Style Content Decomposition-based Data Augmentation for Domain Generalizable Medical Image Segmentation

TL;DR

Domain shifts in medical imaging degrade segmentation generalization. The paper introduces StyCona, a Sty le-Content decomposition-based data augmentation approach that factorizes an image into global style codes and local content maps via SVD and perturbs them to synthesize diverse training samples. Style augmentation blends style codes while content augmentation mixes content maps, all optimized under a standard segmentation loss on augmented data. Across cross-domain cardiac MRI and multi-target fundus segmentation, StyCona achieves state-of-the-art domain generalization, outperforming Fourier-based, MixStyle-like, and RandConv methods while staying a plug-and-play module for existing segmentation models.

Abstract

Due to domain shifts across diverse medical imaging modalities, learned segmentation models often suffer significant performance degradation during deployment. We posit that these domain shifts can generally be categorized into two main components: 1) "style" shifts, referring to global disparities in image properties such as illumination, contrast, and color; and 2) "content" shifts, which involve local discrepancies in anatomical structures. To address the domain shifts in medical image segmentation, we first factorize an image into style codes and content maps, explicitly modeling the "style" and "content" components. Building on this, we introduce a Style-Content decomposition-based data augmentation algorithm (StyCona), which performs augmentation on both the global style and local content of source-domain images, enabling the training of a well-generalized model for domain generalizable medical image segmentation. StyCona is a simple yet effective plug-and-play module that substantially improves model generalization without requiring additional training parameters or modifications to segmentation model architectures. Experiments on cardiac magnetic resonance imaging and fundus photography segmentation tasks, with single and multiple target domains respectively, demonstrate the effectiveness of StyCona and its superiority over state-of-the-art domain generalization methods. The code is available at https://github.com/Senyh/StyCona.

Figures (4)

• Figure 1: Illustration of style codes and content maps for bSSFP and LGE MRI sequences: a) original images showing disparities in both style (global image appearance) and content (local anatomical structures), b) image histograms reflecting domain shifts in terms of pixel intensities, c) style-swapped images generated by swapping the two original images' style codes (top: bSSFP image with LGE style; bottom: LGE image with bSSFP style), d) content maps of the original images, and e) t-SNE van2008visualizing visualization of domain shifts (top: style shift; bottom: content shift).
• Figure 2: Schematic diagram of the proposed style content decomposition-based data augmentation algorithm (StyCona). StyCona includes three steps: 1) style content decomposition, 2) style augmentation, and 3) content augmentation. $x_i$ and $x_j$ represent original images, and $\bar{\tilde{x}}$ denotes a StyCona-augmented sample.
• Figure 3: Qualitative results on the cardiac MRI segmentation and the OC/OD fundus image segmentation tasks.
• Figure 4: Illustration of a) StyCona-augmented images with different numbers of perturbed content maps ($t=8$, $t=16$, and $t=32$) for two randomly selected original images (The orange dashed circles highlight some content-perturbed regions) and b) Segmentation results for $t=8$, $t=16$, and $t=32$ on cardiac MRI segmentation.