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

Yusong Xiao; Yuxuan Wu; Li Xiao; Gang Qu; Haiye Huo; Yu-Ping Wang

cs.CVMar 7, 2026

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

Authors: Yusong Xiao, Yuxuan Wu, Li Xiao, Gang Qu, Haiye Huo, Yu-Ping Wang

Organizations: 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

Abstract

Convolutional neural networks (CNNs) have achieved exciting performance in joint segmentation of optic disc and optic cup on single-institution datasets. However, their clinical translation is hindered by two major challenges: limited availability of large-scale, high-quality annotations and performance degradation caused by domain shift during deployment across heterogeneous imaging protocols and acquisition platforms. While unsupervised domain adaptation (UDA) provides a way to mitigate these limitations, most existing approaches do not address cross-domain interference and intra-domain generalization within a unified framework. In this paper, we present the Dual-Domain Synergy UDA (DDS-UDA), a novel UDA framework that comprises two key modules. First, a bi-directional cross-domain consistency regularization module is enforced to mitigate cross-domain interference through feature-level semantic information exchange guided by a coarse-to-fine dynamic mask generator, suppressing noise propagation while preserving structural coherence. Second, a frequency-driven intra-domain pseudo label learning module is used to enhance intra-domain generalization by synthesizing spectral amplitude-mixed supervision signals, which ensures high-fidelity feature alignment across domains. Implemented within a teacher-student architecture, DDS-UDA disentangles domain-specific biases from domain-invariant feature-level representations, thereby achieving robust adaptation to heterogeneous imaging environments. We conduct a comprehensive evaluation of our proposed method on two multi-domain fundus image datasets, demonstrating that it outperforms several existing UDA based methods and therefore providing an effective way for optic disc and optic cup segmentation.

Figures & tables

Fig. 1: Overview of our proposed DDS-UDA architecture, where we use the teacher-student network as our basic network. Specifically, the bi-directional cross-domain consistency pathway (top) employs a dynamic mask generator (left) that facilitates bi-directional feature-level semantic communication between the source and target domains by generating masks from coarse to fine, which are later trained by adding the same masks at loss computation in a consistency regularization manner. Meanwhile, the intra-domain pseudo label learning pathway (bottom) extracts the average amplitude in the source domain data to synthesize stylized but semantically consistent target domain images, which are fed into the teacher model to generate pseudo labels to supervise the original target domain images.

Domain	Domain 1	Domain 2	Domain 3	Domain 4
Domain	Drishti-GS	RIM-ONE-r3	REFUGE (train)	REFUGE (val)
Training	50	99	320	320
Testing	51	60	80	80
Total	101	159	400	400

Table 1: A summary of the Fundus dataset used in this study.

Method	Domain 1				Domain 2
Method	$HD_{OD}$ $\downarrow$	$Dice_{OD}$ $\uparrow$	$HD_{OC}$ $\downarrow$	$Dice_{OC}$ $\uparrow$	$HD_{OD}$ $\downarrow$	$Dice_{OD}$ $\uparrow$	$HD_{OC}$ $\downarrow$	$Dice_{OC}$ $\uparrow$
Source-only	$9.53_{\scriptscriptstyle 0.62}$	$92.88_{\scriptscriptstyle 0.41}$	$14.01_{\scriptscriptstyle 0.74}$	$82.23_{\scriptscriptstyle 0.53}$	$13.58_{\scriptscriptstyle 0.69}$	$87.54_{\scriptscriptstyle 0.58}$	$16.69_{\scriptscriptstyle 0.83}$	$78.69_{\scriptscriptstyle 0.64}$
Target-only	$7.11_{\scriptscriptstyle 0.45}$	$97.15_{\scriptscriptstyle 0.22}$	$10.37_{\scriptscriptstyle 0.56}$	$87.77_{\scriptscriptstyle 0.40}$	$8.71_{\scriptscriptstyle 0.51}$	$94.74_{\scriptscriptstyle 0.33}$	$9.77_{\scriptscriptstyle 0.48}$	$84.08_{\scriptscriptstyle 0.52}$
FDA [ 47 ]	$9.78_{\scriptscriptstyle 0.58}$	$94.89_{\scriptscriptstyle 0.36}$	$14.39_{\scriptscriptstyle 0.71}$	$84.27_{\scriptscriptstyle 0.55}$	$10.83_{\scriptscriptstyle 0.60}$	$89.79_{\scriptscriptstyle 0.49}$	$12.47_{\scriptscriptstyle 0.63}$	$81.11_{\scriptscriptstyle 0.57}$
DoCR [ 19 ]	$7.39_{\scriptscriptstyle 0.47}$	$95.87_{\scriptscriptstyle 0.30}$	$12.24_{\scriptscriptstyle 0.66}$	$85.33_{\scriptscriptstyle 0.49}$	$9.98_{\scriptscriptstyle 0.55}$	$91.64_{\scriptscriptstyle 0.44}$	$11.59_{\scriptscriptstyle 0.58}$	$83.97_{\scriptscriptstyle 0.50}$
DAFormer [ 18 ]	$7.84_{\scriptscriptstyle 0.61}$	$95.12_{\scriptscriptstyle 0.53}$	$12.87_{\scriptscriptstyle 0.66}$	$84.95_{\scriptscriptstyle 0.95}$	$9.96_{\scriptscriptstyle 0.59}$	$95.87_{\scriptscriptstyle 0.80}$	$12.01_{\scriptscriptstyle 0.66}$	$83.81_{\scriptscriptstyle 0.87}$

Table 2: Comparisons with advanced methods on the Fundus dataset. Metrics are reported as mean std over three runs.

Fig. 2: The visualization results produced by different methods on the Fundus dataset. Red and blue colors denote OC and OD, respectively.

Fig. 3: Boxplot of the OC/OD segmentation results produced by the different components in DDS-UDA for the experiment on the Fundus dataset “Domain 1".

Fig. 4: Accumulative bar diagram for OC/OD segmentation performance of different intra- and cross-domain data processing techniques on the Fundus dataset.

Fig. 5: The influence of the tunable parameters, i.e., $\lambda_{\mathcal{S}}$ , $\lambda_{\mathcal{T}}$ , and $\lambda_{\mathcal{T}\text{-stylized}}$ , on the “Domain 1" scenario of the Fundus dataset.

Source Domain	Domain1 (BinRushed)						Domain2 (Magrabia)						Avg.
Target Domain (MESSIDOR)	Base1		Base2		Base3		Base1		Base2		Base3
Method	OC	OD	OC	OD	OC	OD	OC	OD	OC	OD	OC	OD
Source-only	83.01	94.08	83.76	94.96	83.81	93.69	83.13	93.87	84.77	94.01	83.28	94.06	88.87
IntraOnly	82.66	92.39	84.58	95.28	83.31	92.48	82.98	93.74	84.36	93.68	83.13	94.01	88.55
CrossOnly (Uni)	85.61	94.53	85.68	95.67	85.07	94.77	84.32	95.28	86.30	95.36	86.27	94.89	90.31
CrossOnly	87.25	95.21	86.12	95.82	85.82	94.91	84.58	95.87	86.98	95.75	86.93	95.16	90.87

Table 3: Dice results on the RIGA+ dataset produced by different components in DDS-UDA. “Avg." indicates the average of the results produced by the corresponding method under six experiments.

Fig. 6: Impact of the interval bounds of the parameter $\lambda_{k}$ on the model performance.

Domain	Source		Target (MESSIDOR)
Domain	BinRushed	Magrabia	BASE1	BASE2	BASE3
Training	195	95	138	118	106
Testing	0	0	35	30	27
Unlabeled	–	–	227	238	252
Overall	195	95	400	386	385

Table 4: Statistics from the RIGA+ dataset used in this study.

Panel of Table 5 — Table 5: Performance variation with different values of $d_{min}$ and $d_{max}$ .

Method	Domain1 (BinRushed)
Method	Base1				Base2				Base3
	$HD_{OD}\downarrow$	$Dice_{OD}\uparrow$	$HD_{OC}\downarrow$	$Dice_{OC}\uparrow$	$HD_{OD}\downarrow$	$Dice_{OD}\uparrow$	$HD_{OC}\downarrow$	$Dice_{OC}\uparrow$	$HD_{OD}\downarrow$	$Dice_{OD}\uparrow$	$HD_{OC}\downarrow$	$Dice_{OC}\uparrow$
Source-only	$8.21_{\scriptscriptstyle 0.42}$	$94.08_{\scriptscriptstyle 0.26}$	$10.49_{\scriptscriptstyle 0.21}$	$83.01_{\scriptscriptstyle 0.38}$	$8.13_{\scriptscriptstyle 0.50}$	$94.96_{\scriptscriptstyle 0.54}$	$9.94_{\scriptscriptstyle 0.38}$	$83.76_{\scriptscriptstyle 0.36}$	$8.77_{\scriptscriptstyle 0.46}$	$93.69_{\scriptscriptstyle 0.60}$	$10.64_{\scriptscriptstyle 0.13}$	$83.81_{\scriptscriptstyle 0.39}$
Target-only	$4.98_{\scriptscriptstyle 0.21}$	$95.78_{\scriptscriptstyle 0.23}$	$6.03_{\scriptscriptstyle 0.34}$	$86.91_{\scriptscriptstyle 0.25}$	$4.87_{\scriptscriptstyle 0.10}$	$96.31_{\scriptscriptstyle 0.11}$	$7.36_{\scriptscriptstyle 0.21}$	$86.88_{\scriptscriptstyle 0.14}$	$5.35_{\scriptscriptstyle 0.23}$	$95.03_{\scriptscriptstyle 0.24}$	$6.53_{\scriptscriptstyle 0.26}$	$86.17_{\scriptscriptstyle 0.16}$
FDA [ 47 ]	$7.82_{\scriptscriptstyle 0.29}$	$94.98_{\scriptscriptstyle 0.42}$	$9.73_{\scriptscriptstyle 0.27}$	$83.36_{\scriptscriptstyle 0.47}$	$7.47_{\scriptscriptstyle 0.47}$	$95.37_{\scriptscriptstyle 0.20}$	$10.01_{\scriptscriptstyle 0.39}$	$83.77_{\scriptscriptstyle 0.46}$	$8.38_{\scriptscriptstyle 0.43}$	$94.60_{\scriptscriptstyle 0.54}$	$10.45_{\scriptscriptstyle 0.62}$	$84.39_{\scriptscriptstyle 0.45}$
DoCR [ 19 ]	$6.17_{\scriptscriptstyle 0.41}$	$95.34_{\scriptscriptstyle 0.29}$	$8.13_{\scriptscriptstyle 0.48}$	$85.73_{\scriptscriptstyle 0.41}$	$6.39_{\scriptscriptstyle 0.32}$	$95.75_{\scriptscriptstyle 0.26}$	$8.74_{\scriptscriptstyle 0.51}$	$83.80_{\scriptscriptstyle 0.47}$	$6.54_{\scriptscriptstyle 0.43}$	$94.82_{\scriptscriptstyle 0.31}$	$7.84_{\scriptscriptstyle 0.46}$	$86.62_{\scriptscriptstyle 0.38}$

Table 6: Comparisons with other methods on the RIGA+ dataset. Results are reported as mean std over three runs.

Fig. 7: The visualization results produced by different methods on the RIGA+ dataset. Red and blue colors denote OC and OD, respectively.

Explore similar work

Mar 5, 2026

Tell2Adapt: A Unified Framework for Source Free Unsupervised Domain Adaptation via Vision Foundation Model

Source Free Unsupervised Domain Adaptation (SFUDA) is critical for deploying deep learning models across diverse clinical settings. However, existing methods are typically designed for low-gap, specific domain shifts and cannot generalize into a unified, multi-modalities, and multi-target framework, which presents a major barrier to real-world application. To overcome this issue, we introduce Tell2Adapt, a novel SFUDA framework that harnesses the vast, generalizable knowledge of the Vision Foundation Model (VFM). Our approach ensures high-fidelity VFM prompts through Context-Aware Prompts Regularization (CAPR), which robustly translates varied text prompts into canonical instructions. This enables the generation of high-quality pseudo-labels for efficiently adapting the lightweight student model to target domain. To guarantee clinical reliability, the framework incorporates Visual Plausibility Refinement (VPR), which leverages the VFM's anatomical knowledge to re-ground the adapted model's predictions in target image's low-level visual features, effectively removing noise and false positives. We conduct one of the most extensive SFUDA evaluations to date, validating our framework across 10 domain adaptation directions and 22 anatomical targets, including brain, cardiac, polyp, and abdominal targets. Our results demonstrate that Tell2Adapt consistently outperforms existing approaches, achieving SOTA for a unified SFUDA framework in medical image segmentation. Code are avaliable at https://github.com/derekshiii/Tell2Adapt.

Yulong Shi, Shijie Li, Ziyi Li +1

College of Medicine and Biological Information Engineering, Northeastern University · Engineering Research Center of Medical Imaging and Intelligent Analysis, Ministry of Education · Key Laboratory of Medical Image Computing, Ministry of Education, Northeastern University

Source-Free Domain Adaptation Medical Image Segmentation Vision Foundation Model Adaptation

Ask PDF

May 5, 2026

Dual-Foundation Models for Unsupervised Domain Adaptation

Semantic segmentation provides pixel-level scene understanding essential for autonomous driving and fine-grained perception tasks. However, training segmentation models requires costly, labor-intensive annotations on real-world datasets. Unsupervised Domain Adaptation (UDA) addresses this by training models on labeled synthetic data and adapting them to unlabeled real images. While conceptually simple, adaptation is challenging due to the domain gap, i.e., differences in visual appearance and scene structure between synthetic and real data. Prior approaches bridge this gap through pixel-level mixing or feature-level contrastive learning. Yet, these techniques suffer from two major limitations: (1) reliance on high-confidence pseudo-labels restricts learning to a subset of the target domain, and (2) prototype-based contrastive methods initialize class prototypes from source-trained models, yielding biased and unstable anchors during adaptation. To address these issues, we propose a dual-foundation UDA framework that leverages two complementary foundation models. First, we employ the Segment Anything Model (SAM) with superpixel-guided prompting to enable learning from a broader range of target pixels beyond high-confidence predictions. Second, we incorporate DINOv3 to construct stable, domain-invariant class prototypes through its robust representation learning. Our method achieves consistent improvements of +1.3% and +1.4% mIoU over strong UDA baselines on GTA-to-Cityscapes and SYNTHIA-to-Cityscapes, respectively.

Yerin Cheon, Aruna Balasubramanian, Francois Rameau

Stony Brook University, Stony Brook, NY, USA · SUNY Korea, The State University of New York, Korea

Synthetic-to-Real Domain Adaptation Unsupervised Domain Adaptation Semantic Segmentation

Ask PDF

Jun 19, 2026

One Size does not Fit All: Heterogeneous Latent Space Alignment for Unsupervised Domain Adaptation

Domain shift remains a major obstacle to the reliable deployment of machine learning models in high-stakes environments such as healthcare. While Domain adaptation aims to mitigate these effects, existing approaches suffer from limited expressiveness of latent representations and a reliance on handcrafted, static augmentations. In this work, we address these limitations by proposing a novel deep learning architecture for Unsupervised Domain Adaptation (UDA), specifically optimized for medical image segmentation. Our framework, ADualVUOT, integrates a dual-encoder Variational Autoencoder (VAE) with Continuous Normalizing Flows (CNFs) to increase modeling flexibility and posterior expressiveness. To achieve domain alignment, we leverage Unbalanced Optimal Transport (UOT) through the Gaussian-Gromov-Wasserstein (GGW) distance, which handles structural and topological discrepancies between domains. Furthermore, we incorporate an adversarial augmentation scheme to synthesize worst-case compositions, thus enhancing model robustness. Extensive experiments on medical imaging benchmarks show significant gains over prior OT-based approaches.

Evangelia Koskinioti, Yi Shen, Georgios Stamou +1

Department of Electrical and Computer Engineering, Duke University, NC, USA · Department of Mechanical Engineering, Duke University, NC, USA · Department of Electrical and Computer Engineering, National Technical University of Athens, Athens, Greece.

Unsupervised Domain Adaptation Domain Adaptation Medical Image Segmentation

Ask PDF