Cross-domain low-dose CT image denoising method based on semi-supervised dual-space representation alignment
By using a semi-supervised dual-space representation alignment method, combined with source domain paired annotations and target domain unlabeled data, the SDRA framework is constructed. This solves the performance degradation problem of existing low-dose CT image denoising methods in real-world environments and achieves efficient denoising in multi-center, multi-protocol scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QUFU NORMAL UNIV
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-17
AI Technical Summary
Existing low-dose CT image denoising methods rely on extensive paired label training, which makes it difficult to adapt to the multi-source degradation and diverse heterogeneity of real clinical environments. Furthermore, unsupervised methods fail to fully utilize source domain supervision information, resulting in decreased denoising performance and error accumulation.
A semi-supervised dual-space representation alignment method is adopted, which combines source domain paired labeled data and target domain unlabeled data. Through latent space MMD and GCL representation alignment and image space EDA mechanism, the SDRA framework is constructed for end-to-end training, which improves the model's robustness across domains and denoising quality.
In real-world clinical scenarios involving multiple centers and multiple protocols, it improves denoising robustness and generalization performance while maintaining structural consistency and visual quality, making it suitable for such scenarios.
Smart Images
Figure CN122415374A_ABST