Spatiotemporal Glioblastoma Evolution Visual Prediction

summary

This project explored longitudinal glioblastoma forecasting from multi-modal brain MRI during Spring and Summer 2025 under Dr. Pulkit Grover, Dr. Aswin Sankaranarayanan, and Dr. Matthew J Shepard, MD. My track focused on Neural ODE-based forecasting using FLAIR, T1, T2, and contrast-enhanced T1 from the LUMIERE dataset.

problem

Tumor growth is continuous in time, but clinical MRI arrives as sparse, irregularly sampled snapshots. Forecasting future progression from early scans requires a model that can use limited patient history while still respecting temporal dynamics.

approach

I built a Neural Ordinary Differential Equation framework that models tumor evolution as a continuous-time latent dynamical system. The project went through several branches, but the common structure was:

• encode historical MRI observations into a latent state
• evolve that state forward with a Neural ODE block
• decode a future MRI estimate or growth trajectory
• compare against persistence baselines and simpler temporal setups

The model uses an encoder-Neural ODE-decoder setup with combined Mean Squared Error and Dice losses to balance anatomical fidelity and tumor-focused segmentation quality. A separate benefit of the framework is that the learned latent dynamics expose interpretable progression parameters that can be useful for survival-related analysis.

current result

The main outcome was a working longitudinal MRI forecasting framework and a set of related branches that tested different ways of conditioning on history, patient-specific data, and physics-inspired modeling. The work was strong enough to present at Carnegie Mellon’s 2025 Biomedical Engineering forum, but the project is currently paused rather than completed.

branch evolution

This project evolved through a few related modeling directions.

The earliest recovered version was a rough Neural ODE prototype built from the original source code. That branch established the basic idea of using an attention U-Net encoder, temporal conditioning, and a latent ODE block for glioblastoma forecasting.

I then cleaned that into a runnable neural-ode-implementation branch. That version turned the recovered idea into a proper pipeline for the local patient data, using a 2D slice-based attention U-Net + Neural ODE model, strict holdout evaluation, and a persistence baseline for comparison.

After that, I shifted to a history-conditioned-forecast branch that reframed the task as prefix-history prediction: use all earlier MRI weeks for a patient and forecast the next one. In that version, each historical week is encoded separately, a learned week embedding is added, the latent history is aggregated, and the Neural ODE evolves the state forward in continuous time before decoding the future scan.

The history-conditioned-forecast-slim branch is a slimmer merged version of that same prefix-history line, keeping the forecasting idea while trimming the branch down.

In parallel, there was also a separate physics-dual-patient-rerun branch that explored a physics-informed 3D forecasting pipeline. That approach treated the MRI evolution as a more explicit dynamics problem rather than a learned latent forecast, and it used the same local patient set for comparison.

Across all of these branches, the common goal was the same: model glioblastoma progression from longitudinal MRI data and compare learned forecasts against simple persistence baselines.

limitations

• The dataset size is still small for a high-capacity temporal generative model
• Cross-patient generalization remains harder than within-patient forecasting
• The project was paused before the strongest branch could be fully consolidated and benchmarked

next steps

• consolidate the best branch into one clean training and evaluation pipeline
• add stronger held-out evaluation and survival-linked analysis
• compare Neural ODE forecasting directly against the complementary PDE-based branch