PaperlayerView exact revision on arXiv

Source-linked AI summary

CoDG-Net: Structure-Guided Style Diffusion and Collaborative Learning to Mitigate Catastrophic Forgetting in Medical Image Domain Generalization

Yucheng Song, Jincan Wang, Haokang Ding, Zhiqiang Tian, Kangxu Fan, Zhifang Liao

arXiv:2610.05053v1cs.CV

TL;DR

Medical image domain generalization seeks robust segmentation on unseen domains, while catastrophic forgetting of source-domain knowledge remains overlooked and can matter for clinical diagnosis. CoDG-Net combines structure-guided style diffusion with collaborative learning and bias-guided transfer. It outperforms existing state-of-the-art methods on single-source and multi-source datasets while achieving lower source-domain forgetting.

  • Problem

    Medical image domain generalization seeks robust segmentation on unseen domains, but catastrophic forgetting is overlooked and source data cannot be freely transferred across institutions.

    Source: §1 · p. 1, §2.2 · p. 2

  • Method

    CoDG-Net combines structure-guided diffusion on style-related amplitude spectra with a dual-branch network whose bias-guided strategy regulates knowledge transfer across layers and tasks.

    Source: §1 · p. 2

  • Results

    CoDG-Net outperforms existing state-of-the-art methods on single-source and multi-source domain-generalization datasets and achieves a lower source-domain forgetting rate.

    Source: §1 · p. 2, §5 · p. 7

  • Takeaways & Limitations

    The reported approach balances source-domain knowledge retention and segmentation accuracy within a domain-generalization setting that excludes replay and target-domain data.

    Source: §2.2 · p. 2, §5 · p. 7

Abstract

from arXiv · show

Domain Generalization (DG) for medical image segmentation is both highly challenging and critically important. However, existing medical DG methods largely overlook the issue of Catastrophic Forgetting (CF): \textbf{Models often sacrifice their ability to retain source-domain knowledge while pursuing cross-domain robustness.} This can directly threaten diagnostic safety in already-deployed clinical scenarios. To address this, we investigate data augmentation strategies and catastrophic forgetting for medical image DG segmentation. First, we propose a structure-guided style diffusion augmentation method. Constrained by anatomical structure consistency in the frequency domain, this method performs cross-domain diffusion on the amplitude spectrum, generating samples with more diverse and broader style coverage to better support domain generalization. Then, we design a collaborative learning network with a dual-branch interactive architecture (CoDG-Net), together with a novel learning bias-guided strategy that adaptively regulates knowledge transfer at both the layer level and the task level, thereby effectively mitigating catastrophic forgetting on the source domain. Experiments and ablation studies on single-source and multi-source medical DG benchmark datasets demonstrate that CoDG-Net not only outperforms existing state-of-the-art methods in target-domain segmentation performance, but also achieves a lower forgetting rate on the source-domain data. The code is available at: https://github.com/wangprocess/CoDG-Net.

1 Introduction

Medical image domain generalization seeks robust segmentation on unseen domains, but catastrophic forgetting and constraints on sharing source data complicate this goal. CoDG-Net combines structure-guided style diffusion with collaborative learning and bias-guided knowledge transfer to address generalization and source retention.

1 Introduction

  • Medical image domain generalization aims to support robust segmentation across multi-center and multi-device conditions using source-domain data without target-domain access during training.The paper links this capability to diagnostic accuracy and safety, while noting that catastrophic forgetting has been overlooked.Source: §1 · p. 1
  • Privacy and regulatory constraints limit cross-institutional source-data transfer, making replay and mixed-data continual-learning strategies infeasible in the described medical setting.The paper identifies this as an added challenge for mitigating catastrophic forgetting in medical image domain generalization.Source: §1 · p. 1
  • The paper motivates style-focused augmentation because cross-domain images share anatomical structures while differing mainly in style, and domain-generalization constraints can reduce source-domain distribution fit.The described style differences include contrast, texture, and brightness.Source: §1 · pp. 1–2
  • CoDG-Net pairs source-domain image restoration with generalized segmentation and uses learning bias to regulate knowledge transfer across layers and tasks.The strategy is intended to balance source-domain knowledge preservation with target-domain generalization.Source: §1 · p. 2
  • Structure-guided style diffusion generates broader style coverage by diffusing the amplitude spectrum while preserving anatomical structure consistency.The method uses Fourier signals from pseudo-domain images and recombines generated amplitude with the phase spectrum.Source: §1 · p. 2
  • CoDG-Net outperforms existing state-of-the-art methods on single-source and multi-source domain-generalization datasets while achieving a lower source-domain forgetting rate.The reported comparison covers target-domain segmentation performance and source-domain data retention.Source: §1 · p. 2

2 Related Work

Prior medical image domain-generalization work uses augmentation and feature-based strategies to address cross-domain differences. Catastrophic-forgetting approaches include replay-based continual learning, but the paper's DG setting excludes replay and target-domain data.

2.1 Domain Generalization in Medical Imaging

  • Medical image domain-generalization methods use data augmentation to enrich style variation or use invariant-feature extraction, domain alignment, adversarial learning, and normalization strategies.Examples of augmentation include geometric and color perturbations, frequency-domain perturbations, and style transfer.Source: §2.1 · p. 2

2.2 Catastrophic Forgetting in Continual Learning

  • Continual-learning methods mitigate forgetting through regularization, distillation, or replay, while a cited cross-site method uses replay and assumes historical site data remain accessible.The paper's DG setting has no target-domain data, no post-training source-data access, and no task-level replay, motivating CoDG-Net under those constraints.Source: §2.2 · p. 2

3 Method

The method expands source-domain styles while preserving anatomy, then trains a collaborative segmentation–reconstruction network to generalize to unseen domains. Layer-wise bias guidance regulates which features interact across branches.

3.1 Problem Definition and Overview

  • Training uses source-domain samples to target unseen-domain generalization, first exploring pseudo-domain styles and then feeding style-diverse generated data into CoDG-Net.Pseudo-domain exploration uses traditional Bézier-curve intensity mapping before Fourier-domain style expansion.Source: §3.1 · p. 2, Figure 2 · p. 3

3.2 Structure-Guided Style Diffusion

  • Diffusion diversifies the amplitude spectrum under phase-spectrum structural guidance, then recombines generated amplitude with the original phase to form extended-domain images.The phase guidance is intended to preserve critical lesion and organ boundaries while limiting diffusion generation to the amplitude spectrum.Source: §3.2 · pp. 3–4

3.3 Collaborative Learning Network with Dual-Branch Interaction

  • The reconstruction branch learns to map extended-domain samples back to source appearance, and its intermediate features convey style-mapping cues to the segmentation branch.CoDG-Net trains segmentation and reconstruction branches collaboratively on the extended-domain samples.Source: §3.1 · p. 3, §3.3 · p. 4

Layer-Wise Bias-Guided Strategy

  • Layer-wise guidance compares feature evolution across training steps to mark layers collaborative or independent, then dynamically enables or disables their interaction channels.A stochastic perturbation flips the current state with probability δ to avoid remaining in one state for too long.Source: Layer-Wise Bias-Guided Strategy · p. 4

10 Return f (l)

Task-wise bias guidance assesses feature changes along optimization directions and signs an effective distance to indicate whether the update favors reconstruction or segmentation.

Task-Wise Bias-Guided Strategy

  • The task-bias strategy globally regulates optimization because treating reconstruction and segmentation as equal can let reconstruction dominate updates and damage segmentation.The strategy compares feature changes before and after model updates along different directions.Source: Task-Wise Bias-Guided Strategy · p. 4
  • The smaller Euclidean feature distance is treated as the effective optimization distance because it indicates the direction toward which features tend to converge.The distances are used to evaluate feature effectiveness.Source: Task-Wise Bias-Guided Strategy · p. 5
  • A positive signed effective distance marks an update biased toward reconstruction, while a negative value marks one tending away from reconstruction toward segmentation.The sign is assigned according to the optimization objective.Source: Task-Wise Bias-Guided Strategy · p. 5

4 Return f (l)

The loss function combines task-specific objectives with a collaborative feature loss, whose discrepancy is adjusted by the guidance factor.

3.4 Loss Function

  • The collaborative loss measures intermediate-layer feature-map discrepancies and adjusts them using the layer-wise guidance factor.The factor is described as task-wise guidance in the supplied passage.Source: §3.4 · p. 5
  • Segmentation loss combines Dice loss with collaborative loss to support gradient stability and pixel-level segmentation capability.The two loss terms use fusion weights α and β.Source: §3.4 · p. 5
  • Reconstruction loss combines MSE loss with the collaborative loss, using separate fusion weights α′ and β′.The supplied equation gives Lrec = α′ · LMSE + β′ · LCO.Source: §3.4 · p. 5

4 Experiments

Experiments evaluate CoDG-Net on single-source BraTS and multi-source fundus domain generalization, assessing target-domain segmentation and source-domain forgetting. The reported comparisons and ablations examine both generalization and retention.

4.1 Experimental Setup

  • Experiments use BraTS for single-source DG and fundus images for multi-source optic-disc/cup segmentation, assessing target performance and source forgetting.Target performance is measured with Dice, ASD, and HD; SFM compares source-domain performance against an ERM baseline, with values near zero or negative idealized as preserving knowledge or positive transfer. The setup also specifies data preprocessing and model-training configurations.Source: §4.1 · pp. 5–6

4.2 Comparative Experiments

  • CoDG-Net reports excellent target-domain generalization on both single-source and multi-source tasks, with BraTS and fundus comparisons against existing methods.The comparisons are presented in Tables 1 and 2.Source: §4.2 · p. 6, Table 1 · p. 6, Table 2 · p. 6
  • Compared with existing domain-generalization methods, CoDG-Net mitigates source-domain forgetting; Table 3 compares source performance and forgetting on BraTS.The paper also provides a visual comparison of predictions between the T2 source and T1ce target domains.Source: §4.2 · p. 6, Table 3 · p. 6, Figure 4 · p. 7

4.3 Ablation Study and Analysis

  • Dice rises from 29.69 to 67.59 with augmentation variants, while VA4 reaches SFM 5.19, indicating substantial source-domain forgetting.The full model achieves the best performance among the ablations.Source: §4.3 · p. 6, Table 4 · p. 7
  • Without pseudo-domain style exploration, diffusion remains around the source distribution and cannot effectively generate samples from other domains.The ablation identifies the SE module as necessary for producing broader-domain samples.Source: §4.3 · pp. 6–7
  • Unbiased interactions at all intermediate layers provide limited source protection and can reduce target performance; combining layer-wise and task-wise bias yields the best trade-off.The paper reports that each strategy alone falls short of globally optimal performance.Source: §4.3 · p. 7
  • Shallow-layer interactions activate less often than deeper-layer interactions, consistent with the risk that frequent, unguided interactions introduce noisy interference.Figure 5 shows the proportion of training iterations in which each intermediate layer is open.Source: §4.3 · p. 7, Figure 5 · p. 7
  • With T2 as the BraTS source, δ = 0.05 gives the best performance on Flair, T1, and T1CE, and lower across-seed standard deviation.The T2 source-domain performance is not highly sensitive to δ.Source: §4.3 · p. 7, Figure 6 · p. 7

5 Conclusion

CoDG-Net combines structure-guided style diffusion augmentation with bias-guided collaborative learning to address catastrophic forgetting in medical image domain generalization. Experiments show it outperforms existing state-of-the-art approaches on single-source and multi-source benchmarks.

  • CoDG-Net outperforms existing state-of-the-art approaches on both single-source and multi-source domain-generalization benchmarks.Source: §5 · p. 7
  • Structure-guided style diffusion provides pseudo-domain samples with more diverse styles and broader feature-distribution coverage.Source: §5 · p. 7
  • A dual-branch interaction network uses bias-guided learning to regulate knowledge transfer across layers and tasks, balancing source-domain retention with segmentation accuracy.Source: §5 · p. 7
Loading 2610.05053v1…