Glioblastoma MRI Forecasting

Reproducible Neural ODE forecasting on LUMIERE-style longitudinal MRI

The repository now tracks aggregation code, validation workflow, publication protocol, and cohort-level result tables generated from local run_summary.json files. This page reflects the current reproducibility checkpoint rather than older script-only smoke outputs.

View Repository Aggregated Metrics
81 tracked patient-level runs
+6.7% mean relative improvement
65 positive-improvement patients
7 unit tests in CI

Current Checkpoint

The checked-in aggregate table summarizes the local lumiere_full_v1_* runs using the best available metric split per patient, preferring holdout metrics when present and otherwise falling back to older all-sample summaries.

0.00828323 mean model MSE
0.00876673 mean persistence MSE
16 nonpositive-improvement patients

Tracked artifacts

The repository now includes tracked aggregate outputs in results/lumiere_full_v1_metrics.csv, .json, and .md so the top-line cohort numbers are generated from versioned files rather than copied into prose by hand.

Runtime provenance

Each run summary now records Python, Torch, NumPy, requested device, resolved device, and per-split metric CSV paths. The pipeline also exports train, holdout, and baseline CSV tables for every run directory.

Top Runs

These rows come directly from results/lumiere_full_v1_metrics.md. At the moment they all use the all split because the older full-run summaries predate the holdout-aware schema.

Patient Split Samples Model MSE Baseline MSE Improvement
Patient-028 all 7 0.00386073 0.0123518 +68.7%
Patient-004 all 6 0.00336998 0.0104322 +67.7%
Patient-066 all 8 0.00379933 0.0114351 +66.8%
Patient-077 all 8 0.00324765 0.00911861 +64.4%
Patient-031 all 17 0.00351799 0.00956643 +63.2%

Final publication claims should come from a frozen rerun where every row uses metric_split=holdout.

Manuscript Figures

Representative figures are checked into the repository as static manuscript assets. They remain useful as qualitative examples, but the quantitative claims on this page are driven by the tracked result tables.

Patient-073

Forecast figure for Patient-073

Patient-023

Forecast figure for Patient-023

Patient-015

Forecast figure for Patient-015

Patient-006

Forecast figure for Patient-006

Publication Track

The repository now has a concrete protocol for a frozen rerun. The remaining gap is not infrastructure; it is executing the final cohort run with consistent holdout reporting and manuscript figures regenerated from that run prefix.

What is locked down

Registration manifests, per-run metric CSVs, result aggregation, smoke data generation, explicit device selection, and GitHub Actions validation are all in place.

What still defines publication

A final lumiere_publication_v1 rerun should use the tracked publication protocol, require holdout rows for every included patient, and regenerate the aggregate table with --split holdout.