[Neural Networks] Classification of a sick brain

[themagiciant95]

Hi,

i have a dataset with MRI of patients with a specific disease that affects the brain and another dataset with MRI of healthy patients.

I want to create a classifier (using neural networks) to classify if the MRI of a new patient show the presence of the ill or not.

First of all, i extracted the brain from all the MRIs (the so-called skull stripping) using BET tool found in FSL.

I have three questions for you

1) As the input to the training phase i want to give the whole extracted brains(possibly in the nii format), what kind of preprocessing steps do i need to apply once I've extracted the brains (before passing it to the classifier) ?

2) Do you know any better tool for skull stripping?

3) Do you know a tool (or library) that takes as input a nii files and allows me to create a classifier that uses neural networks ?


Thanks so much

 

[jedishrfu]

Is this basically an image classifier? where you would classify across a set of images?

I found this article on WIRED about fMRI and AI classifiers:

https://www.wired.com/story/fmri-ai-suicide-ideation/

It might give you some leads to research further.

 

[jim mcnamara]

Um, I think you are asking for something that may not exist in a form you can use.
An example: http://taco.cse.tamu.edu/pdfs/enspaper.pdf

This effort classifies neural network datasets. Try contacting the author to see where things stand with regard to out of the box software for your task. His email is in the paper as it seems to be a government agency.

Someone else here on PF may know more...

 

[lomidrevo]

Your questions looks very specific, I have no answers.
But I think I can provide you an advice:

As the input to the training phase i want to give the whole extracted brains

Do not use whole dataset for training! You need to keep some subset of data to validate the accuracy of your model.

 

[Jarvis323]

You can use NiBabel (in python) to load the nii files https://nipy.org/nibabel/gettingstarted.html .

And then you can feed the image into a CNN using TensorFlow https://www.tensorflow.org/overview or Caffe https://caffe.berkeleyvision.org/ or something like that.

You might need to experiment with the architecture a bit since the input is a 3D image rather than a 2D image which is more typical.

In terms of pre-processing, there is a lot you could do (http://www.humanbrainmapping.org/files/2015/Ed Materials/FSL_PreProcessing_Pipeline_OHBM15_Jenkinson.pdf), and it can get pretty complex. But also consider that the classifier only needs to be able to make predictions, not analyze the images from an anatomical and physical perspective.

Some neural networks such as a CNN, will be able to learn how to account for some issues all on its own, for example, orientation scale, etc. You might be fine to just try out a classifier on the raw images first, then begin trying some pre-processing steps to see if it helps.