Spatiotemporal Glioblastoma Evolution Visual Prediction

Research at The ∀ Lab & Image Science Lab

project status: currently paused (January - September 2025) —

Conducted during the Spring and Summer semesters of 2025 under Dr. Pulkit Grover, Dr. Aswin Sankaranarayanan and Dr. Matthew J Shepard, MD, this study is complementary to a Partial Differential Equation based approach by Cynthia Han a fellow graduate student researcher. Both approaches use a multi-modal approach that combine FLAIR, T1, T2 and CT1 modalities of brain MRI from the LUMIERE dataset.

I worked on a Neural Ordinary Differential (N-ODE) network framework designed to model and predict tumor growth dynamics from longitudinal MRI data. By encoding a patient’s tumor size measurements over time into interpretable kinetic parameters, the N-ODE formulates tumor progression as a continuous-time dynamical system, allowing for unbiased and personalized predictions, even from early-stage data.

prelimnary results of my model (actually!) predicting a future timepoint of the tumour growth across all four modalities
  • the model uses a encoder-neural ordinary DE-decoder model with a combined loss approach, integrating Mean Squared Error to match anatomical details and Dice Loss to ensure accurate tumor segmentation, thus optimizing both image fidelity.
  • the model’s encoder-decoder architecture not only improves future tumor size predictions but also produces metrics that strongly predict overall survival.
me presenting this work at carnegie mellon's 2025 forum by the biomedical engineering department