scr iterevol areacontrol

Evolution of shape: restrictions and active contours' parameters

This log file is a continuation of the file scr-iteravolution-vs-leaking.md. The focus of this (hopefully small) log file is to address some issues that occur with the segmentation in the previous file regarding the shape of the cell being segmented and the control we can realistically have over the MATLAB function activecontour.m.

We know about the parameters of activecontour.m that can be addressed, they involve:

• Method: Choose from 'Chan-Vese' and 'edge'
• Iterations: How many times the program is run. Typical values used here are options.iter=25 and options.iter=50.
• Smooth Factor: Parameter that forces the contour to be soft, smooth or regular. Typical values used in this application are options.smoothf= 1.5 and options.smoothf= 2, when the method used is Chan-Vese.
• Contraction Bias: Is the tendency of the contour to expand or shrink.

This function does not allow a finer control over the parameters, the evolution or the finer details (restricting size, volume or area of the contours).

Minor setback exposing a problem: Cell 2

Upon trying to compute the method for all the independent cells, some mistakes in the evolution of the shape were found. And because of the nature of the evolution, where a a frame at t(k+1) is created from the previous t(k+1)=tk+1, the errors in segmentation in one frame can be passed on to the subsequent ones.

For this experiment, the cell at tracklabel=2 was examined in two different experiments: from frames 28:300 and 202:300. In this case, there were many objectives:

• Follow a cell over long span of frames, and see how the following mechanism would cope.
• Follow a single cell even though it might go through clumps.
• See how different would the cells be segmented when starting at different points.
  • How would cell 2 look at frames >230 when starting at frame 28?
  • how would it look upon starting at frame 202?

Some results were found, the .gif files appear below

Frames 28:300

Frames 202:300

The method is able to cope with the following of the cell, but

1. Segmentation errors are passed on between frames and only in specific scenarios does the segmentation fixes itself.
2. In frames where a clump is found, the segmentation falls into even more traps or leaks.
3. Noticeably, the segmentation from frames 230:300 becomes pretty much the same, even with the very different segmentation results at frame 202. The measurements were taken, just as in the script file scr-iteravolution-vs-leaking and the measurements matched after some frames (not shown).

Control of area (and other shape measurements)

Generic examples

For this section of the log, an initial analysis is made with the activecontour function. A generic example is tried.

I = imread('toyobjects.png');
figure(1)
imagesc(I);

mask = roipoly;

figure(2)
imagesc(mask);
title('Initial MASK');

Now the parameters, and the actual function:

acopt.method = 'Chan-Vese';
acopt.iter = 200;
acopt.smoothf = 1.5;
acopt.contractionbias = 0;
acopt.erodenum = 5;

evomask = activecontour(I, mask, acopt.iter, acopt.method, ...
'ContractionBias',acopt.contractionbias,'SmoothFactor', acopt.smoothf);
figure(3)
plotBoundariesAndPoints(I, bwboundaries(mask), bwboundaries(evomask), 'm-');

Real data tests

• Run initscript so we have access to the handles structure.
• Images where some problems can be found for cell 2 are frames 43, 96.

The painless way to run tests:

initscript;

Then just load everything:

a = getdatafromhandles(handles, filenames{43})
clumplab = 2;
wuc = 2;
trackinfo = [tablenet(ismember(tablenet.track, clumplab),[5 1 2 11 13 14]) ...
clumptracktable(ismember(tablenet.track, clumplab),:)];

Plot the image (with current segmentation) and get a different one with roipoly, that will simulate a previous step:

plotBoundariesAndPoints(a.X,...
   bwboundaries(a.dataGL==trackinfo(trackinfo.timeframe==43,:).seglabel));
mask = roipoly;

Get the parameters for the activecontour,

acopt.method = 'Chan-Vese';
acopt.iter = 200;
acopt.smoothf = 1;
acopt.contractionbias = -0.25;
acopt.erodenum = 5;

finally, just run the routine:

evomask = activecontour(a.dataGR, mask, acopt.iter, acopt.method, ...
'ContractionBias',acopt.contractionbias,'SmoothFactor', acopt.smoothf);

The same funciton roipoly was used to simulate a smaller or larger segmented area for frame t(k-1). Then, various tests with the function activecontour were done in order to get a sense

Growing

method	iter	smoothf	contractionbias	erodenum
Chan-Vese	200.00	1.00	-0.25	5.00

Shrinking

method	iter	smoothf	contractionbias	erodenum
Chan-Vese	100.00	1.25	0.10	5.00

These parameters may need to be changed when addressing clumps.

Developing a better area control

The approach described before was implemented in a crude way. At each step, the size (Area) of the new frame is compared with the previous size. A large change results in the adjustment of the

An example of size correction.

Right before a clump.

A clump confusion.

A bigger confusion.

A loss of the original cell because of the confusion.

Some area measurements