Speaker
Description
Glioblastoma Multiforme is an aggressive primary brain tumor with poor prognosis. Understanding the patient-specific phenotype of the disease (e.g., proliferative vs diffusive) can be crucial for treatment planning. However, learning patient-specific parameters in a reaction-diffusion equation of glioma growth is difficult due to the (1) lack of time-series information available for a single patient, (2) magnetic resonance images (MRI) delineating existence of tumor but not encoding tumor cell density, and (3) lack of publicly available datasets. In this talk, I will present recent work on using curriculum learning to recover patient specific parameters. Curriculum learning is a machine learning technique in which we train on examples of increasing complexity. Using synthetic data, we first learn on dense time-series before annealing to recovering our parameters from two time points and fine-tuning at that data sparsity level. We demonstrate the efficacy of curriculum learning over baseline models especially in regions of increased noise and fewer training samples.
