[Mne_analysis] Signs of source solutions with fixed orientation

[Alexandre Gramfort]

hi Ying,

> I was working on a source localization method, and I observed this sign
> issue when I did a simple simulation.
> I simulated a waveform for an ROI in the left auditory cortex
> ("left_aud_dipoles.png"), with a fixed orientation (normal direction),  but
> using MNE,  some of the waveforms of the dipoles in that ROI had flipped
> signs. (see "truth_left_aud.png" and "mne_left_aud.png", near time 0 ms).
> I learned that this is common because of the sulci or gyri structure.
> But is this also common in simulations? Is it inevitable?

it's hard to avoid but if you use a sparse solver that will try to "pick"
the right side of the sulcus it is possible.

see for example the red dipole in the estimates of:

http://martinos.org/mne/stable/auto_examples/inverse/plot_mixed_norm_L21_inverse.html#example-inverse-plot-mixed-norm-l21-inverse-py

on the right hemisphere the auditory response is explained with one
red dipole on the planum temporale.

> If this is true, when I measure the  error of the source solution,  should I
> flip the signs or take the absolute values before comparing with the truth?

There are many different error metrics for inverse methods (L2, peaks
distances, time series correlations, earth-movers distance, F1-score,
ROC etc.). You should always use sign flip trick to average signed
estimates.

HTH
Alex