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

Shramana Dey; Varun Ajith; Abhirup Banerjee; Sushmita Mitra

cs.CVMar 30, 2026

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

Authors: Shramana Dey, Varun Ajith, Abhirup Banerjee, Sushmita Mitra

Organizations: Indian Statistical Institute, Kolkata, West Bengal, India · Department of Engineering Science, University of Oxford, Oxford, United Kingdom · Radcliffe Department of Medicine, University of Oxford, Oxford, United Kingdom

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 & tables

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.

Dataset			Size (pixels)	# Samples (Train-Val-Test)	Field-of- View	Camera
Source – Refuge			2124 $\times$ 2056, 1634 $\times$ 1634	840-120-240	40 ∘	Zeiss Visucam 500, Canon CR-2
Target	Drishti-GS		2896 $\times$ 1944	0-0-101	30 ∘	-
	Gamma		2000 $\times$ 2992, 1934 $\times$ 1956	0-0-100	-	KOWA, Topcon TRC-NW400
	Chákṣu	Bosch	1920 $\times$ 1440	0-0-145	40 ∘	Handheld Fundus Camera
		Forus	2048 $\times$ 1536	0-0-126	40 ∘	3Nethra Classic
		Remidio	2448 $\times$ 3264	0-0-1074	40 ∘	Fundus-on-Phone

Table 1: Source and Target Domain datasets.

Dataset		Metric	WaveSDG	BigAug	SLAug	$\mathbf{C^{2}SDG}$	PCSDG	FreeSDG	MoreStyle	EGSDG
Source – Refuge		$DSC_{OD}$	91.21 ± 4.23 ∗	90.77 ± 5.77	85.41 ± 8.05	89.53 ± 5.12	86.23 ± 8.78	85.09 ± 11.36	85.01± 8.92	86.83 ± 7.35
		$DSC_{OC}$	89.76 ± 5.57 ∗	87.31 ± 6.19	84.33 ± 7.92	88.20 ± 6.71	83.75 ± 10.44	84.45 ± 8.19	85.98 ± 7.12	85.70 ± 7.50
		$HD95_{OD}$	3.86 ± 2.20 ∗	4.98 ± 4.93	6.54 ± 3.38	4.02 ± 2.23	6.70 ± 13.91	7.34 ± 13.47	8.03 ± 3.45	5.69 ± 3.10
		$HD95_{OC}$	3.68 ± 1.78 ∗	5.26 ± 4.47	5.49 ± 2.86	4.31 ± 2.38	5.84 ± 2.97	12.11 ± 10.84	5.13 ± 2.34	5.16 ± 2.76
Target	DrishtiGS	$DSC_{OD}$	80.57 ± 9.30	66.87 ± 17.09	76.51 ± 14.71	80.45 ± 10.54	66.95 ± 11.29	35.89 ± 15.47	74.20 ± 11.74	42.16 ± 24.80
		$DSC_{OC}$	82.06 ± 11.59	63.51 ± 25.08	60.51 ± 18.82	80.73 ± 14.33	62.97 ± 25.30	32.57 ± 16.79	72.78 ± 22.56	45.90 ± 24.67

Table 2: Comparative performance of WaveSDG and seven SOTA models, in joint OD/OC segmentation, using Refuge as the source dataset. Metrics reported as mean ± standard deviation, with bold representing best score and underline second-best score.

Figure 3: Visualization of predicted segmentation mask by WaveSDG and seven 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 OD and OC region, is shown in the remaining columns. Note that for SLAug, false positive regions observed outside the zoomed in region in the Bosch, Forus and Remidio datasets are excluded during visualization.

Dataset		WISER	DS	$DSC_{OD}$	$DSC_{OC}$	$HD95_{OD}$	$HD95_{OC}$
Source – Refuge		✗	✗	90.69 ± 4.81	89.31 ± 5.57	3.99 ± 2.29	3.85 ± 1.92
		✓	✗	90.76 ± 4.59	89.68 ± 5.44	3.98 ± 2.20	3.72 ± 1.81
		✓	✓	91.21 ± 4.23	89.76 ± 5.57	3.86 ± 2.20	3.68 ± 1.78
Target	Drishti-GS	✗	✗	69.10 ± 20.24	63.07 ± 27.49	15.33 ± 9.33	16.95 ± 10.59
		✓	✗	72.86 ± 13.43	64.86 ± 24.01	18.07 ± 23.39	20.32 ± 15.58
		✓	✓	80.57 ± 9.30	82.06 ± 11.59	9.66 ± 3.97	11.44 ± 5.83

Table 3: Effect of WISER module and Deep Supervision (DS) on the performance of WaveSDG

Metric	w/o WISER	w WISER
MMD	0.7953	0.6415
JSD	0.2414	0.2288
Fréchet Distance	686.74	530.01

Table 4: Distances between source and target distributions, with and without WISER

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.

Model	Peak Memory (MB)	MACs (G)	# Params (M)
Baseline	438.39	32.4	18.75
WaveSDG	460.40	36.7	21.99
	( $\uparrow$ ) $5.02\%$	( $\uparrow$ ) $13.27\%$	( $\uparrow$ ) 3.24

Table 5: Resource requirement in WISER

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.

Optic Disc Segmentation

See all 8 papers

Frequency Adapter with SAM for Generalized Medical Image Segmentation
May 11, 2026Phuoc-Nguyen Bui, Van-Nguyen Pham, Duc-Tai Le +2Frequency-Domain Feature Learning Domain Generalization
Sungkyunkwan University, Korea
DDS-UDA: Dual-Domain Synergy for Unsupervised Domain Adaptation in Joint Segmentation of Optic Disc and Optic Cup
Mar 7, 2026Yusong Xiao, Yuxuan Wu, Li Xiao +3Medical Image Segmentation Unsupervised Domain Adaptation
MoE Key Laboratory of Brain-Inspired Intelligence Perception and Cognition, University of Science and Technology of China, Hefei 230052, China · Department of Biomedical Engineering, Tulane University, New Orleans, LA 70118, USA · School of Mathematics and Computer Sciences, Nanchang University, Nanchang 330031, China
ROBUST-WT: Robust Uncertainty-aware Segmentation Transform via Whitening and Training Enhancements
Jun 2, 2026Aqsa Naseer, Maryam Bibi, Syeda Samiya Urooj +1Data Augmentation Medical Image Segmentation
School of Electrical Engineering and Computer Science (SEECS), National University of Sciences and Technology (NUST), Islamabad, Pakistan

May 11, 2026J/K move · Enter open · S save

Frequency Adapter with SAM for Generalized Medical Image Segmentation

Phuoc-Nguyen Bui, Van-Nguyen Pham, Duc-Tai Le, Junghyun Bum+1

Sungkyunkwan University, Korea

Frequency-Domain Feature Learning Domain Generalization Vision Foundation Model Adaptation Medical Image Segmentation

Open paper AskPDF arXiv

Medical image segmentation is a critical task in computer-aided diagnosis and treatment planning. However, deep learning models often struggle to generalize across datasets due to domain shifts arising from variations in imaging protocols, scanner types, and patient populations. Traditional domain generalization (DG) methods utilize causal feature learning, adversarial consistency, and style augmentation to improve segmentation robustness. While effective, these approaches rely on explicit feature alignment, adversarial objectives, or handcrafted augmentations, which may not fully exploit the capabilities of foundation models. Recently, the Segment Anything Model (SAM) has demonstrated strong generalization capabilities in segmentation tasks. SAM-based DG methods attempt to improve medical image segmentation. However, these approaches primarily operate in the spatial domain and overlook frequency-based discrepancies that significantly affect model robustness. In this work, we propose Frequency-based Domain Generalization with SAM (FSAM), a novel framework that integrates Low-Rank Adaptation (LoRA) for efficient fine-tuning and a frequency adapter to incorporate frequency-domain representations for single-source domain generalization. FSAM enhances SAM's segmentation robustness by extracting domain-invariant high-frequency features, mitigating frequency-related domain shifts. Experimental results on fundus and prostate datasets demonstrate that FSAM outperforms existing traditional DG and SAM-based DG approaches in domain generalization. Codes and pre-trained models will be made available on GitHub.

Abstract

Optic Disc Segmentation

Explore similar work

Frequency Adapter with SAM for Generalized Medical Image Segmentation

DDS-UDA: Dual-Domain Synergy for Unsupervised Domain Adaptation in Joint Segmentation of Optic Disc and Optic Cup

ROBUST-WT: Robust Uncertainty-aware Segmentation Transform via Whitening and Training Enhancements