Unconstrained to flat map with PCA method - Details on the process

ginod · July 19, 2022, 9:39am

Hello Brainstormers,

In the Workflow tutorial in Brainstorm (https://neuroimage.usc.edu/brainstorm/Tutorials/Workflows#Unconstrained_cortical_sources), it was suggested to flatten unconstrained cortical sources (Sources > Unconstrained to flat map). This process had two method: for each vertex, take the norm or perform a PCA to extract the orientation which explains the largest part of the variance.

However, I'm not sure I completely understood how is the computation under the PCA method.
Can anyone please had some further explanation on this process?

Have a good day

Francois · July 19, 2022, 1:55pm

The PCA computation is currently performed by this function:

github.com

brainstorm-tools/brainstorm3/blob/master/toolbox/math/bst_scout_value.m#L267

      
        
                msg = ['BST> First component explains ' num2str(explained * 100) '% of the signal'];

                if scoutName

                    msg = [msg ' of cluster ' scoutName];

                end

                disp([msg '.']);

            end

            end

            

            

            %% ===== PCA: FIRST MODE =====

            function [F, explained] = PcaFirstMode(F)

                % Remove average over time for each row

                Fmean = mean(F,2);

                F = bst_bsxfun(@minus, F, Fmean);

                % Signal decomposition

                [U,S,V] = svd(F, 'econ');

                S = diag(S);

                explained = S(1).^2 / sum(S.^2);

                % Find where the first component projects the most over original dimensions

                [tmp__, nmax] = max(abs(U(:,1))); 

                % What's the sign of absolute max amplitude along this dimension?

This option is currently under evaluation, and will probably be subject to changes in the next few months. See the GitHub pull requests:

For statistical analysis, I suggest you chose the NORM option, since the tests used are power tests and do not need to keep the sign of the signals or their frequency contents.
Note that this part is still not very clear, as these tests lead to very poor results as opposed to the t-tests performed on source results with constrained orientations.
See this thread: Group analysis: recommended workflow and statistics · Issue #141 · brainstorm-tools/brainstorm3 · GitHub

@Sylvain @Marc.Lalancette @Raymundo.Cassani

Marc.Lalancette · July 19, 2022, 3:35pm

Indeed the current implementation of PCA for flattening is problematic when run on multiple trials: a sign issue leads to cancellations. This should be fixed hopefully soon.

@Francois, if you have a chance to look at the description of PR559, please let me know if the approach makes sense or if you have a better idea. I was hoping for feedback before continuing working on it.

Francois · July 20, 2022, 12:55pm

@Marc.Lalancette I had a brief look at your PR about the global PCA and this is pretty complex, I won't be able to dive into this before the fall.
I think this thread is more about statistics on unconstrained sources than about PCA. I don't think we would recommend using the PCA for statistics on source maps anyway.

ginod · July 21, 2022, 4:41pm

Thank you all for your responses, I'm impatient to follow the continuity of your work