[Paper Review] A2V: A Semi-Supervised Domain Adaptation Framework for Brain Vessel Segmentation via Two-Phase Training Angiography-to-Venography Translation

Francesco Galati, Daniele Falcetta+5 more|arXiv (Cornell University)|Sep 12, 2023

Fetal and Pediatric Neurological DisordersMedicine3 citations

TL;DR

This paper proposes A2V, a semi-supervised domain adaptation framework for brain vessel segmentation that translates MR angiography to venography via a two-phase training process, leveraging a disentangled latent space to bridge domain gaps. It achieves a 70.4% Dice score on venous vessel segmentation with only limited annotated venographies, outperforming state-of-the-art methods by over 19 percentage points.

ABSTRACT

We present a semi-supervised domain adaptation framework for brain vessel segmentation from different image modalities. Existing state-of-the-art methods focus on a single modality, despite the wide range of available cerebrovascular imaging techniques. This can lead to significant distribution shifts that negatively impact the generalization across modalities. By relying on annotated angiographies and a limited number of annotated venographies, our framework accomplishes image-to-image translation and semantic segmentation, leveraging a disentangled and semantically rich latent space to represent heterogeneous data and perform image-level adaptation from source to target domains. Moreover, we reduce the typical complexity of cycle-based architectures and minimize the use of adversarial training, which allows us to build an efficient and intuitive model with stable training. We evaluate our method on magnetic resonance angiographies and venographies. While achieving state-of-the-art performance in the source domain, our method attains a Dice score coefficient in the target domain that is only 8.9% lower, highlighting its promising potential for robust cerebrovascular image segmentation across different modalities.

Motivation & Objective

• To address the significant domain shift between MR angiography and venography in brain vessel segmentation.
• To enable semi-supervised segmentation of brain veins using only limited annotated venographies and abundant annotated angiographies.
• To reduce the complexity of cycle-based GAN architectures while maintaining stable and efficient training.
• To disentangle volume-related and vessel-related image properties for accurate, appearance-preserving translation.
• To improve cross-modality generalization in 3D cerebrovascular segmentation despite morphological and intensity distribution differences.

Proposed method

• A two-phase training strategy that first learns domain translation and then performs joint segmentation and adaptation.
• A single generator, discriminator, and encoder architecture to reduce complexity compared to multi-component cycle-GANs.
• A disentangled, semantically rich latent space that separates volume structure from vessel-specific features like intensity and texture.
• Image-level domain adaptation via adversarial training with minimal use of GAN loss to stabilize training.
• End-to-end optimization that jointly learns translation and segmentation using weakly supervised signals from source-labeled and target-unlabeled data.
• Use of StyleGAN2-based latent space for high-fidelity, content-preserving image translation from venography to angiography style.

Experimental results

Research questions

• RQ1 Can a semi-supervised domain adaptation framework effectively transfer knowledge from annotated angiographies to unlabeled venographies for brain vessel segmentation?
• RQ2 How can a disentangled latent space improve cross-modality translation and segmentation in the presence of large domain shifts?
• RQ3 To what extent can a simplified, three-component architecture outperform complex cycle-GANs in 3D medical image domain adaptation?
• RQ4 Does reducing adversarial training complexity lead to more stable and efficient training without sacrificing performance?
• RQ5 Can the method generalize to small, complex structures like brain vessels despite significant differences in vessel morphology and image appearance?

Key findings

• The proposed A2V framework achieves a mean Dice score of 70.4% on venous vessel segmentation, outperforming the second-best method (CS-CADA) by 19.0 percentage points.
• The method maintains high performance on the source domain, achieving a 97.5% Dice score on brain structures, indicating strong generalization from source to target.
• The framework reduces the number of components from multiple generators/discriminators in cycle-GANs to just one generator, one discriminator, and one encoder, improving model efficiency and interpretability.
• The two-phase training strategy minimizes adversarial loss usage, resulting in more stable and faster convergence during training.
• Visual results show that the method preserves spatial structure and vessel topology, generating realistic, artery-like appearances in translated venography images.
• The ablation study confirms that disentangled representation learning is critical for accurate translation and segmentation, especially for small, intricate vessels.