RaffeSDG: Random Frequency Filtering enabled Single-source Domain Generalization for Medical Image Segmentation

TL;DR

RaffeSDG tackles the problem of domain shifts in medical image segmentation under severe data scarcity by enabling robust out-of-domain inference from a single source. The method leverages random frequency filtering and homologous sample blending to diversify the single-source domain, paired with a structure saliency-guided, self-supervised segmentation framework to learn domain-invariant representations. Extensive experiments across four modalities and three tissues demonstrate superior cross-domain performance relative to DA, DG, and SDG baselines, with ablations confirming the contribution of augmentation, self-supervision, and attention modules. This approach reduces data collection burdens while delivering practical, scalable generalization for clinical imaging tasks, and the authors provide public code for replication and extension.

Abstract

Deep learning models often encounter challenges in making accurate inferences when there are domain shifts between the source and target data. This issue is particularly pronounced in clinical settings due to the scarcity of annotated data resulting from the professional and private nature of medical data. Although various cross-domain strategies have been explored, including frequency-based approaches that vary appearance while preserving semantics, many remain limited by data constraints and computational cost. To tackle domain shifts in data-scarce medical scenarios, we propose a Random frequency filtering enabled Single-source Domain Generalization algorithm (RaffeSDG), which promises robust out-of-domain inference with segmentation models trained on a single-source domain. A frequency filter-based data augmentation strategy is first proposed to promote domain variability within a single-source domain by introducing variations in frequency space and blending homologous samples. Then Gaussian filter-based structural saliency is also leveraged to learn robust representations across augmented samples, further facilitating the training of generalizable segmentation models. To validate the effectiveness of RaffeSDG, we conducted extensive experiments involving out-of-domain inference on segmentation tasks for three human tissues imaged by four diverse modalities. Through thorough investigations and comparisons, compelling evidence was observed in these experiments, demonstrating the potential and generalizability of RaffeSDG. The code is available at https://github.com/liamheng/Non-IID_Medical_Image_Segmentation.

Figures (10)

• Figure 1: Illustration of single-source domain generalization. (c) RaffeSDG enables generalization within (a) a single-source domain, allowing for inference over (b) a range of unknown target domains.
• Figure 2: Overview of RaffeSDG. Random filtering and sample blending are incorporated to introduce randomization into the single-source domain. (a) The source image $x$ is first augmented by random high-pass frequency filters. (b) Sample blending further enables sub-image augmentation by merging filtered samples $\tilde{x}_m$ obtained from $x$. (c) The segmentor integrates Gaussian filters to facilitate the learning of domain-invariant representations, and leverages attention mechanisms to appropriately forward the representations for segmentation.
• Figure 3: Distributions of the single-source and target domains. (a) Distributions of patches from original samples in the feature space. (b) Frequency filtering extended the variability of the single-source domain. (c) Data augmentation achieved by the proposed strategy.
• Figure 4: Frequency-filtered images. (a) Original image. (b) Function diagram and filtered result using an ideal filter. (c) Function diagram and filtered image using a Butterworth filter. (d) Filtered images using channel-wise Butterworth filters. (e) Samples blended from filtered ones.
• Figure 5: Filtering results with varying $D_0$ and $n$, where $r$ represents the radius of the spectrum of an image.