Strategy for loading in multiple 3D volumes as input [Image Registration]

Hello! I’m wondering if there’s a “field” standard for registering 3D volumes. Namely, I have an input volume image (128 x 128 x 128), and I have a deformation field in each of the 3 axis (dx = (128 x 128 x 128), dy = (128 x 128 x 128), dz = (128 x 128 x 128)) that registers this image to a different time point.

My goal is for a network to learn to “apply” the deformation field to “warp” the input volume into a desired output. As a consequence, my input layer current needs to accommodate four (128 x 128 x 128) volumes, and flattening this results in a massive input layer! I find that downsampling my input volumes doesn’t work well, as it has to be scaled down to such a small 3D volume it is no longer recognizable.

My questions are:

  1. What are common strategies for feeding multiple 3D volumes?
  2. What are strategies to accommodate large, but flat input layers?


1 Like