Why do we need to get a random sub-volume?

It was my understanding from the course that we break the problem down to sub volumes because it is more computationally efficient that way. But I assumed that the sub-volumes would be in sequence (starting at a top corner and ending at the bottom at the opposite corner for example). Can someone please explain to me the reason for getting a random sub-volume that falls anywhere within the 3D image?

1 Like

@Hesham1 Hello and thanks for participating in the community.

You are actually right. In general sequential or stratified overlapping patches/volumes would be probably more accurate. But also certainly computationally more intensive. Random sub volumes will increase variance but from what I read, yes, could lead to lower accuracy.

So it is a tradeoff and therefore sometimes I believe combined (random and non random) strategies are used.


1 Like

Hi @Hesham1,

Great question!

Well the reason behind randomly sampling sub-volumes instead of doing it sequentially is to avoid bias and avoid learning local correlations or dependencies between the sub-volumes.

In the case of bias, the model can sometimes learn unique patterns if we go from top to bottom sequentially and then falls into assumptions.

In the case of dependencies, the model can learn unique correlations because of sub-volume proximity, which might not always be correct.

Therefore, introducing variability and randomness can help in developing a robust model.

I hope that helps!