Controlling MATLAB Memory Use

MATLAB is a very useful tool for analysing medical image data. Unfortunately, it has one significant drawback: MATLAB memory management is terrible. The memory becomes easily fragmented, with no attempts made at compaction (except for the almost useless MATLAB pack function). With large medical image volumes, the results can be disastrous, leading to huge memory use, poor performance, and in the extreme case the termination of your MATLAB process by the operating system.

Controlling MATLAB memory use is a bit of a black art. There are general guidelines that can be followed, but no hard and fast rules that will work in every case. Here are some soft rules:

Re-use memory whenever possible. For example, the EMMA getimages function allows you to specify a block of memory to use. In general, this is the variable that already contains an image, and should be the same size as the images to be loaded:
```
PET = getimages(h,8,1:21,PET);
```
will re-use the memory assigned to the matrix PET, and will not allocate new memory for the returned images. This eliminates the fragmentation associated with reloading a set of images.
Use the EMMA rescale function to multiply matrices by a scalar. This function does not assign new memory, but instead uses the memory already assigned to the matrix. For example, the MATLAB instruction:
```
PET = PET .* 5;
```
will cause fragmentation since the matrix PET will be copied. However, the instruction:
```
rescale (PET,5);
```
has the same effect, but without causing fragmentation.
Initialize large matrices. This is useful when you intend to fill a matrix slowly through a series of operations. For example, if you intended to create a 128x128 image one row at a time in the variable Fake_PET, you should specify:
```
Fake_PET = zeros(128,128);
```
This will initialize the matrix to the correct size, and fill it with zeros. Then, when you fill the matrix in with your values, it will not be copied, and memory fragmentation will not occur.
Use small amounts of memory at a time. This is probably the most important rule, and can potentially solve all memory use problems. For example, when analyzing a dynamic study, consider breaking down your analysis so that it can be performed on one line (or several lines) at a time, instead of on the whole image.

For example consider the case presented in the basic EMMA image manipulation tutorial: we want to generate a volume of integrated images. The following snippet of code demonstrates the wrong way to do this:

h=openimage('/local/matlab/toolbox/emma/examples/yates_19445.mnc');
h2 = newimage('new.mnc',[0 15], ...
     '/local/matlab/toolbox/emma/examples/yates_19445.mnc');

ftimes = getimageinfo (h,'MidFrameTimes');

img = zeros (16384,1);

for j=1:15
   PET = getimages(h,j,1:21);
   img = ntrapz(ftimes,PET')';
   putimages(h2,img,j);
end;

closeimage(h);
closeimage(h2);

This brute force technique violates several of the above guidelines. It does not re-use memory when loading images from the MINC file. It transposes PET, which causes PET to be copied. It then transposes the result from ntrapz, which causes this result to be copied.

A better approach is to load the images for one slice, take the data line by line, and then re-assemble this into a final image. The following piece of MATLAB code has the same final result as the above, but uses much less memory (approximately 4 Megabytes).

h=openimage('/local/matlab/toolbox/emma/examples/yates_19445.mnc');
h2 = newimage('new.mnc',[0 15], ...
     '/local/matlab/toolbox/emma/examples/yates_19445.mnc');

ftimes = getimageinfo (h,'MidFrameTimes');

img = zeros (16384,1);

for j=1:15
   PET = getimages(h,j,1:21,PET);
   for i=1:128:16257;
      line = PET(i:i+127,:);
      img (i:i+127) = ntrapz(ftimes, line')';
   end;

   putimages(h2,img,j);
end;

closeimage(h);
closeimage(h2);

This page was created by Mark Wolforth (wolforth@pet.mni.mcgill.ca)