Hi. I am developing a machine learning model to predict the diagnosis of a subject using his MRI data preprocessed using Voxel-based Morphometry. So, I have two groups of subjects (patients and control), and each person has a 3D volume with 128x128x128 voxels in MNI152 space.
The problem is that the groups have significantly different proportions of gender (verified using Chi-square proportion test). In another type of data, I would "regress out" (using a linear detrending algorithm) the effect of the non-imaging variable. However, I am not sure if this would make sense in this kind of data.
Is it make sense "regress out" the gender (or age and head size) from the intensity value of the voxel? What would be the best way to address the confounding effect of these non-imaging variables?
PS: I am using a deep neural network as the classifier.
If the underlying algorithm was linear, 'regress-out' approach would be somewhat applicable, yet, since this is ML and deep-learning, linear decomposition doesn't really make sence - because, if the relationship between, say, age, and Morphometry metrics is not linear, your regression would basically corrupt the data, when without the regression the non-linar algorithm would probably identify the relationship quine nicely.
So why not just incorporate age into the model? Add another input into your model, and pass age into it. Same with gender and any other characteristic of the subjects. With deep learning it's reasonable to assume the model would make appropriate use of additional inputs, if given enough samples. But, well, the same goes for any ML algorithm
Since this incorporation is a bit different than, say, adding several more voxels, and carries a more 'modulation-like' sence, I would assume that the NN would require more layers/neurons to 'build' the underlying relationship. However I do not know the starting point so the structure used (or planned for use) may easily be enough already.
Significant between-group difference isn't perfect, but is not a deal breaker, in my opinion
- Badges
- Science topic