Add cross-subject spatial correlations to scoring

Scoring/calculateDecompositionAccuracy.m

@@ -57,6 +57,35 @@ end
 
 %--------------------------------------------------------------------------
 
+% Spatial accuracy - cross subject correlations
+scores.P_xs.data   = NaN(params.V, params.N);
+scores.P_xs.name   = 'Spatial maps (cross subject)';
+scores.P_xs.metric = 'Correlation';
+scores.P_xs.range  = [-1.0, 1.0];
+for n = 1:params.N
+    Pn     = NaN(params.V, params.S);
+    inf_Pn = NaN(params.V, params.S);
+    for s = 1:params.S
+        Pn(:,s) = P{s}(:, gt_inds(n));
+        inf_Pn(:,s) = signs(n) .* inf_P{s}(:, inf_inds(n));
+    end
+    % Just the correlations between the two
+    %cP = diag(corrcoef(Pn', inf_Pn'));  % [V, V]: Too big!
+    dim = 2;
+    zPn = zscore(Pn, 1, dim);  % normalise by N, not N-1
+    inf_zPn = zscore(inf_Pn, 1, dim);
+    cP = mean(zPn .* inf_zPn, dim);
+    % Ignore elements with limited true variability
+    cP(std(Pn, 1, dim) == 0.0) = NaN;
+    cP(sum(Pn == 0.0, dim) > 0.2 * params.S) = NaN;
+    
+    % Return scores in order of GT components so comparable across methods
+    scores.P_xs.data(:, gt_inds(n)) = cP;
+end
+
+%--------------------------------------------------------------------------
+
+
 % Repeat for the temporal accuracy
 % Take correlation between true and inferred time courses
 scores.A.data   = NaN(sum(params.R), params.N);