shapeandsom log

Self Organising Maps (SOMs) log file

This document describes the script script_shapeandsom.m inside this package. It derives some of its development from the file script_shapeanalysis.m.

Initialising the SOM network

• Use somGetInputData to get some input data.

kx = 1;
sf.bintensities{kx} = sf.dataGL==clumplab(kx);
[sf.binput{kx}] = somGetInputData(sf.bintensities{kx});

• Bounding box for box inside cells. Define this into the same sf structure, and do it for each nuclei inside the

for kx=1:length(clumplab)
    uxuy = sf.regs(kx).Centroid - sf.regs(kx).EquivDiameter/4;
    wxy = sf.regs(kx).EquivDiameter/2;
    sf.sombb(kx,:) = [uxuy wxy wxy];
end

• Define the structure of the network

sf.netpos = somGetNetworkPositions('supergrid',[10 10], sf.sombb(kx,:));

• Finally, define the network itself:

OG = somBasicNetworks('supergrid', [10 10], sf.netpos);

The separation in this document is intended for explanation purposes. In reality, the wehole processing is made inside the same loop.

for kx=1:length(clumplab)
    % Get the input data from the binary image
    sf.bintensities{kx} = sf.dataGL==clumplab(kx);
    [sf.binput{kx}] = somGetInputData(sf.bintensities{kx});

    uxuy = sf.regs(kx).Centroid - sf.regs(kx).EquivDiameter/4;
    wxy = sf.regs(kx).EquivDiameter/2;
    sf.sombb(kx,:) = [uxuy wxy wxy];

    sf.netpos{kx} = somGetNetworkPositions('supergrid',[10 10], sf.sombb(kx,:));

    sf.OG{kx} = somBasicNetworks('supergrid', [10 10], sf.netpos{kx});
end
clear kx uxuy wxy;

Evolve the SOM to fit the binary cells in sf

Using matlab's functions, we can guarantee to start at a network evolution from the single frame, sf, that is perfectly matched to the binary segmentation of the independent cells.

netsize = [8 8];
ognet = selforgmap(netsize);
ognet.trainParam.showWindow=0;
ognet.trainParam.epochs = 200;

% Create the corresponding bounding boxes
fprintf('\n%s: Getting initial SOMs by fitting them to sf.\n', mfilename);
for kx=1:length(clumplab)
    % Get the input data from the binary image
    sf.bintensities{kx} = sf.dataGL==clumplab(kx);
    [sf.binput{kx}] = somGetInputData(sf.bintensities{kx});
    [sf.net{kx}] = train(ognet,sf.binput{kx}');
    sfpos = sf.net{kx}.IW{1};
    cfpos = sfpos + repmat(cf.xy(kx,:)-sf.xy(kx,:), size(sfpos,1),1);

    cf.OG{kx} = somBasicNetworks('supergrid', netsize, cfpos);
end
fprintf('%s: Initialisation of SOM finished.\n', mfilename);
clear kx uxuy wxy netpos cfpos;

• From here, the networks would need to be moved based on the cross correlation movement vectors.
• Then, a simultaneous evolution of both SOMs could be fitted to the new data.
  • An assumption is made here, that the slow movement of the cells would prevent an ambiguity or confusion between the networks.
• Because of the previous statement, there is room for two tests.
  1. Change sf and cf with time, so that sf always corresponds to t and cf corresponds to t+1.
  2. The other is to fix sf to frame t and move cf to frames t+t0

Now, some plots:

% plot the images and both 'moved' SOM points.
plotGraphonImage(cf.dataGL, cf.OG{2})
plotGraphonImage([], cf.OG{1})
% plot original and moved boundaries for all cells.
plotBoundariesAndPoints([],sf.boundy{1}, sf.boundy{2}, 'c-')
plotBoundariesAndPoints([],[], cf.movedboundy{1}, 'm-')
plotBoundariesAndPoints([],[], cf.movedboundy{2}, 'g-')

Move points over to position in frame frametplusT

This step was performed in the previous section, by taking the auxiliary variables sfpos and cfpos to get the positions from the MATLAB-trained network and traspose them into the new initial position determined by the cross correlation.

Evolve to new frame

There are two possibilities for the implementation

Two possibilities:

• Evolve, by using the clump at cf.
• Evolve by using the intensities and the clump, like cf.thisclump.*cf.dataGR
  • This involves trying the SOM with intensities when initialising from sf.