Hierarchical Self-Supervised Adversarial Training for Robust Vision Models in Histopathology
Researchers have developed a new method called Hierarchical Self-Supervised Adversarial Training (HSAT) to improve the robustness of vision models used in histopathology. This approach specifically leverages the hierarchical structure of medical images, such as patient-slide-patch relationships, which were previously overlooked by other adversarial training techniques. HSAT integrates multi-level contrastive learning to create more effective adversarial examples, leading to significant performance gains in both white-box and black-box attack scenarios on the OpenSRH dataset. AI
IMPACT Enhances the reliability of AI models in critical healthcare applications, potentially leading to more trustworthy diagnostic tools.