Weakly supervised medical image anomaly detection method based on random decoupling and geometric adversarial contrast
By introducing a stochastic decoupling bottleneck and a geometric adversarial contrastive loss function into the encoder-decoder neural network, combined with a topological stability binarization module, the problems of abnormal contamination and unstable localization in weakly supervised medical image anomaly detection are solved, achieving efficient and adaptive anomaly region detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MINJIANG UNIVERSITY
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-17
AI Technical Summary
Existing weakly supervised medical image anomaly detection methods are prone to anomaly contamination, low feature discrimination, and unstable localization results when faced with scarce abnormal samples. They are difficult to effectively identify subtle lesions, and existing methods do not capture global contextual information well, and the localization results rely on manual thresholds and are unstable.
A method based on random decoupling and geometric adversarial comparison is adopted. By combining a global context fusion module, a random decoupling bottleneck module, and a topological stability binarization module with a geometric adversarial comparison loss function, an encoder-decoder neural network is constructed to achieve adaptive anomaly region localization.
It significantly reduced the false alarm rate, improved the model's ability to distinguish subtle lesions, enhanced the automation and consistency of detection, and was able to adapt to medical images from different devices and with different contrasts, effectively capturing global anatomical information.
Smart Images
Figure CN122415487A_ABST