PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.35220v1cs.CV

TL;DR

Medical image segmentation faces composite degradations and multi-scale geometric variation that challenge lesion–background separability, structural consistency, and optimization. O2D-MedNet coordinates dual-domain refinement, selective cross-backbone fusion, and gradient-context optimization. The paper reports consistent improvements across six benchmark datasets and identifies limits in dimensional scope, resolution, computational cost, and clinical validation.

  • Problem

    Composite degradations and multi-scale geometric variation challenge lesion–background separability, structural consistency, and optimization, while existing frameworks often couple these concerns.

    Source: §1 · pp. 1–2

  • Method

    O2D-MedNet combines Fourier- and Wavelet-domain refinement, contour-guided fusion of CNN local and state-space global features, and gradient-context optimization.

    Source: §1 · p. 2

  • Results

    Across six benchmark datasets, O2D-MedNet demonstrates consistent improvements in segmentation accuracy, sensitivity, and topological consistency.

    Source: §5 · p. 14

  • Takeaways & Limitations

    The reported results support O2D-MedNet for robust segmentation across multiple imaging modalities and anatomical targets.

    Source: §5 · p. 14

    The study is restricted to 2D tasks; extending to 3D and temporal imaging requires further investigation, and clinical validation remains necessary.

    Source: §5 · p. 14

Abstract

from arXiv · show

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.

1 INTRODUCTION

Medical image segmentation must handle composite degradations and multi-scale anatomical variation, which challenge separability, structural consistency, and optimization. O2D-MedNet coordinates degradation suppression, local–global representation learning, and training dynamics, with reported improvements across six benchmark datasets.

1 INTRODUCTION

  • Composite degradations and multi-scale geometric variability can jointly impair lesion–background separability, structural consistency, and optimization stability.These conditions also make it difficult to preserve fine-grained boundaries, maintain global coherence, and generalize across heterogeneous samples.Source: §1 · p. 1
  • Existing approaches can couple degradation suppression, representation learning, and optimization, while single-stream pathways and static hybrid fusion limit adaptive coordination.Under composite degradations, anatomical variation, class imbalance, and heterogeneous sample difficulty, this coupling is associated with performance deterioration.Source: §1 · pp. 1–2
  • O2D-MedNet combines Fourier- and Wavelet-domain refinement, contour-guided CNN–state-space fusion, and gradient-context optimization to coordinate degradation suppression, representation learning, and training dynamics.Its stated aims include preserving structural integrity and fine details, improving structural consistency, and stabilizing training under class imbalance and heterogeneous sample difficulty.Source: §1 · p. 2
  • Experiments across six benchmark segmentation datasets spanning three imaging categories and diverse lesions and anatomical structures demonstrate the framework’s effectiveness and generalization.The paper’s conclusion reports consistent improvements in segmentation accuracy, sensitivity, and topological consistency across multiple imaging modalities and anatomical targets.Source: §1 · p. 2

2 RELATED WORK

The related work reviews reference baselines, single-stream approaches, and hybrid-stream frameworks. It identifies limited adaptive coordination of degradation handling, local–global interaction, and optimization as a shared gap.

2.1 Reference Baselines

  • Reference baselines span conventional encoder–decoder, nested multi-scale, and promptable foundation-model paradigms, represented by DeepLabv3+, U2-Net, and MedSAM.They respectively use atrous spatial pyramid pooling with an encoder–decoder, nested residual Ublocks, and a ViT image encoder with prompt encoder and mask decoder.Source: §2.1 · p. 3

2.2 Single-stream Frameworks

  • Single-stream methods include CNN/U-shaped, Transformer, VSSM, and domain-based models, but their single pathways entangle local detail, global structure, and degradation handling.The section describes trade-offs between local sensitivity and global coherence, and notes that domain cues are often combined with fixed or weakly adaptive fusion.Source: §2.2 · p. 3

2.3 Hybrid-stream Frameworks

  • Hybrid-backbone methods combine CNN local modeling with Transformer-style global modeling, yet often depend on static or heuristic fusion without condition-dependent prioritization.The section also notes that hybrid-frequency approaches may not coordinate backbone biases or address optimization instability, while hybrid-task interactions can be task-dependent.Source: §2.3 · pp. 3–4
  • Hybrid-stream frameworks broaden representation diversity but generally lack a unified adaptive mechanism coordinating degradation suppression, cross-stream interaction, and optimization dynamics.Their fusion is often static or weakly adaptive, and optimization may remain external to the architecture.Source: §2.3 · p. 4

3 METHODOLOGY

O2D-MedNet uses dual-domain refinement before parallel CNN and VSSM encoding, with contour-guided selective fusion between local and global representations. During training, gradient-context optimization adapts learning behavior to improve stability.

3 METHODOLOGY

  • The DDI module first processes complementary Fourier- and Wavelet-domain information to enhance feature discriminability.Refined features then enter parallel CNN and VSSM branches, where CGF selectively fuses local and global representations.Source: §3 · p. 4
  • CGF enables selective fusion between local and global representations encoded by parallel CNN and VSSM branches.The architecture description places this fusion after DDI refinement.Source: §3 · p. 4
  • During training, GCO dynamically adapts optimization behavior according to gradient context to improve optimization stability.The framework description presents GCO as its training-stage component.Source: §3 · p. 4

3.1 Degradation-Aware Dual-Domain Interaction

DDI refines features before representation learning, combining Fourier- and Wavelet-domain responses and then using interactive modulation to reduce degradation-induced responses. Its branches provide complementary global structural and multi-resolution local-detail cues.

3.1 Degradation-Aware Dual-Domain Interaction

  • DDI refines features before high-level modeling because degradation-induced responses can blur lesion–background boundaries and destabilize downstream learning.The paper states these responses cannot be reliably suppressed by representation learning alone.Source: §3.1 · p. 4

3.1.1 Dual-Domain Feature Representation

  • Fourier processing targets unstable high-frequency responses and structural consistency, while Wavelet processing preserves multi-resolution local details and boundary-related subbands.The Wavelet branch modulates subband coefficients to suppress degradation-sensitive weak coefficients and enhance boundary-related subbands across scales.Source: §3.1.1 · p. 5

3.1.2 Dual-Domain Interactive Modulation

  • Bidirectional modulation lets regional responses suppress spurious local peaks while preserving boundary-sensitive cues, producing a spatial gate that reduces degradation-induced responses.A sample-level gate estimates global reliability, while spatial modulation refines where dual-domain information is emphasized within each image.Source: §3.1.2 · pp. 5–6

3.2 Cross-Backbone Contour-Guided Fusion

CGF selectively fuses CNN local-detail features and VSSM global representations using contour-sensitive responses. It then refines and reinjects features across both branches to support local consistency and multi-scale structural refinement.

3.2 Cross-Backbone Contour-Guided Fusion

  • Directly fusing CNN and VSSM features assumes compatibility, which may be violated under complex degradations and scale variations.CNN features capture local textures and boundary details; VSSM features model long-range dependencies and global semantics.Source: §3.2 · p. 6

3.2.1 Contour-Guided Selective Fusion

  • Gaussian-residual responses suppress smooth background components and highlight boundary-relevant structures before contour-guided gating selectively fuses CNN and VSSM features.The gating map indicates where to rely more on CNN local details or VSSM global structure.Source: §3.2.1 · p. 6

3.2.2 Local-Consistency Feature Refinement

  • Shift-based local refinement and local self-attention jointly estimate a consistency modulation that recalibrates both branches before their features are combined.Stage-specific refinement features are injected into corresponding encoder and decoder stages through residual connections.Source: §3.2.2 · p. 7

3.3 Stable Gradient-Context Optimization

GCO treats learning-rate scheduling as an adaptive decision process, using loss and gradient context to choose learning rates and update a Q-table online. Its state, action, and reward designs connect optimization behavior to subsequent learning-rate decisions.

3.3 Stable Gradient-Context Optimization

  • GCO adapts learning rates across epochs through a controller that observes the current state, selects an action, receives a reward, and updates its Q-table online.The Q-table is initialized from scratch rather than pretrained on separate tasks.Source: §3.3 · p. 7, §3.3.3 · p. 8

3.3.1 Optimization State Construction

  • The optimization state uses gradient-norm stability, loss progress, and recent learning-rate history to represent training context.Within an epoch, GCO computes the mean and standard deviation of mini-batch gradient norms; a larger stability score indicates more consistent update magnitudes.Source: §3.3.1 · p. 7

3.3.2 Action-based Learning Rate Adjustment

  • An ε-greedy policy selects a learning-rate action from the Q-table, mapping actions to rates in logarithmic space across the allowable interval.The action selection balances random exploration with choosing the action having the highest Q-value.Source: §3.3.2 · p. 8

3.3.3 Reward Design and Online Q-table Update

  • The reward combines recent loss trend and short-term loss reduction with a penalty for unstable gradient context, guiding online Q-table updates.The Q-table uses the next state's estimated future value, the update rate, and the discount factor; exploration decays to a minimum rate.Source: §3.3.3 · p. 8

4 EXPERIMENTAL VALIDATIONS

The experiments evaluate O2D-MedNet across six datasets spanning optical, radiological, and ultrasound imaging, using overlap, detection, and topology metrics. Comparisons cover reference, single-stream, and hybrid-stream methods under shared evaluation settings.

4.1.1 Datasets

  • The evaluation spans six datasets across optical RGB, radiological, and ultrasound imaging, including diverse lesions and anatomical structures.The datasets are Oral, Kvasir-SEG, COVID-19, MLUA, CT2US, and BUSI.Source: §4.1.1 · p. 8

4.1.2 Evaluation Metrics

  • Evaluation uses Dice, mIoU, Sensitivity, and Betti Error to measure overlap, foreground detection, and topological consistency.For multiclass tasks, Dice and mIoU are macroaveraged over foreground classes; for binary tasks, all metrics are computed on the foreground class.Source: §4.1.2 · p. 9

4.1.3 Comparison Methods

  • Comparisons group representative methods into reference baselines, single-stream frameworks, and hybrid-stream frameworks.The compared methods include conventional segmentation architectures, CNN-, Transformer-, VSSM-, and frequency-enhanced models, and hybrid architectures.Source: §4.1.3 · p. 9

4.1.4 Implementation Details

  • All datasets use five-fold cross-validation, 224 × 224 inputs, and 500 training epochs; GCO adaptively controls the learning-rate schedule.The base optimizer is Adam, and images are converted to three channels when necessary and normalized to [0, 1].Source: §4.1.4 · pp. 9–10
  • Comparison methods are evaluated on the six benchmarks using mIoU and Dice, with official protocols followed where available and otherwise unified settings.Unified settings include data splits, input resolutions, and optimization schedules.Source: §4.1.4 · p. 10

4.2 Comparison Performance

O2D-MedNet achieves the highest mIoU and Dice across six heterogeneous datasets, with improvements over each dataset’s strongest competing method. Qualitative comparisons report cleaner boundaries, fewer false positives, and fewer missed regions across modalities.

4.2.1 Quantitative Comparison

  • O2D-MedNet achieves the highest mIoU and Dice on all six datasets, outperforming the highest-competing method on each.Its mIoU/Dice gains are 5.0/3.8 on Oral, 3.4/2.0 on Kvasir-SEG, 3.2/1.9 on COVID-19, 3.8/2.8 on MLUA, 0.5/0.2 on CT2US, and 2.2/2.9 on BUSI.Source: §4.2.1 · pp. 10–11

4.2.2 Qualitative Comparison

  • Across modalities, O2D-MedNet produces cleaner boundaries, fewer false positives, and fewer missed regions than the compared methods.Reported qualitative patterns include improved lesion-background separation for optical images, reduced fragmentation and omission in radiological images, and more continuous ultrasound boundaries under speckle noise.Source: §4.2.2 · p. 11

4.3 Ablation Studies

The component ablations on Oral and COVID-19 show successive improvements in overlap and sensitivity metrics alongside reductions in Betti Error as the model is built from a CNN baseline to the full system. Qualitative results describe progressively more complete, structurally consistent, and compact predictions.

4.3.1 Quantitative Analysis

  • Adding components from the CNN baseline through the full model consistently improves mIoU and Dice while reducing Betti Error on Oral and COVID-19.The ablation pipeline adds the VSSM branch, CGF, DDI, and GCO in sequence.Source: §4.3.1 · p. 11
  • The VSSM branch reduces Betti Error from 13.0 to 10.3 on Oral and from 15.8 to 11.4 on COVID-19, while Oral sensitivity falls from 92.3 to 91.1.The reported results associate the branch with improved global structural coherence, while noting that global modeling alone is insufficient for stable detection of subtle lesions.Source: §4.3.1 · p. 11
  • After CGF, Oral Betti Error decreases to 7.9 and sensitivity recovers to 93.3.The authors describe CGF as coordinating local and global features.Source: §4.3.1 · p. 11
  • Adding DDI raises COVID-19 Dice from 88.9 to 90.1 and reduces Betti Error from 8.8 to 5.8.The reported changes are consistent with DDI suppressing degradation-sensitive responses on composite distortions.Source: §4.3.1 · p. 11
  • GCO reduces Oral Betti Error from 5.3 to 3.4 and raises sensitivity from 93.2 to 95.5.The authors attribute its contribution mainly to alleviating under-convergence and training oscillation, improving recall and structural stability.Source: §4.3.1 · p. 11

4.3.2 Qualitative Analysis

  • Qualitatively, the full model yields the most balanced predictions, preserving lesion structures while reducing lesion-background interference in oral and CT cases.The sequence progresses from fragmented CNN-only predictions to more coherent structures with the VSSM branch, more stable interfaces with CGF, and cleaner boundary responses with DDI.Source: §4.3.2 · p. 11

4.4 Effectiveness Analysis

The effectiveness analysis examines how DDI refines lesion responses, how CGF coordinates local and global features, and how GCO stabilizes predictions during training. Results show improved separability and fusion performance, alongside more stable segmentation predictions with GCO.

4.4.1 DDI Module

  • DDI reduces surrounding-region activation while preserving lesion contours, smoothing response profiles and improving boundary-band separation.The changes accompany consistent increases in AUC and CNR, supporting improved lesion-background separability.Source: §4.4.1 · p. 12

4.4.2 CGF Module

  • CGF achieves the strongest reported overlap and topology metrics on Oral and COVID-19 among the compared fusion modules.On Oral, CGF records Dice 96.8, Sensitivity 95.5, mIoU 91.9, and Betti Error 3.4; on COVID-19, it records Dice 90.4, Sensitivity 89.9, mIoU 82.6, and Betti Error 4.0.Source: Figure 9 · p. 12, §4.4.2 · p. 12

4.4.3 GCO Module

  • GCO produces more stable prediction evolution across epochs, preserving lesion localization and major structures in Oral and COVID-19 samples.Without GCO, predictions fluctuate in lesion size and location; with GCO, predictions remain more consistent with ground truth across the sampled epochs.Source: §4.4.3 · pp. 12–13, Figure 10 · p. 12

4.4 Effectiveness Analysis

  • The fused branch combines local detail and global structure into more complete, lesion-aligned activations during decoding.The analysis describes this as evidence of effective integration of the CNN and VSSM branches.Source: Figure 11 · p. 13

4.5 Activation Response Analysis

Activation analysis describes complementary CNN and VSSM responses and lesion-aligned fused features. Complexity results report O2D-MedNet’s compute and latency, while cross-backbone tests show GCO improving mIoU across evaluated architectures.

4.5 Activation Response Analysis

  • CNN activations emphasize local textures and boundary cues, VSSM activations capture broader context, and fused features preserve lesion responses while suppressing irrelevant ones.The reported activation patterns indicate that fusion combines local discriminability with contextual modeling for segmentation.Source: §4.5 · p. 13

4.6 Computational Complexity

  • O2D-MedNet uses 158.31M parameters, 12.533 GFLOPs, and 37.902 milliseconds per image, with lower GFLOPs than several high-capacity baselines.The authors characterize its performance-efficiency balance as favorable despite added parameters and latency from the dual-backbone design.Source: §4.6 · p. 13, Table 4 · p. 13

4.7 Further Cross-Architecture Validation of GCO

  • GCO improves mIoU across six CNN-, Transformer-, and VSSM-based backbones, averaging gains of 3.2 points on Oral and 2.1 on COVID-19.The authors report that these gains are not tied to a specific network architecture.Source: §4.7 · p. 14, Figure 12 · p. 14

5 CONCLUSION AND FUTURE WORK

O2D-MedNet coordinates dual-domain feature refinement, contour-guided cross-backbone interaction, and gradient-context optimization for medical image segmentation. Experiments across six datasets show consistent improvements, while the current study’s scope and design retain practical limitations.

5 CONCLUSION AND FUTURE WORK

  • Across six benchmark datasets, O2D-MedNet improves segmentation accuracy, sensitivity, and topological consistency across imaging modalities and anatomical targets.The conclusion attributes the reported effectiveness to jointly modeling feature representation and optimization dynamics.Source: §5 · p. 14
  • The evaluation is limited to 2D segmentation, and fixed-resolution resizing may compromise extremely fine-grained structures.The dual-backbone design also adds parameters and inference latency compared with lightweight architectures; further clinical validation is needed.Source: §5 · p. 14
  • Future work will extend the framework to 3D and longitudinal imaging, improve robustness under limited annotations and cross-center domain shifts, and incorporate uncertainty-aware modeling.Source: §5 · p. 14

Related papers

Loading 2610.35220v1…