Towards Trusted AI for Healthcare: from Theory to Practice

With the development of artificial intelligence (AI) for healthcare, people started to pay more attention to the challenges and risks associated with AI models to be used in sensitive healthcare applications. Currently, there are still many practical barriers to the promotion of AI in healthcare applications, such as insufficient training samples, difficulties in data sharing and labeling, lack of transparency in AI models, etc. To overcome these barriers and accelerate the application of AI in healthcare, we must work on developing more accurate models while also improving AI's trustworthiness. Trusted AI has emerged as an important trend in AI research and industry. Trusted AI can be characterized by privacy, interpretability, robustness, and fairness. This presentation will discuss how to build a new generation of AI-powered healthcare systems and focus on the ongoing progress in both theories and practice towards this goal. First, I will address the issue of data security and data sharing by presenting our theoretical and algorithmic advances in federated learning to build secure AI systems. Second, to address the AI model interpretability issues, I will discuss how to improve the effectiveness and reliability of AI interpretable tools and show their applications in image segmentation and disease diagnosis. Finally, I will briefly introduce the preliminary exploration in the direction of robustness and fairness and discuss the opportunities and challenges faced by trusted AI in healthcare applications.