MDCL-UNet: A Multi-Domain Collaborative Learning Method for Medical Image Segmentation

Engineering notes

Extensive experiments on retinal vessel, optic disc/cup, and abdominal multi-organ segmentation datasets demonstrate that MDCL-UNet achieves consistently higher Dice scores and lower HD95 distances than single-dataset baselines and existing cross-dataset collaborative learning methods. The source code will be available at https://github.com/lqr41710085/MDCL-UNET .

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Medical image segmentation, which delineates anatomical structures and pathological regions at the pixel level, plays an important role in computer-aided diagnosis. Existing segmentation methods are typically developed and evaluated under a single-dataset setting, and their performance often degrades when applied to images from different medical centers, scanners, or acquisition protocols due to substantial domain gaps. While cross-dataset collaborative learning methods can train a unified model from multiple datasets, they generally do not explicitly distinguish domain-specific appearance variations from domain-invariant semantic information, placing a heavy burden on shared layers when the domain gap is large. To address this issue, we propose MDCL-UNet, a supervised multi-domain collaborative learning framework based on domain feature disentanglement. MDCL-UNet adopts a two-branched encoder in which a domain-specific branch equipped with the proposed Domain Style Instance Normalization (DSIN) module and a domain-invariant branch jointly disentangle domain features. A domain adversarial classifier and an MMD-based domain decoupling loss are used to ensure that the two branches learn complementary and orthogonal representations. Instead of discarding domain-specific features as in conventional domain generalization methods, a Domain Fusion Attention Module (DFAM) is further introduced to adaptively fuse domain-specific style cues with domain-invariant semantic features through attention-based integration, enabling effective reuse of cross-domain information. Extensive experiments on retinal vessel, optic disc/cup, and abdominal multi-organ segmentation datasets demonstrate that MDCL-UNet achieves consistently higher Dice scores and lower HD95 distances than single-dataset baselines and existing cross-dataset collaborative learning methods. Moreover, MDCL-UNet maintains stable training and favorable scalability with increasing numbers of domains. The source code will be available at https://github.com/lqr41710085/MDCL-UNET .

Links and sources

• PDF from original source

Comments