1
$\begingroup$

I trained an infection segmentation models on a large dataset of CT scans, and want to extend it to other datasets to show the ability of the model to generalize. What I found though, is that CT scans look very different. For example,

  1. CT scan from the first datasets,

  2. Image from another dataset

I tried three things:

  1. Global mean/standard deviation (subtract from each image the same vector $m$ and divide by $s$),
  2. Local mean/standard deviation (from each image subtract its $m$ and divide by its $s$, os that every image is normally distributed with mean $m$ and std $s$,
  3. Gamma correction ($\gamma=0.01$), see the image at the bottom.

None of which worked. My question is: is it possible to 'normalize' other CT scans, 'normalize' in a sense make them more similar to the one on which I trained the model?

enter image description here enter image description here

enter image description here

$\endgroup$
4
  • $\begingroup$ What specifically are the differences? Could you provide some examples of images from the datasets you are using? $\endgroup$
    – Ben
    Oct 2, 2020 at 0:32
  • $\begingroup$ There are 3 images in the post $\endgroup$
    – Alex
    Oct 2, 2020 at 0:49
  • $\begingroup$ I assumed these images represent the three preprocessing methods you tried, considering you say in bullet three to see the third image. Would you be able to provide examples of raw scans from the different datasets so we can see what the significant differences are? (Or did I misunderstand and the three images do not correspond to the three preprocessing methods?) $\endgroup$
    – Ben
    Oct 2, 2020 at 1:27
  • $\begingroup$ The first one is the one I used for training, the second one is the raw image, the third one is the second after processing $\endgroup$
    – Alex
    Oct 2, 2020 at 9:15

1 Answer 1

0
$\begingroup$

There is an extensive literate and established software to normalize CT scans. It is required for medical research to conduct group level analysis. The most common process involves transforming to standard template. The paper "Age-specific CT and MRI templates for spatial normalization " goes into greater detail.

$\endgroup$
1
  • $\begingroup$ This would have been my instinct as well. I was going to say normalize to a common intensity histogram. $\endgroup$ Nov 8, 2021 at 15:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.