I was looking at 3D CNNs that process volumetric data, e.g. for MRI images of brain, where the input is a 4D tensor, and I couldn't find images from the filters of the first layer.
Suppose that detecting a spherical shape is very important for the underlying task (e.g. classifying if a person has a tumor cancer or not). In that case, isn't detecting planes than edges more relevant? So, should the learned filters look like planes in 3D?
Moreover, if is it true that all CNNs learn to detect edges in the first layer, does this mean the kernel size dictates the number of filters we should use? For example, if our input is a grayscale image of dimensions $25 \times 25$ (we use a 2D CNN) and we use a kernel size of $3 \times 3$, is it meaningful to use more than $3$ filters? I mean the only edges that can be detected are those at angles: $0$ (vertical edge), $90$ (horizontal edge) and $45, -45$ (diagonal edges).
