How-to guides#

Task-focused recipes for things you already understand and just want to do. For the concepts behind them, see Learn; for the full signatures, see the API.

Accelerate a batch on the GPU#

The CPU backend is the default and computes everything on its own. For large cohorts, install the gpu extra and select the GPU backend; nothing about the analysis code changes.

import spectralbrain as sb
# sb.NumpyBackend is the default; the gpu extra adds CUDA-backed solvers.

Harmonize across scanners before group statistics#

When a reference or patient cohort spans multiple sites, harmonize descriptors with ComBat / ComBat-GAM so that group differences reflect biology, not scanner. This is the dual sensitivity question — always report harmonized and raw where batch and group are confounded.

from spectralbrain.statistics import analysis   # harmonization + group tests

Vertex-wise statistics with real family-wise error control#

Use max-statistic permutation for genuine FWE control, with FDR and TFCE as alternatives, plus partial correlations with the correct degrees of freedom.

from spectralbrain.statistics import analysis

Fit a Bayesian model#

Six PyMC models ship with the bayesian extra, with posterior and diagnostic plots in viz.

from spectralbrain.statistics import bayesian
from spectralbrain.viz import bayes as bayes_viz

The MTLE-HS hippocampus workflow#

The library’s primary use case: load HippUnfold hippocampal surfaces, compute spectral descriptors, build a normative reference, and z-score patients, lateralizing left vs. right. Tutorials 06–10 assemble this end to end.