I'm working with a problem where I only have several hundred training samples. That already sounds absurdly low, but I'm thinking: The relevant details for the problem are local patches within the 3d volume, rather than a single global feature. Since I have a 3D volume, the amount of data is effectively scaled up by the number of slices. So if I have 40 slices, it's more like I have several thousand samples rather than several hundred. With augmentation I can push that up further. Am I being overly hopeful here, or is there some truth in my thesis? Any papers to support?

Would a 3D CNN require less training samples than a corresponding 2D CNN?

I’m working with a problem where I only have several hundred training samples. That already sounds absurdly low, but I’m thinking:

The relevant details for the problem are local patches within the 3d volume, rather than a single global feature.
Since I have a 3D volume, the amount of data is effectively scaled up by the number of slices. So if I have 40 slices, it’s more like I have several thousand samples rather than several hundred.
With augmentation I can push that up further.

Am I being overly hopeful here, or is there some truth in my thesis? Any papers to support?

Get help from others!

Recent Answers