We are also interested in gaining insights of AI techniques, as well as investigating AI technologies and developing and translating AI methodologies for medical applications. We are currently focusing on the following two areas.
Human-like parameter tuning for optimization problems: A wide spectrum of tasks in medicine can be formulated as solving optimization problems. Examples include iterative CT reconstruction and radiotherapy treatment planning. In the optimization models for these problems, there are typically parameters, which govern solution quality, e.g. image quality or plan quality.
Conventionally, these parameters are manually tuned for the best results. Yet the tedious process and very often sub-optimal parameter selection impede clinical utilization of these optimization models. We are exploring reinforcement learning techniques to develop a smart system that can automatically tune parameters with human-level intelligence.
Distributed deep learning to break data barrier: Behind the success of each deep-learning study lies a tremendous amount of data for model training and validation. Nonetheless, when it comes to deep learning for a medical problem, the amount of data is sometimes limited.
This issue is exacerbated by the difficulty of sharing data among different medical institutions due to issues like patient privacy concerns. Recently, distributed deep learning offers a new angle to solve this problem. With the deep learning algorithm running at each institution using private patient data, while communicating with each other to share non-private algorithm data, it becomes possible to collectively perform learning activities effectively using all the data.
In this project, we are developing the infrastructure and distributed deep-learning algorithms to enable sharing knowledge among different institutions without sharing data. This will facilitate a number of deep learning methods for medical applications.
- Shen, C., Gonzalez, Y., Chen, L., Jiang, S., & Jia, X. (2017). Intelligent Parameter Tuning in Optimization-based Iterative CT Reconstruction via Deep Reinforcement Learning. arXiv preprint arXiv:1711.00414. (arXiv)