O2D-MedNet: An Optimization-Aware Dual-Domain Framework for Robust Medical Image Segmentation

Shipeng Hu; Miaohui Wang; Jing Qin; Sukun Tian; Yuchun Sun

cs.CVOct 26, 2026

O2D-MedNet: An Optimization-Aware Dual-Domain Framework for Robust Medical Image Segmentation

Authors: Shipeng Hu, Miaohui Wang, Jing Qin, Sukun Tian, Yuchun Sun

Organizations: Center of Digital Dentistry, Peking University School and Hospital of Stomatology. · State Key Laboratory of Radio Frequency Heterogeneous Integration and Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University. · School of Nursing, Hong Kong Polytechnic University.

Abstract

Medical image segmentation (MIS) remains challenging due to composite degradations and multi-scale geometric variations, which jointly impair lesion-background separability, structural consistency, and optimization stability. Existing single-stream frameworks entangle heterogeneous cues within a single representation pathway, while hybrid-stream approaches often rely on generic or static fusion, limiting effective coordination across domains and failing to explicitly address optimization instability. To overcome these limitations, we propose a new optimization-aware dual-domain framework for MIS (O2D-MedNet). First, a degradation-aware dualdomain interaction (DDI) module integrates Fourier - and Wavelet-domain features to suppress degradation-sensitive responses while preserving structural and fine-detail information. Second, a cross-backbone contour-guided fusion (CGF) module enables selective interaction between CNN-based local detail modeling and visual state-space-based global structure modeling, improving boundary accuracy and representation consistency. Third, a gradient-context optimization (GCO) strategy stabilizes training by adaptively adjusting learning dynamics based on loss trends and gradient statistics, mitigating instability caused by class imbalance and heterogeneous sample difficulty. Extensive experiments on six benchmarks demonstrate consistent improvements over state-of-the-art MIS methods across multiple imaging modalities. Further analyses confirm enhanced sensitivity, structural consistency, and optimization stability, validating the effectiveness of the proposed O2D-MedNet for robust MIS under heterogeneous conditions.

Figures & tables

Fig. 1: Comparison of representative medical image segmentation (MIS) paradigms. (a) Single-stream frameworks process either original or domain-based inputs through a single pathway, limiting their ability to disentangle heterogeneous cues [ 6 , 7 , 8 ] . (b) Hybrid-stream frameworks introduce multiple backbone branches and fuse their features to exploit complementary representations, but the interaction is typically generic or static [ 9 , 10 ] . (c) In contrast, O2D-MedNet integrates dual-domain refinement, explicit local–global decoupling, contour-guided selective fusion, and optimization-aware learning in a unified framework.

Method	Category	Representation	Local	Global	Domain	Cross-stream	Degradation	Optimization
Method	Category	Representation	Detail	Context	Knowledge	Interaction	Modeling	Awareness
Reference Baselines
DeepLabv3+ [ 16 ]	Single	Atrous CNN + ASPP	Explicit	Explicit	No	No	No	No
U 2 -Net [ 17 ]	Single	Nested U-shaped CNN	Explicit	Explicit	No	No	No	No
MedSAM [ 18 ]	Hybrid	ViT + Box Prompt	Implicit	Explicit	No	Explicit	No	No
Single-stream Frameworks

TABLE I: Taxonomy and comparison of representative medical image segmentation (MIS) frameworks. Existing methods are categorized according to their dominant representation and stream architecture, with emphasis on local-global modeling, domain knowledge, cross-stream interaction, degradation modeling, and optimization awareness.

Fig. 2: Overview of the proposed dual-domain cross-backbone interaction framework for MIS ( O2D-MedNet ). A dual-domain refinement (DDI) module integrates global spectral and local multi-resolution cues to suppress degradations and enhance feature separability. Parallel CNN and VSSM branches then capture local details and global structures, respectively. A contour-guided fusion (CGF) module enables selective cross-backbone interaction in skip connections to improve multi-scale structural consistency and reduce redundant propagation. Finally, a gradient-context optimization (GCO) strategy adapts training based on loss evolution and gradient statistics to improve optimization stability and generalization.

Fig. 3: Illustration of the proposed DDI module. Input features are processed in parallel by Fourier and Wavelet branches to capture global structural and multi-resolution detail cues, respectively. Similarity-aware gating adaptively fuses the responses, followed by cross-branch interaction to enhance spatial complementarity.

Fig. 4: Illustration of the CGF module. CNN and VSSM features are transformed into contour-enhanced responses using Gaussian-based high-pass filtering. A contour-guided gating map enables selective cross-backbone fusion, balancing local details and global structures. The fused features are reinjected into both branches and refined by lightweight local attention to enhance multi-scale structural consistency and suppress redundant responses.

Fig. 5: Foreground-ratio distributions across the six datasets. The evident cross-dataset shifts indicate substantial variation in target-scale statistics.

Group

Method

Optical RGB Dataset

Radiological Dataset

Ultrasound Dataset

Oral

Kvasir-SEG

COVID-19

MLUA

CT2US

BUSI

mIoU

\uparrow

Dice

\uparrow

mIoU

\uparrow

Dice

\uparrow

mIoU

\uparrow

Dice

\uparrow

mIoU

\uparrow

Dice

\uparrow

mIoU

\uparrow

Dice

\uparrow

mIoU

\uparrow

Dice

\uparrow

Reference Baselines

U 2 -Net [ 17 ]

82.3

\pm

3.1

90.3

\pm

1.7

78.5

\pm

3.8

88.0

\pm

3.2

79.4

\pm

3.8

88.5

\pm

3.3

59.5

\pm

5.8

68.5

\pm

5.2

95.3

\pm

4.0

97.6

\pm

2.7

61.9

\pm

4.1

76.4

\pm

3.6

DeepLabV3+ [ 16 ]

81.6

\pm

2.3

89.8

\pm

1.4

50.1

\pm

6.7

66.8

\pm

6.1

78.3

\pm

4.6

87.8

\pm

4.3

59.1

\pm

4.8

74.3

\pm

4.6

95.1

\pm

3.9

97.5

\pm

1.7

61.1

\pm

5.6

75.8

\pm

5.6

SFM [ 57 ]

70.0

\pm

3.0

82.4

\pm

2.4

68.8

\pm

5.1

81.5

\pm

4.7

75.4

\pm

5.3

85.9

\pm

3.2

45.6

\pm

6.2

62.6

\pm

5.6

94.4

\pm

2.1

97.1

\pm

1.4

58.7

\pm

3.9

74.0

\pm

3.6

TherNet [ 58 ]

82.5

\pm

2.9

90.4

\pm

1.9

74.4

\pm

3.8

85.3

\pm

2.9

78.3

\pm

2.9

87.8

\pm

2.5

46.9

\pm

6.7

63.8

\pm

6.7

95.0

\pm

4.1

97.4

\pm

2.2

50.9

\pm

6.8

67.4

\pm

6.4

TABLE II: Quantitative comparison in terms of mIoU and Dice on the Oral , Kvasir-SEG , COVID-19 , MLUA , CT2US , and BUSI datasets. Comparison methods are grouped into Reference Baselines , Single-stream Frameworks , and Hybrid-stream Frameworks . The best and second-best results in each column are highlighted in red and blue , respectively.

Fig. 6: Qualitative comparison across different modalities of medical imaging data. For representative coverage, the competitive methods consist of Reference Baselines , Single-stream Frameworks , and Hybrid-stream Frameworks . White indicates correct overlap, red indicates false positives, and green indicates false negatives.

Panel of Table 9 — TABLE III: Ablation study on the Oral and COVID-19 datasets. Higher values indicate better performance for mIoU , Dice , and Sensitivity (‘Sens.’), while lower values indicate better performance for Betti Error (BE).

Fig. 7: Qualitative ablation results on representative samples from Oral and COVID-19 datasets. The predictions show progressively improved completeness and structural consistency.

Fig. 8: Effectiveness of DDI. Comparison of local ROIs before and after DDI, along with Sobel gradient distributions and scanline intensity profiles. DDI suppresses spurious high-frequency responses while preserving true boundary structures, thereby improving lesion-background separability.

Fig. 9: Effectiveness of CGF. Comparison with DFM , CRIMs , CRFM , and CGAF on Oral and COVID-19 . CGF achieves the best performance in both overlap accuracy and topological consistency.

Fig. 10: Effectiveness of GCO. Prediction evolution across training epochs on representative samples from Oral and COVID-19 datasets.

Fig. 11: Activation responses of the proposed local–global decoupling and scale-wise interaction. Compared with the CNN and VSSM branches, the fused branch produces more complete and lesion-aligned activations during decoding, indicating effective integration of local detail and global structure.

Method	Params (M) $\downarrow$	GFLOPs $\downarrow$	Time (msec.) $\downarrow$
U 2 -Net	1.131	9.751	17.047
DeepLabV3+	40.347	13.268	7.433
SFM	25.874	7.871	19.055
TherNet	48.072	27.835	24.513
MWCNN	24.929	51.907	4.619
MedSAM	90.689	35.565	18.478

TABLE IV: Computational complexity and inference efficiency under a unified input resolution and hardware setting. Lower values indicate better performance. Best results are highlighted in bold .

Fig. 12: Cross-backbone generalization of GCO on the Oral and COVID-19 datasets.

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 788 papers

Med-DisSeg: Dispersion-Driven Representation Learning for Fine-Grained Medical Image Segmentation
May 14, 2026Zhiquan Chen, Haitao Wang, Guowei Zou +1Image Segmentation Metric Learning
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, 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
When Fusion Fails: Corruption-Aware Rebalanced Fusion for Multi-Modal Medical Image Segmentation
Sep 9, 2026Yuchen Pei, Xiaoyu Hu, Yixiong Zou +5Cross-Modality Medical Image Segmentation Multi-Source Image Fusion
Central China Normal University Wuhan, China · Huazhong University of Science and Technology Wuhan, China · Guangzhou University of Chinese Medicine Guangzhou, China +2

May 14, 2026J/K move · Enter open · S save

Med-DisSeg: Dispersion-Driven Representation Learning for Fine-Grained Medical Image Segmentation

Zhiquan Chen, Haitao Wang, Guowei Zou, Hejun Wu

School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, China

Image Segmentation Metric Learning Medical Image Segmentation Medical Image Representation Learning

Open paper AskPDF arXiv

Accurate medical image segmentation is fundamental to precision medicine, yet robust delineation remains challenging under heterogeneous appearances, ambiguous boundaries, and large anatomical variability. Similar intensity and texture patterns between targets and surrounding tissues often lead to blurred activations and unreliable separation. We attribute these failures to representation collapse during encoding and insufficient fine grained multi scale decoding. To address these issues, we propose Med DisSeg, a dispersion driven medical image segmentation framework that jointly improves representation learning and anatomical delineation. Med DisSeg combines a lightweight Dispersive Loss with adaptive attention for fine grained structure segmentation. The Dispersive Loss enlarges inter sample margins by treating in batch hidden representations as negative pairs, producing well dispersed and boundary aware embeddings with negligible overhead. Based on these enhanced representations, the encoder strengthens structure sensitive responses, while the decoder performs adaptive multi scale calibration to preserve complementary local texture and global shape information. Extensive experiments on five datasets spanning three imaging modalities demonstrate consistent state of the art performance. Moreover, Med DisSeg achieves competitive results on multi organ CT segmentation, supporting its robustness and cross task applicability.

Abstract

Medical Image Segmentation

Explore similar work

Med-DisSeg: Dispersion-Driven Representation Learning for Fine-Grained Medical Image Segmentation

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

When Fusion Fails: Corruption-Aware Rebalanced Fusion for Multi-Modal Medical Image Segmentation