Decoupling Wavelet Sub-bands for Single Source Domain Generalization in Fundus Image Segmentation

Abstract

Domain generalization in fundus imaging is challenging due to variations in acquisition conditions across devices and clinical settings. The inability to adapt to these variations causes performance degradation on unseen domains for deep learning models. Besides, obtaining annotated data across domains is often expensive and privacy constraints restricts their availability. Although single-source domain generalization (SDG) offers a realistic solution to this problem, the existing approaches frequently fail to capture anatomical topology or decouple appearance from anatomical features. This research introduces WaveSDG, a new wavelet-guided segmentation network for SDG. It decouples anatomical structure from domain-specific appearance through a wavelet sub-band decomposition. A novel Wavelet-based Invariant Structure Extraction and Refinement (WISER) module is proposed to process encoder features by leveraging distinct semantic roles of each wavelet sub-band. The module refines low-frequency components to anchor global anatomy, while selectively enhancing directional edges and suppressing noise within the high-frequency sub-bands. Extensive ablation studies validate the effectiveness of the WISER module and its decoupling strategy. Our evaluations on optic cup and optic disc segmentation across one source and five unseen target datasets show that WaveSDG consistently outperforms seven state-of-the-art methods. Notably, it achieves the best balanced Dice score and lowest 95th percentile Hausdorff distance with reduced variance, indicating improved accuracy, robustness, and cross-domain stability.

Figures (5)

• Figure 1: Representative wavelet decomposition of (a) Fundus image, with (b) extracted Optic Disc-Optic Cup regions, and its corresponding representation in wavelet sub-bands (c) LL, (d) LH, (e) HL, and (f) HH.
• Figure 2: The proposed WaveSDG architecture. The lower and upper rectangles illustrate the overall network and the novel WISER module, respectively. The model is trained exclusively on the source dataset, while the target dataset is used only during evaluation. The WISER module filters the encoder features prior to it being passed to the decoder. Dotted rectangles highlight the internal components of the WISER module. Solid lines denote forward propagation path and the dotted lines represent the point from which gradients flow during back-propagation starts.
• Figure 3: Visualization of predicted segmentation mask by WaveSDG and six other competing methods, along with the corresponding ground truth. Green represents the predicted OD region, while blue shows predicted OC region. Zoomed view of the cropped region (in white) in the original image, containing the entire predicted region, is shown in the remaining columns.
• Figure 4: Visualization of the effect of various components in WISER. (a) Feature channel and its wavelet-decomposed sub-bands, prior to the WISER module. (b) Intermediate computation of the scaling factor $\mathcal{E}_1^S$, normalized edge energy $\mathcal{E}_1^{norm}$, boosting factor $\mathcal{E}_1^{eff}$, and their element-wise product. (c) Corresponding WISER-transformed feature channel and filtered sub-bands.
• Figure 5: Visualization of LL sub-band decomposition, with row (a) showing column-wise, (i) query image, with its decoupled (ii) anatomy and (iii) style components. Row (b) displays the respective images, nearest to the query image, in the content embedding space. Row (c) refers to the corresponding images, nearest to the query image, in the style embedding space.