Prediction gallery¶
This gallery shows predicted diffusion volumes produced by NiFreeze’s models on several real datasets spanning the acquisition-scheme spectrum — a simple, legacy DTI dataset, a single-shell HARDI dataset, a multi-shell dataset, and a DSI dataset. For each dataset, every applicable model is run in two modes:
LOVO (leave-one-volume-out): the model is fit on every other volume and used to predict the held-out orientation, so the prediction is unbiased with respect to the target volume.
single-fit: the model is fit once on all volumes and then predicts.
The whole gallery is driven by a single declarative matrix
(a support harness under docs/sphinxext/gallery) of
(dataset × model × mode) cells. A model’s
capability contract (BaseModel) decides, before
fitting, which cells apply — so each page also reports a coverage table
recording what was exercised and why any cell was skipped (e.g. DKI needs
multiple shells; the shell-averaging model has no single-fit mode). This makes
the gallery a living record of what the models are validated to do on real data.
Note
The datasets are fetched from OpenNeuro. Two labels are worth stating plainly:
ds000114 is single-shell at b=1000 s/mm² (high angular resolution, milder
than textbook high-b HARDI), and ds004737 is compressed-sensing DSI (a
sub-sampled q-space grid, not a full 258-point acquisition).
Note
The pages below are generated by the gallery workflow, which fits each
(dataset × model × mode) cell in its own job and renders the panels. They
are plain figures over that output — nothing is computed at documentation
build time.