I am working with MRI volumes and I need to train a 2d segmentation network. I have multiple slices from each volume in my dataset. Is it fine to have multiple slices coming from the same volume in the same mini-batch? I am preprocessing by doing a zero-mean unit variance per volume based on the foreground voxel statistics, as is the common practice for most literature in deep learning for MRI in 3D (Nvidia,nnU-Net).
I believe I read somewhere some time ago that data in the same mini-batch should be uncorrelated but I cannot find that resource anymore. Any insights into this would be really helpful.