MultiVesSeg: Domain Adaptation for Brain Vessels

Updated 4 July 2026

The paper introduces MultiVesSeg, an out-of-the-box framework that uses label-preserving image-to-image translation to maintain spatial vessel structures during domain adaptation.
Its architecture leverages a StyleGAN2-based shared latent space with dynamic skip connections, enabling effective feature disentanglement across varying imaging modalities and centers.
Evaluation on multiple transfer settings shows improved segmentation accuracy and annotation efficiency, highlighting robust vessel continuity preservation even under significant domain shifts.

MultiVesSeg is an out-of-the-box semi-supervised domain adaptation framework for brain vessel segmentation across domains. It is designed for segmentation of the cerebrovascular tree across different medical centers, different imaging modalities, different vessel types, and different acquisition characteristics, while avoiding domain-specific model design and data harmonization between the source and the target domain. Its central mechanism is feature disentanglement implemented through a StyleGAN2-based shared latent space and label-preserving image-to-image translation, so that vessel appearance can be manipulated during adaptation while spatial information such as shapes and locations is preserved (Galati et al., 1 Oct 2025).

1. Problem setting and domain-shift model

MultiVesSeg addresses brain vessel segmentation under domain shift, where a model trained on one domain does not generalize well to another. The paper studies three concrete transfer settings: multi-center MRA from TOF MRA from OASIS-3 $\to$ TOF MRA from IXI, MRA-to-CTA from TOF MRA $\to$ CTA, and MRA-to-MRV from TOF MRA arterial scans $\to$ SWI venography. These shifts are not treated as only low-level appearance changes. The framework is explicitly motivated by variation in imaging modality, medical center / scanner / protocol, vessel type, and acquisition and appearance factors such as intensity range, contrast, noise, pixel spacing / spatial resolution, texture, and vessel brightness polarity (Galati et al., 1 Oct 2025).

The problem is difficult because cerebral vessels are small and sparse relative to the full 3D volume, vascular morphology is complex and highly branching, and domain gaps differ in severity. The paper characterizes MRA vs MRA from another center as mild, MRA vs CTA as moderate, and MRA vs MRV/SWI as very large, because arteries and veins differ not only in appearance but also in geometry and anatomical distribution. A naive image translation is therefore inadequate: if translation changes vessel position, thickness, density, or shape, the source labels cease to be valid. MultiVesSeg is built around the constraint that adaptation should alter vessel appearance in a label-preserving manner, rather than maximize visual realism.

The paper also distinguishes volume-level properties and vessel-level properties. Volume-level properties include spacing, scale, contrast, and resolution; vessel-level properties include intensities, texture, shapes, locations, and densities. This distinction is important because the method does not simply translate all domain-specific factors. Instead, it seeks to translate only a subset of factors compatible with segmentation supervision.

2. Disentangled latent space and model architecture

The framework has four principal components: a Generator $G$ based on StyleGAN2, a Discriminator $D$ used during latent-space learning, an Encoder $E$ that maps an image into the StyleGAN latent space, and a label-synthesis branch $G_{lsb}$ adapted from DatasetGAN. The generator learns a shared latent space $\mathcal{W}$ , more precisely $\mathcal{W}^+$ on the encoder side, spanning both source and target domains. The encoder receives a binary domain flag $d$ , with $\to$ 0 meaning encode toward the source-domain latent representation and $\to$ 1 meaning encode toward the target-domain latent representation. The label-synthesis branch is implemented as three fully connected layers attached to generator features and predicts segmentation masks aligned with generated images (Galati et al., 1 Oct 2025).

The encoder uses a ResNet backbone inspired by pixel2style2pixel / residual encoder designs. It predicts both a latent code in $\to$ 2 and a feature tensor, and it uses a dynamic skip connection module to connect encoder features into the generator and preserve fine content correspondence. The paper’s architectural interpretation is that these residual and skip pathways are essential for preserving spatial information during domain translation.

Disentanglement is induced functionally rather than through explicitly separated content and style vectors. The intended separation is between domain-invariant structural/spatial information—including vessel positions, vessel shapes, and anatomy needed for segmentation—and domain-specific appearance information such as intensities, vessel darkness/brightness, textures, and contrast. The method is careful not to force translation of properties that would violate label preservation, especially in artery-to-vein adaptation, where vessel geometry and density may differ intrinsically.

The latent space is regularized by path length regularization. The Phase 1 objective is

$\to$ 3

where $\to$ 4 is the non-saturating GAN loss, $\to$ 5 is $\to$ 6 regularization, and

$\to$ 7

encourages small latent changes to produce smooth, semantically structured image changes. In the paper’s formulation, this regularization is what makes independent manipulation of image properties possible.

3. Label-preserving translation and optimization

Training proceeds in two phases. In Phase 1, $\to$ 8 and $\to$ 9 are trained adversarially on both source and target images for 250,000 iterations in order to learn a unified latent space $\to$ 0 covering both domains. In Phase 2, $\to$ 1 is frozen except for $\to$ 2, and $\to$ 3 is trained for 20,000 iterations to perform label-preserving image-to-image translation and segmentation. The implementation details also mention a preliminary phase of 15,000 iterations before Phase 2 using only source data (Galati et al., 1 Oct 2025).

The semi-supervised setting uses three sets: $\to$ 4 with $\to$ 5. No paired source-target scans are required. For a source image $\to$ 6, the encoder can produce both an intra-domain reconstruction and an inter-domain translation: $\to$ 7

$\to$ 8

The same logic is used symmetrically for target images. The label-synthesis branch predicts segmentation masks for both reconstruction and translation outputs, and both are supervised by the same reference annotation whenever translation is intended to be label-preserving.

The segmentation loss is

$\to$ 9

with

$G$ 0

The reconstruction loss is

$G$ 1

The paper does not provide an explicit single weighted Phase 2 formula, but it states that $G$ 2 is updated using $G$ 3 and $G$ 4, while $G$ 5 is updated using $G$ 6.

At test time, a target image $G$ 7 is encoded twice: once with $G$ 8 to obtain a target-side prediction $G$ 9, and once with $D$ 0 to obtain a source-side prediction $D$ 1. The final segmentation is produced by averaging $D$ 2 and $D$ 3, followed by argmax. This dual-view inference is one of the concrete mechanisms through which MultiVesSeg combines source and target latent interpretations.

The framework operates in a 2.5D approach, meaning volumetric data are processed as a series of multi-channel 2D slices rather than full 3D volumes. This is a methodological choice rather than an incidental implementation detail, and it directly conditions the model’s computational profile and its handling of cross-slice continuity.

4. Data, preprocessing, and evaluation protocol

The source domain is OASIS-3, comprising 49 TOF MRA volumes with median size $D$ 4 and median voxel size $D$ 5 mm. The target domains are IXI with 50 TOF MRA volumes, TopCoW with 40 CTA volumes, and a private SWI venography dataset with 28 SWI volumes. For OASIS-3, the split is 35 volumes train, 7 validation, 7 test. For the target domains, the paper uses unlabeled training sets $D$ 6, validation sets, test sets, and a very small labeled subset $D$ 7: 3 midpoint slices for IXI, 3 whole target volumes for TopCoW because annotations only cover the circle of Willis, and 3 midpoint slices for the SWI dataset (Galati et al., 1 Oct 2025).

Preprocessing is done separately per dataset, without inter-domain harmonization. The steps are: skull stripping / brain masks via SynthStrip; resampling each dataset to a fixed spacing equal to that dataset’s median spacing, with minor increases if needed to fit in $D$ 8; normalization using mean/std; intensity clipping to the 0.1 and 99.9 percentiles; rescaling to $D$ 9; and one-hot encoding segmentation masks into three classes: background, brain, and vessels. The paper stresses that it avoids explicit cross-domain harmonization.

The evaluation uses Dice, clDice, ASSD, Precision, and Recall. The baselines are 2D U-Net, CycleGAN, SIFA, SynthSeg, UniverSeg, AADG, DCDA, and CS-CADA. The paper notes fairness caveats because some comparison methods were originally designed for larger structures or 2D settings, but they were adapted where possible.

The main transfer settings are summarized below.

Scenario	Target domain	Target supervision
Multi-center MRA	IXI	$E$ 0: 3 midpoint slices
MRA-to-CTA	TopCoW	$E$ 1: 3 whole target volumes
MRA-to-MRV	SWI private dataset	$E$ 2: 3 midpoint slices

This experimental design makes the method explicitly semi-supervised domain adaptation rather than domain generalization. The paper also studies unsupervised adaptation when $E$ 3 and a few-shot target-only segmentation condition when $E$ 4, but the main framework is formulated for the SSDA regime.

5. Quantitative results and ablation findings

On the source OASIS-3 domain, the reported intra-domain performance is Dice 73.7 ± 2.8, Precision 66.9 ± 4.8, Recall 82.5 ± 3.7, and clDice 76.9 ± 5.2. The authors state that this is comparable to state-of-the-art artery segmentation methods. In the transfer experiments, MultiVesSeg achieves Dice 69.9 ± 2.3, clDice 76.8 ± 2.9, and ASSD 1.55 ± 0.57 for OASIS-3 $E$ 5 IXI; Dice 74.5 ± 4.2, clDice 78.0 ± 8.7, and ASSD 1.46 ± 0.49 for OASIS-3 $E$ 6 TopCoW; and Dice 67.5 ± 1.7, clDice 69.9 ± 2.7, and ASSD 1.42 ± 0.20 for OASIS-3 $E$ 7 SWI (Galati et al., 1 Oct 2025).

Transfer setting	MultiVesSeg result	Noted comparison
OASIS-3 $E$ 8 IXI	Dice 69.9 ± 2.3, clDice 76.8 ± 2.9, ASSD 1.55 ± 0.57	best or near-best among all methods
OASIS-3 $E$ 9 TopCoW	Dice 74.5 ± 4.2, clDice 78.0 ± 8.7, ASSD 1.46 ± 0.49	better than U-Net (70.5 Dice) and SIFA (60.9)
OASIS-3 $G_{lsb}$ 0 SWI	Dice 67.5 ± 1.7, clDice 69.9 ± 2.7, ASSD 1.42 ± 0.20	huge margin over all baselines

The MRA-to-MRV setting is the most striking. The reported baselines there are U-Net 29.1 Dice, SynthSeg 10.9, CycleGAN 5.1, SIFA 0.8, DCDA 0.0, CS-CADA 0.4, and UniverSeg 3.6. The paper interprets this as evidence that disentanglement plus label-preserving translation is especially effective when the domain shift involves both appearance and vessel-type changes.

The ablation studies are concentrated on the hardest MRA-to-MRV setup. Increasing target annotations from $G_{lsb}$ 1 to $G_{lsb}$ 2 slice gives a large improvement, after which gains saturate; the authors therefore use $G_{lsb}$ 3 for the main experiments. On the source side, going from $G_{lsb}$ 4 to $G_{lsb}$ 5 source volumes gives a sharp improvement, while increasing to $G_{lsb}$ 6 yields only a modest further boost of +3.4% Dice in the hardest setting. This suggests that MultiVesSeg is annotation-efficient on the target domain but still strongly benefits from source supervision.

Architectural ablations further identify the decisive components. In MRA-to-MRV, the full model achieves Dice 72.2 ± 2.5 and clDice 75.4 ± 3.3. Removing residual connections reduces Dice to 14.4 ± 3.4. Removing DSBN gives Dice 69.3; removing balanced data sampling gives Dice 71.2; removing intensity inversion gives Dice 71.8. The paper interprets this as showing that residual connections are crucial, likely because they preserve spatial information and support fine correspondences. In the unsupervised MRA-to-MRV case, removing intensity inversion causes a severe collapse, indicating that some appearance guidance is needed when the shift is extremely wide.

Qualitative findings are aligned with the quantitative results. The paper emphasizes improved preservation of vessel continuity, complex branching patterns, spatial alignment, and thin vessel recovery relative to alternatives. It also states that the translated images are often hybrid translations rather than perfectly source-like or target-like images. This is not treated as a flaw; the paper is explicit that segmentation utility rather than visual realism is the objective.

6. Position within vessel-segmentation research, limitations, and future directions

MultiVesSeg occupies a specific position within current segmentation research. It is a domain adaptation method for brain vessels, not a universal one-model-no-retraining framework. This distinguishes it from UVSM, which targets multi-modality retinal vessel segmentation across CF, MC, FA, FAF, and IR/NIR using a canonical-domain translation pipeline and a topology-aware segmentation stage without target-modality finetuning (Wen et al., 10 Feb 2025). It is also distinct from MDSVM-UNet, which addresses 3D coronary artery segmentation from CTA through multidirectional snake convolution, residual visual Mamba, and a two-stage coarse-to-fine pipeline rather than cross-domain adaptation (Yuan et al., 23 Mar 2026). In another direction, MultiFlow addresses multi-vessel classification and segmentation from phase-contrast MRI in single-ventricle patients, combining joint vessel-plane classification, 2D+time segmentation, and downstream deep temporal clustering (Yao et al., 17 Feb 2025). The name should also not be conflated with MultiverSeg, an interactive biomedical segmentation system based on in-context guidance rather than vessel-domain adaptation (Wong et al., 2024).

The paper explicitly discusses several limitations. First, MultiVesSeg still requires retraining for each new target domain. It is therefore not a domain generalization method. Second, for very large shifts—especially MRA-to-MRV—the method still benefits substantially from a small number of target labels, and completely unsupervised adaptation may require weak appearance guidance such as intensity inversion. Third, training is computationally costly: approximately 194 hours (~8 days) in total on two NVIDIA GeForce RTX 2080 Ti GPUs, 12 GB each, with Phase 1 taking about 166 hours and Phase 2 about 28 hours. Fourth, the method is 2.5D rather than full 3D, which may affect long-range vessel continuity across slices. Fifth, several implementation details remain unspecified in the manuscript, including the optimizer for MultiVesSeg itself, learning rates, weight decay, exact Phase 2 loss weights, and the full augmentation protocol.

The proposed future directions are correspondingly clear. The paper points to moving beyond repeated domain adaptation toward domain generalization or in-context learning, leveraging foundation models pre-trained on large collections of tree-like structures, reducing dependence on target annotations, and improving cross-domain vessel linking without manual guidance. A plausible implication is that MultiVesSeg is best understood as a strong transitional framework: it shows that label-preserving translation in a disentangled latent space can bridge large shifts across centers, modalities, and vessel types, while also exposing the remaining gap between semi-supervised adaptation and truly domain-agnostic cerebrovascular segmentation.