Skip to content

fmri_data_pca.md

Latest commit

 

History

History
94 lines (68 loc) · 3.6 KB

fmri_data_pca.md

File metadata and controls

94 lines (68 loc) · 3.6 KB
Raw

fmri_data.pca — principal components decomposition of an image set

← back to fmri_data methods · Object methods index · Recasting objects

Principal components analysis on the [images × voxels] matrix (obj.dat'). Returns image-level scores, an image-vector / fmri_data object whose images are the eigenmaps (canonical brain patterns), and the variance explained by each component. Useful for unsupervised exploration, denoising, and as a precursor to PCR / ICA pipelines.

Quick example

imgs = load_image_set('emotionreg');
[scores, eigenmap_obj, explained] = pca(imgs, 'k', 5);

pca sample

Code map

pca code map

Editable PowerPoint version

Usage

[component_scores, eigenmap_obj, explained] = pca(obj, ['noplot'], ['k', num_components])

Voxels are mean-centered but not scaled — PCA operates on the covariance matrix, so high-variance voxels have proportionally more influence on the eigenmaps. The mean-centered data can be reconstructed as X = component_scores * eigenmap_obj.dat'.

Inputs

Argument Type Description
obj image_vector / fmri_data Multi-image object (e.g. a time series or a set of contrast images).
'k', num_components int Optional. Number of components to compute. Default 3; lower is faster. Synonym: 'numcomponents'.
'noplot' flag Suppress the default score-trace and eigenmap montages.

Outputs

Output Type Description
component_scores [images × k] matrix Score for each image on each component — feed to downstream tests, plot against design regressors, etc.
eigenmap_obj image_vector / fmri_data Same shape as obj, with one image per component holding voxel weights (eigenmap).
explained column Percent variance explained by each component.

Notes

  • The covariance-based decomposition means you may want to z-score voxels beforehand (e.g. with rescale(obj, 'zscoreimages')) if all voxels should contribute equally — i.e. switching to a correlation-based PCA.
  • When the input is a time-series dataset, plot the FFT of each component to spot scanner artefacts vs. neural rhythms (see example).
  • pca will silently choose the smaller of k and the number of images via MATLAB's built-in pca function with 'Economy', true.

Example: PCA of an emotion-regulation contrast set

% Load 30 single-subject contrast images
obj = load_image_set('emotionreg');

% Default: 3 components, with score traces and montages of each eigenmap
[component_scores, eigenmap_obj, explained] = pca(obj);

% Re-display the eigenmaps as transparent overlays
montage(eigenmap_obj, 'trans')

% Variance-explained scree plot
figure; bar(explained); xlabel('Component'); ylabel('% variance');

Other examples

% Frequency content of component scores from a TR = 1.8 s time series
[scores, eigenmap_obj] = pca(ts_obj, 'k', 10, 'noplot');
create_figure('fft'); [fftmag, fftfreq, h] = fft_calc(scores, 1.8);

See also