MorVess: Morphology-Aware Vessel Segmentation

• MorVess is a morphology-aware pulmonary vessel segmentation network that reframes vessel extraction as a geometric and topological reconstruction task.
• It integrates a SAM-based backbone, a lightweight 2.5D adapter, and geometric priors (VDM & VTM) to boost small vessel recovery, boundary precision, and continuity.
• Its multi-head design and two-stage training strategy yield improved performance metrics on chest CT datasets while maintaining efficiency and structural fidelity.

Searching arXiv for the specified paper and related disambiguation. Searching arXiv for "MorVess pulmonary vessel segmentation". MorVess is a morphology-aware pulmonary vessel segmentation network for chest CT that treats vessel parsing as a geometric and topological reconstruction problem rather than only a voxel-wise foreground–background classification problem. Introduced in "MorVess: Morphology-Aware Pulmonary Vessel Segmentation Network" (Mao et al., 23 Jun 2026), it combines a SAM-based backbone, a lightweight 2.5D adapter for volumetric context, differentiable geometric prior supervision, and a global-local fusion mechanism to improve small-vessel recovery, centerline continuity, boundary precision, and diameter consistency. The term should be distinguished from V-MORALS, a separate visual reachability-analysis method in robotics (Aladin et al., 26 Feb 2026).

1. Scope, nomenclature, and problem setting

MorVess addresses pulmonary vessel segmentation in chest CT, a setting characterized by sparse and tree-like anatomy, high tortuosity, extensive branching, multi-scale structure, low contrast, noise, and partial volume effects (Mao et al., 23 Jun 2026). The paper identifies several recurrent failure modes in standard segmentation networks: missing small peripheral vessels, broken or disconnected centerlines, jagged or blurred boundaries, and poor preservation of topology and diameter continuity.

The method’s central premise is that conventional voxel-wise supervision, such as binary mask supervision with Dice or cross-entropy losses, is insufficient because it frames vessel segmentation as a local classification problem rather than a geometric and topological reconstruction problem (Mao et al., 23 Jun 2026). MorVess therefore introduces explicit morphology-aware supervision. It jointly predicts three outputs—vessel mask, vessel distance map, and vessel thickness map—to encode boundary localization, centerline continuity, and smooth caliber transitions within a unified framework.

This design places MorVess within a broader class of geometry-aware medical image segmentation systems, but its specific contribution is to integrate differentiable geometric priors with large-scale foundation model adaptation. The paper’s emphasis is not only on overlap accuracy, but also on topology integrity and clinically relevant structural consistency.

2. Architectural organization

MorVess is a SAM-based segmentation framework adapted to 3D pulmonary CT through a lightweight 2.5D adapter and a multi-head decoder (Mao et al., 23 Jun 2026). The architecture is organized around three major components: geometric prior generation, 2.5D SAM adaptation, and a global-local fusion block. Training is performed in two stages: stage I for adaptation of 2.5D and geometric heads, and stage II for topological refinement.

The network jointly predicts a semantic vessel mask, a vessel distance map (VDM), and a vessel thickness map (VTM). This multi-head design extends the original SAM mask decoder with two additional geometric branches. The geometric heads use the mean embedding of mask tokens,

$$\bar{h}_{\text{mask}}=\text{Mean}(h_{\text{mask}}),$$

which is then passed through separate MLPs to generate dynamic weights for the VDM and VTM heads (Mao et al., 23 Jun 2026). The paper presents this as a lightweight way to inject global mask context into geometric prediction.

A key adaptation layer is the 2.5D adapter, which bridges volumetric CT input and SAM’s 2D representation. The input volume is reshaped into a sequence of slices with shape $[B, N, H, W]$, and the adapter uses downsampling, a $3\times1\times1$ 3D convolution, upsampling, and a residual connection (Mao et al., 23 Jun 2026). This mechanism is designed to inject inter-slice depth context while preserving the efficiency and representational strengths of the frozen SAM ViT encoder.

The global-local fusion block (GLFB) is introduced because the SAM decoder primarily uses final-layer features, which may miss delicate vessel branches. GLFB fuses five sources:

1. $F_{\text{dec}}$, the decoder feature;
2. $F_{\text{shallow}}$, the shallow encoder feature;
3. $F_{\text{depth}}$, the deep encoder feature;
4. VDM;
5. VTM.

These are concatenated as

$$F_{\text{aggregated}}=\text{concat}\left[F_{\text{dec}}, F_{\text{shallow}}, F_{\text{depth}}, \text{VDM}, \text{VTM}\right],$$

then processed to produce $F_{\text{fused}}$ (Mao et al., 23 Jun 2026). A vessel token representation $h_{\text{vessel}}$ is mapped to channel-wise weights,

$W_{\text{vessel}}=\text{MLP}(h_{\text{vessel}}),$

and the fused features are modulated to yield vessel logits, followed by residual refinement:

$[B, N, H, W]$0

The paper argues that this combination of global semantics, local texture, and explicit geometry is particularly important for recovering small branches and refining topology.

3. Geometric priors and optimization objective

MorVess formalizes morphology-aware supervision through two differentiable geometric prior maps: the vessel distance map and the vessel thickness map (Mao et al., 23 Jun 2026).

The VDM is intended to convert vessel boundaries into a smooth, differentiable potential field. Boundary voxels receive stronger supervision, and the resulting field is designed to improve boundary precision while reducing jagged contours and weak edge learning. The final supervisory field is defined through exponential decay,

$[B, N, H, W]$1

with $[B, N, H, W]$2 (Mao et al., 23 Jun 2026).

The VTM encodes global vessel diameter consistency. It is derived from an internal distance field, a topology-preserving skeleton $[B, N, H, W]$3, and the radius at each skeleton point. The resulting map assigns vessel thickness to foreground voxels based on the nearest skeleton point, thereby imposing a global diameter prior (Mao et al., 23 Jun 2026). According to the paper, this encourages smooth diameter transitions, discourages non-physical oscillations in vessel width, and improves distal branch continuity.

These geometric priors are coupled to the segmentation objective through a weighted total loss,

$[B, N, H, W]$4

with

$[B, N, H, W]$5

The semantic terms are cross-entropy, Dice loss, and clDice loss, while the geometric regression terms supervise distance and thickness maps (Mao et al., 23 Jun 2026).

The thickness loss is formulated as a scale-invariant $[B, N, H, W]$6 loss between normalized predictions and ground truth:

$[B, N, H, W]$7

with predicted thickness logits passed through Softplus to ensure non-negativity (Mao et al., 23 Jun 2026).

The paper interprets the supervision terms as follows: VDM provides boundary supervision, clDice enforces centerline consistency, and VTM regularizes smooth diameter transitions and global morphology consistency. Taken together, these losses move the model beyond mask fitting toward anatomical reconstruction.

4. Two-stage training and implementation protocol

MorVess uses a two-stage fine-tuning strategy designed to stabilize adaptation from a pretrained foundation model to volumetric pulmonary CT (Mao et al., 23 Jun 2026). In stage I, the SAM ViT encoder is frozen while the 2.5D adapter and the multi-head decoder are trained. This stage uses a learning rate of $[B, N, H, W]$8 for 400 epochs. In stage II, the converged 2.5D adapter is frozen and the geometric and fusion heads are refined using a learning rate of $[B, N, H, W]$9 for 200 epochs.

The implementation is based on PyTorch 2.0, CUDA 11.8, and Ubuntu 20.04, and training is reported on a single NVIDIA L40 GPU with 40 GB memory, with TF32 and AMP enabled (Mao et al., 23 Jun 2026). The backbone is SAM ViT-Base with official pretrained weights.

Input preprocessing is also specified. CT intensities are clipped to $3\times1\times1$0 HU and normalized to $3\times1\times1$1. Five adjacent slices are used as a 2.5D input stack, the input size is $3\times1\times1$2, the batch size is 4, and evaluation uses 5-fold cross-validation (Mao et al., 23 Jun 2026).

This training protocol reflects a specific trade-off. The adapter introduces quasi-3D context without retraining the full backbone, and the staged optimization isolates the volumetric adaptation problem from the later topological refinement problem. A plausible implication is that the authors seek to preserve pretrained visual priors while constraining the number of trainable parameters and the optimization instability typical of full 3D foundation-model fine-tuning.

5. Datasets, evaluation metrics, and quantitative performance

The primary evaluation is conducted on two pulmonary CT benchmarks: Parse2022 and AIIB2023 (Mao et al., 23 Jun 2026). Parse2022 is the MICCAI 2022 Pulmonary Artery Segmentation Challenge dataset and contains 100 high-resolution chest CT scans with finely annotated vessel ground truth. AIIB2023, the Airway-Informed Imaging Biomarker 2023 dataset, contains 300 chest CT scans from pulmonary fibrosis patients and is described as challenging because of distorted morphology and density inhomogeneity. The paper also reports cross-domain tests on HiPas and ATM2022.

The evaluation metric suite includes Dice (DSC), clDice, HD95, AMR, DBR, and DLR (Mao et al., 23 Jun 2026). These jointly assess voxel overlap, centerline or topology overlap, boundary quality, false-negative sensitivity, branch recovery, and detected vessel length.

Benchmark	Core scores	Structural scores
Parse2022	Dice 86.84, clDice 83.22, HD95 4.53 mm	AMR 0.12, DBR 0.80, DLR 0.83
AIIB2023	Dice 94.31, clDice 89.34, HD95 3.24 mm	AMR 0.07, DBR 0.86, DLR 0.89

On Parse2022, MorVess improves Dice from 76.85 to 86.84 relative to Swin-UNETR and from 83.27 to 86.84 relative to COMMA; clDice improves from 80.10 to 83.22, and HD95 improves from 5.11 to 4.53 (Mao et al., 23 Jun 2026). On AIIB2023, the paper states that MorVess outperforms all baselines, including nnU-Net, Swin-UNETR, SegMamba, Diff-UNet, DSCNet, and COMMA.

The paper reports bootstrap paired resampling with 10,000 iterations and states that topology-related gains are statistically significant, especially for clDice, DBR, and DLR (Mao et al., 23 Jun 2026). It further emphasizes that the strongest gains occur in small-vessel recovery, global connectivity preservation, fragmentation reduction, and centerline continuity. Qualitative analysis is described as showing that many baselines miss terminal vessels, break fine branches, and produce discontinuities in difficult regions, whereas MorVess preserves these regions more effectively.

These results support the paper’s claim that geometry-aware supervision improves not only overlap metrics but also the structural fidelity of the vascular tree. For pulmonary vessel analysis, this distinction is consequential because branch continuity and diameter consistency are directly tied to downstream quantification.

6. Ablations, efficiency, generalization, and limitations

The ablation studies isolate the contribution of geometric priors, adaptation modules, and foundation-model initialization (Mao et al., 23 Jun 2026). On Parse2022, the geometric-prior study reports the following progression:

Variant	Dice / clDice / HD95
Baseline	82.4 / 74.24 / 6.85
+VDM	83.9 / 79.32 / 5.72
+VTM	84.1 / 78.75 / 5.59
Full model	86.84 / 83.22 / 4.53

The paper interprets these results by attributing boundary sharpness and continuity gains primarily to VDM, and diameter consistency and topology gains primarily to VTM (Mao et al., 23 Jun 2026). Their combination yields the best overall performance, suggesting complementary roles.

A component ablation reports a baseline Dice of 0.6844, rising to 0.7233 with the 2.5D adapter, 0.7481 with GLFB, 0.7626 with 2.5D plus GLFB, and 0.8544 for the full model (Mao et al., 23 Jun 2026). The same section states that adding SAM pretrained weights significantly boosts performance. The paper’s interpretation is that the 2.5D adapter is crucial for inter-slice continuity, GLFB is crucial for multi-scale fusion, and SAM pretraining provides strong transferable visual priors.

A rank ablation for FacT/LoRA-style tuning evaluates ranks $3\times1\times1$3 and reports that performance improves with rank and saturates around 32, which is selected as the best trade-off between performance and parameter efficiency (Mao et al., 23 Jun 2026).

Efficiency is another explicit design goal. MorVess is reported to use about 1.0M trainable parameters, 42 GMACs per 5-slice stack, and 4.2 GB VRAM, compared with nnU-Net at 32M parameters, 180 GMACs, and 18 GB, and Diff-UNet at 64M parameters, 340 GMACs, and 32 GB (Mao et al., 23 Jun 2026). This indicates that the method’s geometry-aware improvements are not achieved by simply scaling parameter count.

Cross-dataset evaluation further examines whether the learned priors generalize beyond training distributions. For Parse2022 to HiPas transfer, MorVess reports Dice 81.14, clDice 78.42, and HD95 7.18; for AIIB2023 to ATM2022 transfer, it reports Dice 89.25, clDice 86.75, and HD95 4.22 (Mao et al., 23 Jun 2026). The paper interprets this as evidence that VDM and VTM encode shape and topology priors rather than only dataset-specific texture cues.

The study also includes geometric analysis using VMTK centerlines and reports improved consistency in total vessel volume, diameter distribution, and small-vessel volume fraction (Mao et al., 23 Jun 2026). This suggests that the segmentation improvements translate into more reliable structural measurements, especially for small vessels and diameter gradients.

The limitations stated or implied in the paper remain important. Pulmonary vessel segmentation is difficult because of low contrast, noise, and multi-scale branching; MorVess addresses these issues through geometry-aware supervision, but its claims are bounded by the evaluated datasets and the specific SAM-based adaptation pipeline. A plausible implication is that the approach is strongest when anatomical tubular priors are informative and when topology-preserving supervision can be estimated reliably from annotations.