Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Medical Image Segmentation via Single-Source Domain Generalization with Random Amplitude Spectrum Synthesis

Qiang Qiao, Wenyu Wang, Meixia Qu, Kun Su, Bin Jiang, Qiang Guo

TL;DR

Medical image segmentation often suffers from domain shifts under single-source domain generalization (SSDG). This work introduces RAS$^{4}$DG, a framework that combines Random Amplitude Spectrum Synthesis (RASS) with Random Mask Shuffle and Reconstruction (RMS/RSD) to promote domain-invariant learning by perturbing the amplitude spectrum in a frequency-aware manner and enforcing robust structural representations. The amplitude perturbation uses $\delta[m,n,p] \sim \mathcal{N}(1,\sigma^2[m,n,p])$ with $\sigma[m,n,p] = (2\alpha \sqrt{\frac{m^2+n^2+p^2}{H^2+W^2+D^2}})^{\gamma} + \beta$, emphasizing high-frequency components, while reconstructing images via the original phase. Evaluations on 3D fetal brain (Atlases → FeTA2021) and 2D fundus (DRIVE → IOSTAR/LES-AV) show that RAS$^{4}$DG outperforms state-of-the-art SSDG and several DA methods, with ablations indicating RASS as the main contributor and RMS/RSD providing additional gains, thereby narrowing the gap to supervised performance and demonstrating practical potential for privacy-conscious medical image analysis.

Abstract

The field of medical image segmentation is challenged by domain generalization (DG) due to domain shifts in clinical datasets. The DG challenge is exacerbated by the scarcity of medical data and privacy concerns. Traditional single-source domain generalization (SSDG) methods primarily rely on stacking data augmentation techniques to minimize domain discrepancies. In this paper, we propose Random Amplitude Spectrum Synthesis (RASS) as a training augmentation for medical images. RASS enhances model generalization by simulating distribution changes from a frequency perspective. This strategy introduces variability by applying amplitude-dependent perturbations to ensure broad coverage of potential domain variations. Furthermore, we propose random mask shuffle and reconstruction components, which can enhance the ability of the backbone to process structural information and increase resilience intra- and cross-domain changes. The proposed Random Amplitude Spectrum Synthesis for Single-Source Domain Generalization (RAS^4DG) is validated on 3D fetal brain images and 2D fundus photography, and achieves an improved DG segmentation performance compared to other SSDG models.

Medical Image Segmentation via Single-Source Domain Generalization with Random Amplitude Spectrum Synthesis

TL;DR

Medical image segmentation often suffers from domain shifts under single-source domain generalization (SSDG). This work introduces RASDG, a framework that combines Random Amplitude Spectrum Synthesis (RASS) with Random Mask Shuffle and Reconstruction (RMS/RSD) to promote domain-invariant learning by perturbing the amplitude spectrum in a frequency-aware manner and enforcing robust structural representations. The amplitude perturbation uses with , emphasizing high-frequency components, while reconstructing images via the original phase. Evaluations on 3D fetal brain (Atlases → FeTA2021) and 2D fundus (DRIVE → IOSTAR/LES-AV) show that RASDG outperforms state-of-the-art SSDG and several DA methods, with ablations indicating RASS as the main contributor and RMS/RSD providing additional gains, thereby narrowing the gap to supervised performance and demonstrating practical potential for privacy-conscious medical image analysis.

Abstract

The field of medical image segmentation is challenged by domain generalization (DG) due to domain shifts in clinical datasets. The DG challenge is exacerbated by the scarcity of medical data and privacy concerns. Traditional single-source domain generalization (SSDG) methods primarily rely on stacking data augmentation techniques to minimize domain discrepancies. In this paper, we propose Random Amplitude Spectrum Synthesis (RASS) as a training augmentation for medical images. RASS enhances model generalization by simulating distribution changes from a frequency perspective. This strategy introduces variability by applying amplitude-dependent perturbations to ensure broad coverage of potential domain variations. Furthermore, we propose random mask shuffle and reconstruction components, which can enhance the ability of the backbone to process structural information and increase resilience intra- and cross-domain changes. The proposed Random Amplitude Spectrum Synthesis for Single-Source Domain Generalization (RAS^4DG) is validated on 3D fetal brain images and 2D fundus photography, and achieves an improved DG segmentation performance compared to other SSDG models.
Paper Structure (12 sections, 2 equations, 4 figures, 2 tables)

This paper contains 12 sections, 2 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Random Amplitude Spectrum Synthesis
  4. Random Mask Shuffle and Reconstruction
  5. Experiments and Results
  6. Dataset.
  7. Implementation Details.
  8. Comparison with the State-of-the-arts.
  9. Ablation Study
  10. Conclusion
  11. Acknowledgments.
  12. Disclosure of Interests.

Figures (4)

  • Figure 1: Visualization of various results. (a) Results of different methods on the Atlases dataset. (b) 2D datasets and (c) 3D datasets used in our experiments and the statistical analysis of the image amplitude for high frequency (HF) and low frequency (LF).
  • Figure 2: Pipeline of our RAS$^{4}$DG. The bottleneck of the network is shown in right side.
  • Figure 3: Detail view of RSD. $F_{b}$ is reconstructed from spatial and channel dimensions.
  • Figure 4: Visualization of segmentation masks predicted by different methods in different datasets. (a) FeTA2021, (b) IOSTAR, and (c) LES-AV.