Background: Oncology is a major focus of precision medicine, with medical imaging, e.g., Magnetic Resonance Imaging (MRI) and Computed Tomography (CT), gaining an increasingly important role. Artificial Intelligence (AI), particularly deep learning, has made significant advances in developing predictive models from medical imaging. Currently, in medical imaging, AI models are manually constructed through a heuristic trial-and-error process, which suffers various drawbacks (e.g. time consuming, requiring expert knowledge, low reproducibility, suboptimal performance).
Aim: The aim of this research is to overcome these drawback by developing AutoML methods to automatically optimize model construction per oncological application. You will work on a variety of oncological diseases, with a focus on rare cancers (e.g. sarcoma, primary liver cancer, colorectal cancer, melanoma, glioma), where in each application there is a classification task to be solved. You will have to design a search space of promising deep learning solutions (e.g. CNNs, transformers) and develop an AutoML algorithm to optimize model construction. We aim to make the optimization strategy model- and task-agnostic, such that various deep learning models can be tried, and other tasks besides segmentation such as classification can be used. The project is quite technical: however, the resulting biomarkers can have substantial clinical impact.
Student task description
- Familiarize yourself with the literature on active learning.
- Write a report detailing the advantages and limitations of this approach.
- Weekly meetings with the supervisor
12 months with 36 hours per week.