Intra-tumor molecular heterogeneity is a leading cause of treatment failure for aggressive cancers such as glioblastoma (GBM), the most aggressive type of brain cancer. The invasive nature of biopsy makes it impossible to sample every sub-region to understand the regionally-specific molecular & genetic characteristics. Neuroimaging such as magnetic resonance imaging (MRI) portrays the entire brain non-invasively, providing the opportunity to estimate the spatial molecular & genetic distributions across each individual tumor, such as the spatial distributions of tumor cell density, aberrations of driver genes, and immune cell abundance. Such capability will revolutionize treatment: surgical resection can be better guided; radiation therapy can be spatially optimized to avoid over- and under-treating certain regions of the brain; therapeutics can be adapted to regional genetic aberration patterns; treatment response can be precisely tracked to differentiate treatment failure from pseudo-progression.
Methodologies:
- Fusion of machine learning and mechanistic models
- Knowledge-infused global-local data fusion
- Semi-supervised learning & weakly-supervised learning
- Uncertainty quantification
- Posterior regularization