Suitable for 3D segmentation about iic HOT 1 CLOSED

It could potentially be applied to 3D segmentation. There are a number of ways you could modify our current pipeline. The simplest is just to slice the 3D image into disjoint 2D images and use our method as it currently is. Or you could make the slices overlap, so each point appears in more than one image, and get the average prediction (for more robustness). And/or, try to change the local spatial invariance (which is currently 2D, because it maximises MI between patches and nearby patches in x,y space) into 3D local spatial invariance. The more of these modifications applied, the more I would expect performance to improve.

from iic.