Artificial intelligence is increasingly becoming part of everyday life, driven in large part by advances in deep learning. The medical domain is no exception. But developing an effective deep neural network for a specific task still requires considerable expertise, time and computing power.
AI that helps design better AI
Together with research partner Leiden University Medical Center (LUMC) and industry partners Elekta and Ortec Logiqcare, CWI researcher Peter Bosman, who leads the project, aims to develop new technology that can automate the search for suitable deep neural network architectures. The goal is to create methods that scale better than existing approaches.
The project will combine promising ideas from neural architecture search with key strengths of another branch of AI: evolutionary algorithms. More specifically, the researchers will build on GOMEA, a relatively recent type of evolutionary algorithm developed as part of Bosman’s long-running research line on scalable model-based evolutionary algorithms.
Making medical AI work across hospitals
Although the technology is expected to be broadly applicable, an important motivation for the project comes from medical applications. In medicine, high-quality AI often depends on large amounts of training data. Yet individual hospitals usually have limited datasets, and sharing patient data between hospitals is often difficult because of privacy, legal and organisational constraints.
For that reason, the researchers will develop variants of the new technology for situations in which data is spread across multiple locations. They will investigate solutions for cases where data can be pooled in one place, as well as for scenarios in which only information about learned models can be shared.
Applications in radiation oncology
The project will demonstrate the potential of the new technology through two applications in radiation oncology: automatic contouring of medical images and the prediction of dose distributions for brachytherapy cancer treatment planning.
Both applications are closely connected to Bosman’s broader research on using AI to create accurate, interpretable and automatic brachytherapy treatment plans. Software developed within that research line was recently used for the first time at Amsterdam UMC to support the treatment of a patient with prostate cancer.
Research team and international collaboration
The CWI team will include one postdoc, one PhD student and one scientific programmer. A second PhD student will be based at LUMC.
The project also includes planned international collaborations with researchers at the University of Texas at Austin and Virginia Commonwealth University in the United States, the University of Queensland in Australia, and Victoria University of Wellington in New Zealand.
Photo: Ivo van der Bent
