Contamination-aware chest x-ray lesion detection
By constructing an initial memory library and using greedy kernel set sampling, axis projection modules, and dynamic attention modules, the performance degradation of the model caused by contaminated samples in chest X-ray images was resolved, improving the accuracy and robustness of lesion detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MINJIANG UNIVERSITY
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-17
AI Technical Summary
Existing unsupervised lesion detection methods suffer from performance degradation due to the presence of mislabeled abnormal images in the training set, particularly in chest X-ray images. This is caused by visual overlap between lesions and normal tissues and the influence of contaminated samples, resulting in insufficient detection accuracy.
An initial memory library is constructed and a greedy core set subsampling strategy is adopted to reduce redundancy. Combined with an axis projection module and a memory-based dynamic attention module, tainted features are pushed away from normal features through symmetric weighted contrastive loss, thereby improving the robustness of feature representation.
It improves the anomaly detection performance of chest X-ray images, increases the lesion recognition rate, reduces the risk of false negatives, and enhances the diagnostic accuracy of the model on real datasets.
Smart Images
Figure CN122415562A_ABST