Glioblastoma MRI Forecasting

Unified branch findings for the recovered forecasting work

A compact project page covering the physics-informed baseline, the Neural ODE implementation, the prefix-history refinement, and the simple persistence comparison that still wins on the tiny local cohort.

View Repository Read README Open Pipeline
2 local patients used in runs
91 patients in the public LUMIERE cohort
0 learned branches beat persistence
30.33 GB size of the public LUMIERE archive

Project Summary

The repository keeps the runnable Neural ODE pipeline and its recorded outputs, while the branch history documents two other comparison points: the physics-informed 3D attempt and the persistence baseline. The practical conclusion is consistent across all of them: the learned models are functional, but the tiny cohort is too small for them to surpass a simple last-scan predictor.

Main Code

The active pipeline is scripts/run_neural_ode_pipeline.py. It supports full prefix history, sliding windows, strict future holdout, and separate-per-patient runs.

Data Context

The current local working set only contains patient_007 and patient_067. The larger LUMIERE cohort is the right next step for any claim about generalization.

Branch Timeline

The work evolved from a physics-informed experiment into a cleaned Neural ODE pipeline, then into a prefix-history formulation that made the forecasting objective explicit.

Initial Physics Branch

A runnable 3D physics-informed attempt that acted as a historical comparison point, but did not beat the persistence baseline.

Cleaned Neural ODE Branch

The recovered Neural ODE idea was turned into a reproducible pipeline with strict future-week holdout and baseline evaluation.

Prefix-History Refinement

The sequence-to-one prefix formulation clarified the task and produced stable forecasts, but still lagged behind persistence.

Slim Main Branch

The current branch keeps the runnable pipeline, result artifacts, and a compact GitHub Pages summary.

Final Comparison

The key takeaway is that the model runs end to end, but the last-scan baseline remains the strongest method on this small local cohort.

Method Patient Held-out week MSE Baseline MSE MAE Baseline MAE
Neural ODE strict holdout patient_007 105 0.04119 0.00399 0.16927 0.03153
Neural ODE strict holdout patient_067 152 0.02426 0.00752 0.14075 0.04469
Persistence baseline patient_007 105 0.00399 0.00399 0.03153 0.03153
Persistence baseline patient_067 152 0.00752 0.00752 0.04469 0.04469

The learned model improves the forecasting setup, but not the final metric outcome on this cohort.

Results Gallery

The existing smoke-run outputs are kept in the repository and shown here directly so the page reflects actual experiments rather than only summary text.

Patient 007

Prediction and target for patient 007
Prediction versus target for week 105.
Training loss for patient 007

Patient 067

Prediction and target for patient 067
Prediction versus target for week 152.
Training loss for patient 067

Interpretation

The physics-informed branch was a useful comparison point, but the practical non-Neural-ODE baseline is the persistence predictor. That simple approach still beats the learned methods here, which is why the repo is best read as a feasibility record rather than a final predictive claim.

What the learned models show

They demonstrate that the history-conditioned forecasting pipeline is implementable and that the models can produce plausible outputs for all four modalities.

What the baseline shows

The latest observed scan is already a very strong predictor on this small cohort, so any new model has to clear a high bar just to look useful.