braindatalab / bsi-zoo

A brain source imaging zoo!

Welcome to the Brain Source Imaging (BSI) Zoo

This package allows for a standardized comparison of BSI methods with custom simulations.

✔️ source estimation at controlled SNR levels

✔️ support for both random and semi-real data generation

✔️ support for fixed and free orientation

✔️ performance evaluation

✔️ parameter optimisation

.. and ships with several BSI methods!

Setup

python setup.py install

Documentation

Coming Soon 🚀

bsi-zoo's Issues

Add test for metrics

• Test cases for metrics
• check the free orientation case is combined dipole

Add fake solver to benchmark as a baseline

Disregard rows with NaN metric scores in benchmark data

Some estimators predict fewer vertices than the ground truth for certain parameter combinations. This cannot be evaluated by all current metrics, and is stored as NaN is the benchmark pkl files.

add MCE estimator

EMD score from stc

Added EMD score for metrics using the true and estimated stc using this ref.

1. The resulting scores are all over the range. How to normalize something like this?
2. Also, how to think about this emd score intuitively, given that we don't give any geometry info.

@AliHashemi-ai anything to add?

@agramfort

Fix github actions to run again

Hi @agramfort
you can see an example here on the #63 branch
https://github.com/braindatalab/BSI-Zoo/actions/runs/5660531247/job/15336427160

It has something to do with the PEP versions.

Util to get paths

Add extreme estimator alpha values in CV grid

• all methods
• diag noise only
• extreme noise case
• 3 runs, 1 subjects

Benchmark pipeline doesn't support free orientation data

F1 and EMD score zero for gamma map

• Check this!

New plot idea (SNR vs Residual noise)

So that we have a proper diagonal in the figs:

• plot SNR vs residual noise

• 2D gaussian for extrapolating this and filling empty spots

Hypothesis:
on the diagonal, we should get the best performance
Underfit and overfit behavior on either sides

Euclidean distance test for nnz>1

TODO: Extract coordinates of n=nnz vertices with the highest amplitudes from stc.

Spatial CV for type II methods

ValueError('operands could not be broadcast together with shapes (204,204) (136,136) '),

cleanup

Orientation support in NLL metric

currently only supports fixed orient

Add eLORETA

we need another linear estimator in the pipeline