A retinopathy recognition method and system based on a diffusion model
By constructing a diffusion model to generate high-quality normal samples and combining it with an anomaly detection model, the problem of insufficient accuracy and generalization ability in the identification of small-sample retinal lesions was solved, and efficient identification and stable detection of retinal lesions were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUJIAN UNIV OF TECH
- Filing Date
- 2026-05-22
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies suffer from low classification accuracy and insufficient generalization ability in the identification of retinal lesions with small or zero samples, especially with poor model performance stability under cross-device, cross-population, or cross-collection conditions.
A diffusion model-based method for identifying retinal lesions is constructed. High-quality normal samples are generated through the diffusion model, and retinal lesion identification is performed by combining the anomaly detection model. The method includes image quality assessment, diffusion model training, and anomaly detection model construction. The diffusion model is used to generate potential representations that conform to normal anatomical structures, and anomaly detection is performed by reconstruction error or feature comparison.
It improves the detection accuracy and generalization ability of retinal lesion identification, reduces the dependence on labeled data, is suitable for medical imaging scenarios with high labeling costs and scarce abnormal samples, and improves the robustness and stability of the model.
Smart Images
Figure CN122415581A_ABST