Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Francesco Galati, Daniele Falcetta, Rosa Cortese, Barbara Casolla, Ferran Prados, Ninon Burgos, Maria A. Zuluaga

TL;DR

While achieving state-of-the-art performance in the source domain, the 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.

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.

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

TL;DR

While achieving state-of-the-art performance in the source domain, the 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.

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.
Paper Structure (7 sections, 4 figures, 1 table)

This paper contains 7 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Experiments and Results
  5. Experimental Setup
  6. Results
  7. Discussion and Conclusion

Figures (4)

  • Figure 1: Maximum intensity projection (MIP) of a magnetic resonance angiography (left) and minimum intensity projection (mIP) of a magnetic resonance venography (right) along the three spatial axes.
  • Figure 2: Two-phase training algorithm. Images $x_i$ from domains $\mathcal{S}$ and $\mathcal{T}$ are fed into the model, composed of the generator $G$, the discriminator $D$ and the encoder $E$. The modules are trained in two separate phases. Phase 1 (top) trains $G$ in an adversarial fashion to learn a smooth and semantically rich latent space. Phase 2 (bottom) trains $E$ to perform image-to-image translation and refines $G$ to also generate segmentation masks $\hat{y}_i^t$ and $\hat{y}_i^s$.
  • Figure 3: Phase 2 of the training algorithm alternating intra-domain (2.1) and inter-domain (2.2 and 2.3) configurations. We compute the sum of mean squared error and LPIPS LPIPS for reconstruction and cycle-consistency losses $\mathcal{L}_{\textrm{r}}$, and the sum of Dice and cross-entropy for segmentation losses $\mathcal{L}_{\textrm{s}}$. The backpropagation of $\mathcal{L}_{\textrm{r}}$ updates only the weights of $E$, while $\mathcal{L}_{\textrm{s}}$ affects both $E$ and $G$.
  • Figure 4: Visual comparison of the results produced by different methods for brain and vein segmentation from SWI images. Brain masks are indicated in red, vessels in green. From the 1st to the 3rd row we display in order a top-level, a middle-level and a bottom-level slice.