Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unsupervised Domain Adaptation for Brain Vessel Segmentation through Transwarp Contrastive Learning

Fengming Lin, Yan Xia, Michael MacRaild, Yash Deo, Haoran Dou, Qiongyao Liu, Kun Wu, Nishant Ravikumar, Alejandro F. Frangi

TL;DR

This work tackles cross-modality brain vessel segmentation under unsupervised domain adaptation by introducing a symmetric cross-modality framework that fuses fully-supervised, semi-supervised, and transwarp contrastive learning. The core ideas are Transwarp contrastive learning, which aligns content in the time domain while stabilizing style in the Fourier domain, and Homocentric Squares Domain Adaptation (HSDA), which uses a 3D square Gaussian mask to exchange low-frequency amplitude information. A student–teacher architecture with EMA updates and a Swin-UNet backbone learns domain-invariant features from labeled 3DRA data and unlabeled MRA data, evaluated on Aneurist and SMILE datasets. Results show state-of-the-art performance among UDA methods for cross-modality cerebral vessel segmentation, with ablations confirming the value of each component, though fully supervised baselines remain superior when labels are available; this approach reduces labeling needs and enhances clinical applicability in multi-center vascular imaging.

Abstract

Unsupervised domain adaptation (UDA) aims to align the labelled source distribution with the unlabelled target distribution to obtain domain-invariant predictive models. Since cross-modality medical data exhibit significant intra and inter-domain shifts and most are unlabelled, UDA is more important while challenging in medical image analysis. This paper proposes a simple yet potent contrastive learning framework for UDA to narrow the inter-domain gap between labelled source and unlabelled target distribution. Our method is validated on cerebral vessel datasets. Experimental results show that our approach can learn latent features from labelled 3DRA modality data and improve vessel segmentation performance in unlabelled MRA modality data.

Unsupervised Domain Adaptation for Brain Vessel Segmentation through Transwarp Contrastive Learning

TL;DR

This work tackles cross-modality brain vessel segmentation under unsupervised domain adaptation by introducing a symmetric cross-modality framework that fuses fully-supervised, semi-supervised, and transwarp contrastive learning. The core ideas are Transwarp contrastive learning, which aligns content in the time domain while stabilizing style in the Fourier domain, and Homocentric Squares Domain Adaptation (HSDA), which uses a 3D square Gaussian mask to exchange low-frequency amplitude information. A student–teacher architecture with EMA updates and a Swin-UNet backbone learns domain-invariant features from labeled 3DRA data and unlabeled MRA data, evaluated on Aneurist and SMILE datasets. Results show state-of-the-art performance among UDA methods for cross-modality cerebral vessel segmentation, with ablations confirming the value of each component, though fully supervised baselines remain superior when labels are available; this approach reduces labeling needs and enhances clinical applicability in multi-center vascular imaging.

Abstract

Unsupervised domain adaptation (UDA) aims to align the labelled source distribution with the unlabelled target distribution to obtain domain-invariant predictive models. Since cross-modality medical data exhibit significant intra and inter-domain shifts and most are unlabelled, UDA is more important while challenging in medical image analysis. This paper proposes a simple yet potent contrastive learning framework for UDA to narrow the inter-domain gap between labelled source and unlabelled target distribution. Our method is validated on cerebral vessel datasets. Experimental results show that our approach can learn latent features from labelled 3DRA modality data and improve vessel segmentation performance in unlabelled MRA modality data.
Paper Structure (14 sections, 11 equations, 4 figures, 2 tables)

This paper contains 14 sections, 11 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Method Overview and Problem Formulation
  4. Homocentric Squares Domain Adaptation (HSDA)
  5. Fully-supervised Learning
  6. Semi-supervised Learning
  7. Transwarp contrastive learning
  8. Overall framework and training objective
  9. Experiments
  10. Datasets
  11. Experimental Setup
  12. Quantitative Results
  13. Visual Inspection
  14. Conclusion

Figures (4)

  • Figure 1: Visualization of 3DRA and MRA data reveals significant intra- and inter- domain shifts.
  • Figure 2: Schematic of the proposed method. The method utilizes a composite loss function incorporating fully supervised, semi-supervised, and transwarp Contrastive Learning.
  • Figure 3: Homocentric squares Gaussian kernel $\mathcal{K}_{HSG}$ for image adaptation on 3DRA (source) and MRA (target) vessel patch.
  • Figure 4: Visualisation comparison on MIP maps. Ours shows less over-segmentation on local area.