How do I use indexing (vectorization?) rather than for loops for dynamic field referencing into a cell array?

I think there's a way to eliminate for loops by indexing, or perhaps it's called vectorizing. How do I do this?

For example, I'm trying to plot several vectors of different lengths in the same 3D scatter plot. This code stores 34 DICOM ROI items' double arrays in a 1x34 cell array:

info = dicominfo('filename.dcm');
numberofarrays = 34;
datacells = cell(1,numberofarrays);
names = fieldnames(info.ROIContourSequence.Item_9.ContourSequence);
for loop = 1:numberofarrays
   datacells(loop) = {info.ROIContourSequence.Item_9.ContourSequence.(names{loop}).ContourData(:,1)};
end

but this code results in the following error:

datacellsfaster(1:34) = {info.ROIContourSequence.Item_9.ContourSequence.(names{1:34}).ContourData(:,1)};

Expected one output from a curly brace or dot indexing expression, but there were 34 results.

Why was it expecting only one output? How do I eliminate for loops through indexing?

Answer 1

[Guillaume]

I'm suprised you get Expected one output from a curly brace or dot indexing expression as an error and not Argument to dynamic structure reference must evaluate to a valid field name.

Unfortunately, you cannot pass a cell array as dynamic field names, The dynamic name must be a scalar string so,

.ContourSequence.(name{1:34})

is never going to work.

If all the ContourData have the same size you could eliminate the loop with struct2cell and by concatenating all the CountourData into one big matrix. There's no guarantee that it'd be faster, and it certainly would use up a lot more memory.

You're better off with the loop. In any case, with only 34 iterations that loop should execute very quickly.

[Daniel Bridges]

Thank you for the reply.

What do you mean by the phrase 'scalar string'?

The MathWorks Empty Matrices, Scalars, and Vectors page says that a scalar value is a 1-by-1 matrix, whereas their String handling page says a string is a vector containing Unicode. The former page says a vector is a 1-by-n or n-by-1 matrix (n>1). Hence a string cannot be a scalar, and so I do not know what you mean by 'scalar string'.

Incidentally, I have since learned a little more about indexing. For example:

for loop = 1:numberofarrays
  %index = find(z==eachdepth(loop),1); % find the z column with value matching depth
  % store all x and y values for that depth
  xpoints(loop) = {totalROIpoints(totalROIpoints(:,3)==eachdepth(loop),1)};
  ypoints(loop) = {totalROIpoints(totalROIpoints(:,3)==eachdepth(loop),2)};
end
%%scratchwork
% totalROIpoints = [x;y;z]';
% totalROIpoints(1:30,:);
% totalROIpoints(totalROIpoints(:,3)==-1647,:);
%ans1 = totalROIpoints(totalROIpoints(:,3)==-1647,:);
%ans1(:,1);
%ans2 = totalROIpoints(totalROIpoints(:,3)==-1647,:)
%ans3 = [totalROIpoints(totalROIpoints(:,3)==-1647,1) totalROIpoints(totalROIpoints(:,3)==-1647,2) totalROIpoints(totalROIpoints(:,3)==-1647,3)]

[Guillaume]

'scalar string' is an abuse of language, yes. I mean a single string as opposed to a cell array of string.

I'm not sure what you're trying to say with your latest example. You can replace the loop by using ismember and accumarray but I'm not sure it will be any faster:

for loop = 1:numberofarrays
   xpoints(loop) = {totalROIpoints(totalROIpoints(:,3)==eachdepth(loop),1)};
   ypoints(loop) = {totalROIpoints(totalROIpoints(:,3)==eachdepth(loop),2)};
end

is equivalent to

%assuming eachdepth has numberofarrays elements
[indepth, row] = ismember(totalROIPoints(:, 3), eachdepth);
xpoints = accumarray(row(indepth), totalROIPoints(indepth, 1), [], @(x) {x});
ypoints = accumarray(row(indepth), totalROIPoints(indepth, 2), [], @(x) {x});

[Daniel Bridges]

Thank you for clarifying and introducing me to two additional commands. My goal is to pull multiple rows with the same "z" value of an n-by-3 data matrix*, so ismember appears more elegant than looping over the same data repeatedly. Is it faster, though? It might require the same number of array-scans, but would they occur faster from not being in a top-level (user) script? accumarray seems not what I want to do, though, since it sums the resulting value each time it is referenced; rather, I'm seeking to merely extract a series of coordinates. (I'm using cell arrays because each depth has a different number of coordinates.)

My latest example is merely to show progress in knowledge of index use, i.e. using logical indices to pull data. I had initially thought to loop using find, but then with that scratchwork figured out how to do it via indices instead. This relates to my OP because I was actually seeking clarity both for dynamic field referencing and using indices instead of loops to create faster, more elegant code.

*I constructed the n-by-3 data matrix from a cell array from x1y1z1...xnynzn vectors of different lengths stored in a structure. Of course it would be better to pull the code directly, but searching documentation and online it seemed necessary to import the data from structures to cell arrays ...

[Guillaume]

ismember, as a built-in function, is going to be faster than a loop. The bottleneck is going to be the accumarray calls due to the use of a non-standard accumulation function (the @(x) {x} anonymous function).

In my example, accumarray does not sum the resulting values, it extracts the series of coordinates exactly as you want. The accumulation function I've specified simply wrap the accumulated vectors (all the values that match a depth) into a cell array.

find is indeed often unnecessary and using logical indexing directly will speed up your code by a small amount indeed.

By the way, I forgot to say, another way to write your initial loop

for loop = 1:numberofarrays
   datacells(loop) = {info.ROIContourSequence.Item_9.ContourSequence.(names{loop}).ContourData(:,1)};
end

would be, assuming that numberofarrays == numel(fieldnames(info.ROIContourSequence.Item_9.ContourSequence)):

datacells = structfun(@(seq) {seq.ContourData(:, 1)}, info.ROIContourSequence.Item_9.ContourSequence);

structfun iterates over all the fields of ContourSequence and, again, the anonymous function simply wrap the ContourData of each field into a cell array. I don't expect any significant gain in speed with that syntax, the looping may be faster but you now have the cost of a function call (to the anonymous function).

Answer 2

[geotocho]

Rather than curly brace indexing, the vectorized way of assigning a command to all cells uses parentheses. Much like the traditional array but in this example A and B are cell arrays of the same size. I wish to assign the row of cells in A to B. In this manner dot indexing is allowed.

B(1,:) = A(1,:);

[Daniel Bridges]

Thank you for this comment. I will try to remember it if I should encounter this situation in the future.