3D convolution is not invariant to the plane of MRI

Hi All,

I am applying 3D CNN to 3D-MRI data to solve a regression problem. It is worth mentioning that MRI acquisition of the data I am using is 3D (unlike 2D MR acquisition which is also referred as 3D in deep learning context due to the availability of multiple slices). What I have experienced is that training CNN on different planes (axial, sagittal, coronal) gives different results, and axial performs better on training while for testing I get good results if I change plane to sagittal.

These type of differences does make sense in 2D data with 2D CNN, because 2D CNN sees only one slice at a time, but I do not understand why 3D convolution is not invariant to spatial planes that are used for training and testing.

Any information in this regard, or literature that analyses the effect of imaging planes on the performance of 3D CNN would be very helpful.

Hi Iram,

Are the voxels isotropic for the 3D MRI volume? Also, what part of the body has been imaged?


Have you tried a 3D CNN that is rotation invariant?

3D CNN is designed to be rotation invariant

It is contrast-enhanced T1w isotropic high-resolution (1mm3) 3D brain MRI


I have not tried it yet. However will a simple pre-processing would be enough to solve the problem such as bringing all volumes to same orientation in both x,y and z plane?

Yep, it will certainly be a simpler approach to align all volumes to the same orientation. The rotation invariant approach may be a bit more involved.

1 Like
Category icons made by Freepik from www.flaticon.com