Oncology, MRI
A tumor doesn't move on a screen by magic — it grows according to an equation. The datasets in this category are generated by actually solving that equation: the Fisher-Kolmogorov reaction-diffusion model, the standard mathematical description of how a glioma spreads through tissue, integrated directly on the same spatial grid an MRI scanner would produce.
That means every voxel of tumor you see in the data is there because the underlying growth model put it there, not because a generative network learned what tumors tend to look like. The ground truth — the exact tumor boundary, the exact concentration field — is known at every timepoint, which is the one thing a real oncology imaging dataset can never give you: nobody has a biopsy-verified map of where every cancer cell actually is.
What this makes possible
- Segmentation training with exact ground truth — no inter-rater disagreement, no ambiguous tumor margin, because the margin is a mathematical boundary, not a radiologist's best guess.
- Radiomics feature validation — test whether an imaging feature actually tracks the underlying biology, since the biology is known exactly.
- Longitudinal growth series — multiple snapshots of the same simulated tumor over time, for models that need to learn progression rather than a single scan.
The passport for each dataset states exactly which physical effects the model includes and which it deliberately leaves out — growth kinetics are one thing, full radiological image formation is another, and we don't blur that line.