Skip to contents

fmridataset

Changelog · Canonical data model · API audiences · Issues · Contributing

fmridataset is an R package for keeping fMRI arrays aligned with what their rows and columns mean. It represents each assay as observations by features and carries stable IDs, annotations, entities, relations, and explicit spatial identity through views, transformations, and storage round trips.

Use it when raw time series, beta estimates, parcel values, surface data, or latent representations must retain their meaning as they move between analysis steps and storage systems.

Status: The 0.10.0 development line is the road to 1.0 and requires R 4.3 or newer. fmri_frame() is the only data container. The pre-frame 0.x dataset and backend API has been removed; the last commit carrying it is 3ae565e. APIs may still change before 1.0.

Installation 

install.packages("remotes")
remotes::install_github("bbuchsbaum/fmridataset")

fmridataset is not on CRAN. The published R-universe build is version 0.8.9 and documents the old dataset API, not this one.

Quick start

Create a small volume-backed frame, then select observations and voxels by their stable IDs:

library(fmridataset)

signal <- matrix(seq_len(24), nrow = 6, ncol = 4)
voxel_space <- volume_space(
  dim = c(2, 2, 1),
  affine = diag(4),
  template = "toy"
)

frame <- fmri_frame(
  assays = list(signal = signal),
  observations = data.frame(
    .obs_id = paste0("volume-", seq_len(6)),
    run_id = rep(c("run-1", "run-2"), each = 3),
    TR = 2
  ),
  space = voxel_space
)

view <- frame[c("volume-6", "volume-1"), c("voxel-4", "voxel-2")]

collect_assay(view)
#>      [,1] [,2]
#> [1,]   24   12
#> [2,]   19    7

observation_ids(view)
#> [1] "volume-6" "volume-1"

feature_ids(view)
#> [1] "voxel-4" "voxel-2"

The numerical view, observation metadata, feature metadata, and restricted volume space all retain the requested order. Nothing is read until collect_assay() asks for values, and explain(frame) describes the frame, its sources, and the cost of realizing it without reading any.

Axis IDs are required. Importers derive reproducible IDs from declared keys with axis_frame(id_policy = "deterministic", ...). Exploratory session-only IDs require id_policy = "ephemeral"; they are visibly marked and must be replaced before FDS persistence or semantic certification.

Runs and timing

Run structure, TR, and censoring are observation metadata under a validated schema rather than separate state:

has_temporal_schema(frame)
#> [1] TRUE

schema <- temporal_schema(frame)
schema$run_lengths
#> run-1 run-2
#>     3     3

as_sampling_frame(frame)   # an fmrihrf::sampling_frame, when runs are contiguous

The run column is discovered from a run-typed relation, scan_id, or run_id, or named explicitly with run_col. Frames without acquisition structure, such as beta estimates, simply report has_temporal_schema() as FALSE.

Load one BIDS subject

With bidser 0.5.0 or newer, a subject’s fMRIPrep BOLD runs open as one lazy frame:

bold <- read_bids_bold(
  "/data/my-study",
  subject = "01",
  task = "memory",
  space = "MNI152NLin2009cAsym"
)

run_1 <- filter_obs(bold, run_id == "run-1")
map <- spatial_map(bold, observation = 1)

Construction reads BOLD headers and the matching run masks, but not BOLD values. By default the frame uses the intersection of the run masks. No resampling or cross-space alignment is performed implicitly; ambiguous spaces, masks, or multi-echo selections produce an error requiring an explicit choice.

Persist and reopen 

path <- write_frame(frame, "frame.h5")   # atomic, manifest-backed HDF5 via fmristore
again <- open_frame(path)                # assays reopen as lazy sources

What it covers

  • Aligned assays: keep one or more numerical assays tied to the same observation and feature axes, with explicit roles, units, and provenance.
  • Annotated domains: attach scalar metadata, multivariate blocks, experimental entities, and validated relations without copying assay data.
  • Spatial identity: represent volume, surface, parcel, basis, and composite feature spaces; compatibility is checked by identity rather than dimensions.
  • Explicit transformations: map between feature spaces with validated, serializable operators and derivation provenance.
  • Bounded execution: read lazy in-memory, NIfTI, sharded, HDF5-backed, and experimental Zarr sources through observation-by-feature selections, with realization budgets enforced before any read.
  • Collections and studies: group native-space frames that cannot share a feature axis, and link heterogeneous frames through shared entities, relations, and typed maps.
  • Portable semantics: serialize the logical frame contract with FDS v1 and bind it to physical storage without changing axis or spatial identity.

Fit and boundaries

fmridataset owns semantic containers, alignment, views, sources, spaces, and the logical FDS schema. Companion packages own adjacent responsibilities:

  • neuroim2 provides native neuroimaging objects.
  • delarr provides lazy numerical plans and bounded execution.
  • fmristore provides certified HDF5 layouts, atomic writes, append, and recovery.
  • multidesign and fmrigds own design compilation and statistical execution.
  • bidser provides BIDS discovery used by read_bids_bold().

HDF5 is the certified persistence direction for 1.0. Zarr support remains experimental.

Documentation

Contributing

See CONTRIBUTING.md for development setup and checks. Changes to public behavior should include behavioral tests, updated roxygen documentation, and a NEWS.md entry.

License

GPL (>= 3)