Jun132013June 13, 2013October 4, 201313Commentsby L.

Local cross corr images

Posted in Software

click to go to papers... this is from the Supplement to Smith and Hausser 2010

One key step in analyzing data form calcium imaging experiments is segmentation (i.e., drawing regions of interest, ROIs). There have been some nice discussions in the literature (1, 2, 3, 4). Experiments with calcium dyes like OGB-1-AM, and those with genetically encoded indicators like GCaMP, require different approaches for drawing ROIs.

Here’s some code I’ve used to generate a first pass of ROIs. The simple idea is that the pixels that are located within cells will be highly correlated because during a fluorescence transient they will all exhibit similar time courses, and other regions (e.g., neuropil, blood vessels, etc.) will not be as correlated.

The input is an X x Y x T 3D matrix (time or frame is the third dim). The output is an X x Y cross corr image. As written below, it looks at the 3 x 3 local neighborhood. If you increase the variable w to 2, then it will look at a 5 x 5 local neighborhood, and so forth.


function [ccimage]=CrossCorrImage(tc)

w=1; % window size

% Initialize and set up parameters
ymax=size(tc,1);
xmax=size(tc,2);
numFrames=size(tc,3);
ccimage=zeros(ymax,xmax);

tic
for y=1+w:ymax-w
    for x=1+w:xmax-w
        % Center pixel
        thing1 = reshape(tc(y,x,:)-mean(tc(y,x,:),3),[1 1 numFrames]); % Extract center pixel's time course and subtract its mean
        ad_a   = sum(thing1.*thing1,3);    % Auto corr, for normalization later
        
        % Neighborhood
        a = tc(y-w:y+w,x-w:x+w,:);         % Extract the neighborhood
        b = mean(tc(y-w:y+w,x-w:x+w,:),3); % Get its mean
        thing2 = bsxfun(@minus,a,b);       % Subtract its mean
        ad_b = sum(thing2.*thing2,3);      % Auto corr, for normalization later
        
        % Cross corr
        ccs = sum(bsxfun(@times,thing1,thing2),3)./sqrt(bsxfun(@times,ad_a,ad_b)); % Cross corr with normalization
        ccs((numel(ccs)+1)/2) = [];        % Delete the middle point
        ccimage(y,x) = mean(ccs(:));       % Get the mean cross corr of the local neighborhood
    end
end
toc

I then use an adaptive threshold algorithm and the MATLAB function “bwlabel” to get a quick, first pass set of ROIs. Once you have your ROIs, you can select weights for the pixels within the ROI, if need be.

Tags: