1
$\begingroup$

I'm trying to solve a medical imaging regression problem using a CNN. Each of the samples in my data set consists of one, two, or three of the following file types:

  • flair.nii.gz
  • mprage.nii.gz
  • swi.nii.gz

Each of the files is a three- or four-dimensional matrix of voxel values between 0 and 255.

I don't want to throw information away since I have a limited number of data samples. Is there a technique for working around those cases which have a 'missing' file (or two)? For example, could I construct a matrix consisting of, say, all zeroes of the correct dimension and size and use this as a replacement for the file? Would this work or would it lead to problems with the CNN? Of course, the samples that I want to make inferences on might also have missing files, so, maybe, a matrix with all zeroes would be a type of information.

Anyway, what does one do in a case like this?

Improve this question
$\endgroup$

2 Answers 2

Reset to default
0
$\begingroup$

The general version of your problem is described a bit in this stackexchange post. However, it's hard to know how to address the issue unless the missing file dropout is better understood. For example, if the missing files are truly iid, then zeroing out the tensor might be ok. If one of your files drops out 2 or 3x more than another, zeroing out will skew towards the other features.

You also specified that the amount of data you have is small, which might make dropout less effective with a net, and you might have better luck with hand crafted features with a decision tree.

Ultimately, you just have to try things. I would start by making sure you can overfit on the small set of examples without missing data and then try to overfit again with the missing data to see how, or if, your metrics degrade. This will give you better insight for what to try when you train with a proper train / test / validation split.

Improve this answer
$\endgroup$
-1
$\begingroup$

I am not sure if I understood your question. Basically, you have data with different dimensions. For example you have image data HxWxc (3dimension) and you have video data, HxWxcxt (4dimension). What you are saying is that padding zeros to the image to make a video. That does not really make sense to me.

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ As it’s currently written, your answer is unclear. Please edit to add additional details that will help others understand how this addresses the question asked. You can find more information on how to write good answers in the help center. $\endgroup$
    – Community Bot
    Commented Jul 30, 2022 at 16:34
  • 1
    $\begingroup$ No, I mean that some files are missing altogether. $\endgroup$
    – Paul Reiners
    Commented Jul 30, 2022 at 18:42

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .