Query regarding using image3Dinputlayer

I am designing a 3D CNN. I have 3D images whose first two dimenisons are height and width. The third dimesion is number of images which keeps on varying. Hence my 3D images are of varying length.

The format of image3Dinputlayer is image3dInputLayer(inputSize) where input size is [h w d c] i.e. height, width, depth and number of channels.

Can we replace the third dimension 'depth' by number of images?

Since, my number of images will keep on varying , so I want to ask:

Can any of the dimension height width or depth keep on varying while training 3DCNN ?

In case, requiring any clarity in the question feel free to comment.

Answer 1

[Mahesh Taparia]

Hi

To train a CNN, the input size needs to be same. The possible workaround in your case can be, try to append a matrix of zeros/ any features in order to have the number of depth channels to be equal to maximum possible depth channels. Else, try to reduce the dimensions to fix number of channels using dimensional reduction technique like pca. Hope it will help!