniappan et al., 1 May 2026, Patroumpas et al., 8 Mar 2026).

The term also has an important lexical breadth. In biomedical work, a vessel is a tubular anatomical structure; in maritime work, a vessel is a ship. The commonality is not the object class but the methodological move toward explicit semantics. In both settings, raw signals are regarded as insufficient: medical methods seek anatomy, continuity, and topology, while maritime methods seek mobility states, environmental context, hazard meaning, and human-interpretable action rationales (Song et al., 2024, Patroumpas et al., 8 Mar 2026, Christensen et al., 30 Dec 2025).

2. Anatomy-aware and topology-preserving segmentation

In medical imaging, one major SemanticVessel theme is that vessels should be segmented with explicit anatomical or topological bias rather than treated as generic foreground. "VasoMIM" introduces Vascular anatomy-aware Masked Image Modeling for X-ray angiograms because conventional MIM often fails to capture vascular anatomy under severe class imbalance between vessel and background pixels (Huang et al., 14 Aug 2025). Its two specified components are an anatomy-guided masking strategy, which preferentially masks vessel-containing patches, and an anatomical consistency loss, which enforces consistency in vascular semantics between original and reconstructed images; the abstract reports state-of-the-art performance across three datasets (Huang et al., 14 Aug 2025).

Temporal structure is another route to semantics. "Sequential vessel segmentation via deep channel attention network" formulates X-ray coronary angiography segmentation as prediction of the current-frame mask from a short 2D+t window, explicitly using $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$ to segment $F_i$ (Hao et al., 2021). The architecture combines 3D temporal-spatial feature extraction, a feature fusion operation that converts 3D encoder features to 2D decoder features, channel attention blocks in the decoder, and Dice loss for class imbalance. On its test data, the reported averages were DR 0.7640, P 0.8615, F 0.8056 for a 2D naive model; DR 0.7959, P 0.8640, F 0.8255 for a 3D naive model; and DR 0.8424, P 0.8492, F 0.8428 for 3D + CAB, indicating that temporal context and channel-wise refinement improve vascular discrimination in noisy XCA sequences (Hao et al., 2021).

A more explicitly topology-oriented formulation appears in "Optimized Vessel Segmentation: A Structure-Agnostic Approach with Small Vessel Enhancement and Morphological Correction," which proposes OVS-Net as a structure-agnostic multi-modality framework with five modules: macro vessel extraction, micro vessel enhancement, prompt encoder, mask prediction, and a morphological correction-based post-processing network (Song et al., 2024). The post-processing stage is trained to repair broken connections while preserving genuine discontinuities. On a 17-dataset benchmark spanning X-ray coronary artery, fundus retina, OCTA retina, and X-ray pelvic iliac artery data, OVS-Net reports average Dice 88.05, average IoU 79.38, average clDice 90.14, and average $\beta_0$ 0.34, together with a 34.6% improvement in connectivity relative to the second-best SAM-based baseline (Song et al., 2024).

SPIRONet pursues the same objective from a different angle. It uses dual encoders to capture local spatial and global frequency vessel features, a cross-attention fusion module, and a topological channel interaction module based on graph neural networks (Huang et al., 2024). On CADSA, CAXF, DCA1, and XCAD, it reports state-of-the-art performance and an inference speed of 21 FPS with a $512 \times 512$ input size, above the cited clinical real-time requirement of 6–12 FPS (Huang et al., 2024).

A recurrent misconception is that high overlap alone is sufficient for vessel understanding. The literature repeatedly rejects that view. OVS-Net introduces $\beta_0$ as a connectivity measure, SPIRONet adds topological channel interaction, and VasoMIM argues that vessel-aware pretraining is necessary because class imbalance otherwise weakens vascular representations (Song et al., 2024, Huang et al., 2024, Huang et al., 14 Aug 2025).

3. Foundation-model adaptation and anatomical prompting

A second major line of work adapts foundation models, especially SAM, to vessel-specific semantics. "VesSAM: Efficient Multi-Prompting for Segmenting Complex Vessel" argues that standard prompts such as full-image boxes or random sparse points are too weak for thin, branching, low-contrast vascular structures (Fu et al., 2 Nov 2025). VesSAM therefore integrates a convolutional adapter for local texture features, a multi-prompt encoder that fuses skeletons, bifurcation points, and segment midpoints via hierarchical cross-attention, and a lightweight mask decoder designed to reduce jagged artifacts. It also provides an automated pipeline that skeletonizes the vessel mask, extracts bifurcation points as positions with degree at least 3, partitions the skeleton into vessel segments, and derives segment midpoints (Fu et al., 2 Nov 2025).

The reported benchmark spans eight datasets across five imaging modalities. In the in-distribution setting, VesSAM achieves an average 79.89 IoU and 88.41 Dice; in the out-of-distribution setting, it reports 48.11 IoU and 61.40 Dice (Fu et al., 2 Nov 2025). The paper states that VesSAM beats SAMed by more than 10% Dice and 13% IoU, and its ablations show that skeletons, bifurcation points, and midpoints are complementary rather than redundant (Fu et al., 2 Nov 2025). This is a particularly clear SemanticVessel formulation: topology is injected directly through prompt design.

"VesselSAM: Leveraging SAM for Aortic Vessel Segmentation with AtrousLoRA" adapts SAM to aortic CTA by combining Low-Rank Adaptation with an atrous attention module (Iltaf et al., 25 Feb 2025). The central update is

$h = W_0 x + W_d \cdot \text{Atrous}(W_e x),$

where $W_0$ is frozen and $W_e$, $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$0 form the low-rank trainable path. The atrous module supplies multi-scale context, while LoRA limits the trainable footprint. On the AVT and TBAD datasets, VesselSAM reports DSC values of 93.50%, 93.25%, 93.02%, and 93.26% across multi-center datasets, and the details attribute to it 6.8M trainable parameters, or 7.2% (Iltaf et al., 25 Feb 2025).

These methods are notable for a common design principle: a foundation model is not assumed to be semantically adequate for vessels by default. Instead, vessel semantics are added through structured prompts, low-rank adaptation, atrous context aggregation, shallow feature injection, or connectivity-oriented post-processing (Fu et al., 2 Nov 2025, Iltaf et al., 25 Feb 2025, Song et al., 2024).

4. Scarce labels, domain shift, and latent vessel semantics

SemanticVessel research is also strongly shaped by annotation scarcity and domain heterogeneity. "VPBSD:Vessel-Pattern-Based Semi-Supervised Distillation for Efficient 3D Microscopic Cerebrovascular Segmentation" proposes a two-stage pipeline in which a teacher is pretrained on unlabeled data, a vessel-pattern codebook is learned by vector quantization, and a lightweight student is trained with supervised Dice, pseudo-label Dice, and feature distillation (Lin et al., 2024). On VesSep2020, VpbSD reports DSC 0.852, Accuracy 0.974, HD95 1.974, Jaccard 0.751, GWD 0.231, and clDice 0.900, improving over a baseline light UNet that had DSC 0.823 and clDice 0.873 (Lin et al., 2024). The codebook is presented as a discrete vocabulary of vessel patterns learned from unlabeled data.

Semi-supervised retinal segmentation under extreme label scarcity is addressed by "Semi-Supervised Semantic Segmentation of Vessel Images using Leaking Perturbations" (Hou et al., 2021). The framework combines a GAN-based semi-supervised architecture, a leaking mechanism that injects generator features into the discriminator, and a Mean Teacher regularizer with focal consistency. With only 8 labeled images, the reported results include Acc 95.74, Sp 86.72, Se 97.50 on DRIVE; Acc 95.65, Sp 91.86, Se 91.02 on STARE; and Acc 96.83, Sp 92.21, Se 94.72 on CHASE_DB1 (Hou et al., 2021). The cross-domain results further indicate that the same design is intended not only to exploit unlabeled data but also to resist dataset shift.

Domain generalization is made explicit in "VesselMorph: Domain-Generalized Retinal Vessel Segmentation via Shape-Aware Representation," which introduces a Hessian-based bipolar tensor field to represent morphology (Hu et al., 2023). The method uses multiscale Hessian analysis, learns latent representations from both the intensity image and the tensor field, and fuses them through a weight-balancing patch-swapping trick. The paper evaluates on six public datasets and reports superior performance in most out-of-domain settings, especially cross-resolution HRF and cross-modality OCTA500 and ROSE (Hu et al., 2023). The underlying claim is that intensity is domain-dependent whereas tubular morphology is domain-agnostic.

Cross-modality brain-vessel transfer is handled in "A2V: A Semi-Supervised Domain Adaptation Framework for Brain Vessel Segmentation via Two-Phase Training Angiography-to-Venography Translation" (Galati et al., 2023). A2V learns a disentangled latent space through a two-phase StyleGAN2-based pipeline and uses source angiography labels together with a small number of target venography labels. On its source-domain angiography test set, it reports Dice $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$1 and clDice $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$2; on target SWI venographies, Dice $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$3 and clDice $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$4, which the abstract summarizes as target Dice only 8.9% lower than source Dice (Galati et al., 2023). "Multi-Domain Brain Vessel Segmentation Through Feature Disentanglement" generalizes this direction by manipulating vessel appearances during adaptation while preserving spatial information such as shapes and locations, with evaluations spanning multi-center MRA, MRA-to-CTA, and MRA-to-MRV settings (Galati et al., 1 Oct 2025).

Across these papers, the semantic content of a vessel is increasingly displaced from raw intensity into codebooks, tensor fields, latent spaces, and label-preserving translations. This suggests that domain robustness in vessel analysis is often achieved by making vessel structure more explicit rather than by stronger appearance modeling alone.

5. From masks to graphs, splines, and controllable 3D models

A further expansion of SemanticVessel replaces dense masks with explicit structural representations. "Vesselpose: Vessel Graph Reconstruction from Learned Voxel-wise Direction Vectors in 3D Vascular Images" criticizes the segment-then-fix paradigm for topologically accurate vascular network reconstruction (Palaniappan et al., 1 May 2026). Its 3D U-Net jointly predicts a vessel mask and a voxel-wise direction vector field, and a direction-vector-guided extension of TEASAR then reconstructs the vascular graph. The method also introduces false merges and false splits as interpretable topological error measures. On the single-tree synthetic dataset, it reports edge F1 92.25, precision 95.49, recall 89.24, and branch-level F1 81.50; on PARSE 2022, it improves edge F1 from 39.46 to 57.52 and branch F1 from 23.46 to 35.33 relative to Trexplorer Super; on multi-tree synthetic data, it reaches 0.80 edge F1 versus 0.46 for U-Net+TEASAR or vesselFM (Palaniappan et al., 1 May 2026).

"Parametric shape models for vessels learned from segmentations via differentiable voxelization" joins voxels, meshes, and parametric shape models through differentiable transformations (Dima et al., 3 Jul 2025). The vessel is parameterized as centerlines and radii using cubic B-splines, optionally with radial adjustments for non-circular cross-sections, and supervision is applied through soft voxelization and Dice loss without explicit ground-truth shape parameters. The reported mean Dice scores are $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$5 on Aorta24, $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$6 on TopCoW, and $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$7 on MouseAneurysm (Dima et al., 3 Jul 2025). Because centerlines and radii are explicit, the fitted shape remains editable after fitting.

"Global Control for Local SO(3)-Equivariant Scale-Invariant Vessel Segmentation" adds user control to this structural turn (Rygiel et al., 2024). It combines a global controller, which supplies seed points and stopping criteria from coarse segmentation masks, with a local iterative vessel tracker and a signed-distance neural field for watertight surface reconstruction. On abdominal aortic aneurysm CTA, the proposed method achieves $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$8 Dice versus $$f(x^{(img)}) = y,\qquad f(x^{(txt)}) = y,\qquad f(x^{(img)}) = f(x^{(txt)}).$$9 for nnU-Net, while also producing watertight, smooth, sub-voxel 3D vessel models and allowing the superior extent of the aorta to be changed by modifying the global controller rather than retraining (Rygiel et al., 2024).

A common misconception is that vessel analysis terminates at segmentation. These papers argue the opposite. Graphs, centerlines, radii, branches, roots, false splits, and false merges are treated as first-class outputs because clinical or analytic use often depends on topology, geometry, and editability rather than on voxel overlap alone (Palaniappan et al., 1 May 2026, Dima et al., 3 Jul 2025, Rygiel et al., 2024).

6. Maritime semantic vessels: trajectories, hazards, and fallback action

In maritime computing, SemanticVessel takes a different but structurally analogous form. "Context-Enriched Natural Language Descriptions of Vessel Trajectories" transforms raw AIS sequences into semantic trajectories by segmenting them into trips and mobility-annotated episodes, then enriching each episode with nearby geographic entities, offshore navigation features, weather conditions, and bathymetry (Patroumpas et al., 8 Mar 2026). The episode types are STOPPED, TURNING, MANEUVERING, COMMUNICATION GAP, and SAILING. The reported dataset comprises 460 trips between 10 ports, totaling 212,203 raw AIS locations and yielding 3,332 episodes (Patroumpas et al., 8 Mar 2026). Five open models were then evaluated for controlled natural-language generation from structured JSON, and the strongest model, gpt-oss-120B, achieved relevance $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$0, faithfulness $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$1, and correctness $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$2 on a 1–5 scale (Patroumpas et al., 8 Mar 2026).

The same transition from geometry to semantics appears in autonomy. "Foundation models on the bridge: Semantic hazard detection and safety maneuvers for maritime autonomy with vision-LLMs" introduces Semantic Lookout, a camera-only, candidate-constrained fallback maneuver selector for the alert-to-takeover gap under the draft IMO MASS Code (Christensen et al., 30 Dec 2025). The system first segments water, builds a pixel-space clearance map,

$(F_{i-2}, F_{i-1}, F_i, F_{i+1})$3

retains only water-valid projected candidates satisfying $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$4 and $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$5, then asks a VLM to choose one numbered maneuver or station-keeping from the overlay image (Christensen et al., 30 Dec 2025). On 40 harbor scenes, the best scene-understanding result is awareness $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$6 for gpt-5-high at $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$7 s, while sub-10 s models such as gpt-4.1 and gpt-5-minimal retain most of that awareness at $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$8 s and $(F_{i-2}, F_{i-1}, F_i, F_{i+1})$9 s, respectively (Christensen et al., 30 Dec 2025). For action alignment, gpt-5-low achieves Accept@1 = 0.68 [0.52, 0.80] and Best@1 = 0.48 [0.33, 0.63], outperforming geometry-only baselines; on fire scenes, FB-3 (gpt-5-low) improves average standoff distance by +1.1 m at 30 s and +3.5 m at 60 s (Christensen et al., 30 Dec 2025).

The maritime papers make explicit what is only implicit in many medical ones: semantic modeling is valuable because raw coordinates or raw pixels do not explain what the vessel is doing or what action is warranted. In AIS summarization, this yields human-readable trip narratives grounded in structured episodes. In fallback autonomy, it yields a human-overridable maneuver chosen for scene meaning rather than for geometry alone (Patroumpas et al., 8 Mar 2026, Christensen et al., 30 Dec 2025).

Taken together, the literature presents SemanticVessel not as a single architecture but as a recurring research strategy: encode vessel-specific meaning explicitly, preserve that meaning across representations, and evaluate success in terms of continuity, topology, agreement, and operational usefulness rather than overlap alone.