DDS-UDA: Dual-Domain Synergy for Unsupervised Domain Adaptation in Joint Segmentation of Optic Disc and Optic Cup
New framework uses dual-domain synergy to adapt AI models across different medical scanners without new labels.
A research team led by Yusong Xiao has introduced DDS-UDA (Dual-Domain Synergy for Unsupervised Domain Adaptation), a new AI framework designed to solve a critical bottleneck in medical AI: domain shift. When convolutional neural networks (CNNs) trained to segment the optic disc and cup (key for glaucoma diagnosis) on one hospital's imaging system are deployed on another's, their performance often plummets due to differences in scanners and protocols. DDS-UDA tackles this without needing costly new manual annotations for each new domain.
The framework's innovation lies in two synergistic modules. First, a bi-directional cross-domain consistency module uses a coarse-to-fine dynamic mask generator to facilitate clean semantic information exchange between source and target domains, suppressing noise. Second, a frequency-driven intra-domain pseudo label learning module synthesizes spectral amplitude-mixed signals to enhance generalization within a domain. Built on a teacher-student architecture, DDS-UDA effectively disentangles domain-specific biases from the core, invariant features needed for accurate segmentation.
In comprehensive evaluations on multi-domain fundus image datasets, DDS-UDA demonstrated superior performance over existing unsupervised domain adaptation methods. This represents a significant step toward clinically robust AI that can maintain accuracy across the heterogeneous landscape of real-world medical imaging hardware, moving beyond proof-of-concepts on single, curated datasets.
- Solves 'domain shift' where AI models fail on data from different medical scanners than they were trained on.
- Uses a dual-module approach: cross-domain consistency for clean feature exchange and frequency-driven pseudo-labels for intra-domain generalization.
- Outperforms existing UDA methods in tests on multi-domain fundus image datasets for optic disc/cup segmentation.
Why It Matters
Enables more reliable deployment of diagnostic AI across diverse hospital systems, a major hurdle for clinical adoption.