Multi-Domain Brain Vessel Segmentation Through Feature Disentanglement

TL;DR

This work tackles the challenge of segmenting brain vessels across multiple imaging modalities and centers by learning a disentangled latent representation that separates domain-specific appearance from vessel geometry. It introduces a two-phase framework that uses StyleGAN2-based disentanglement, a label-preservation mechanism with a label-synthesis branch, and cycle-consistency to enable label-preserving image translations across domains without homogenizing preprocessing. Comprehensive ablations show how annotations and architectural choices affect performance, and the method achieves strong results across three challenging domain shifts (multi-center MRA, MRA-to-CTA, and MRA-to-MRV), including veins, which are typically harder to segment. The approach advances practical cerebrovascular segmentation by enabling robust, cross-domain performance while reducing the need for domain-specific tuning, with open-source code provided for reproducibility and further development.

Abstract

The intricate morphology of brain vessels poses significant challenges for automatic segmentation models, which usually focus on a single imaging modality. However, accurately treating brain-related conditions requires a comprehensive understanding of the cerebrovascular tree, regardless of the specific acquisition procedure. Our framework effectively segments brain arteries and veins in various datasets through image-to-image translation while avoiding domain-specific model design and data harmonization between the source and the target domain. This is accomplished by employing disentanglement techniques to independently manipulate different image properties, allowing them to move from one domain to another in a label-preserving manner. Specifically, we focus on manipulating vessel appearances during adaptation while preserving spatial information, such as shapes and locations, which are crucial for correct segmentation. Our evaluation effectively bridges large and varied domain gaps across medical centers, image modalities, and vessel types. Additionally, we conduct ablation studies on the optimal number of required annotations and other architectural choices. The results highlight our framework's robustness and versatility, demonstrating the potential of domain adaptation methodologies to perform cerebrovascular image segmentation in multiple scenarios accurately. Our code is available at https://github.com/i-vesseg/MultiVesSeg.

Figures (7)

• Figure 1: Maximum intensity projection (MIP) of a magnetic resonance angiography (left), MIP of a computed tomography angiography (center), and minimum intensity projection (mIP) of a magnetic resonance venography (right). All images are skull-stripped and viewed from the axial perspective.
• Figure 2: During the two-phase training algorithm, images $x_i$ from domains $\mathcal{S}$ and $\mathcal{T}$ are input into our model consisting of the generator $G$, discriminator $D$, and encoder $E$. The training process is split into two distinct phases. In Phase 1 (left), $G$ undergoes adversarial training with $D$ to build a unified latent space that is both disentangled and semantically rich. In Phase 2 (right), the encoder $E$ is trained for label-preserving image-to-image translation, while $G$ is refined to generate segmentation masks $\hat{y}_i^t$ and $\hat{y}_i^s$.
• Figure 3: In Phase 2 of our training algorithm, we perform both source and target reconstructions (first row, source domain on the left and target domain on the right) and source-to-target and target-to-source translations (second and third rows). The backpropagation of $\mathcal{L}_{\textrm{r}}$ exclusively updates the weights of $E$, while $\mathcal{L}_{\textrm{s}}$ influences both $E$ and $G$.
• Figure 4: Vessel segmentation performance with varying target annotations $m$ (left) and source annotations $N$ (right). Vertical error bars represent the standard deviation across the testing set.
• Figure 5: Comparison of the segmentation results for brain and vessels in the target MRA, CTA, and SWI images using different methods. Red indicates brain masks, while green represents vessels. The rows display slices at varying levels: top, middle, and bottom.