Is it possible to perform both 2d and 3d convolutions in the same deep network?

I have a set of stereo images, each of size HXWX3. I want to use them and learn transformation between them. I used N branches (by skip connection) of several stride-1 2D convolutions and obtained the corresponding N volumes with HXWXM dimensions. Now, is there any way, so that I can arrange these N volumes in 4D as HXWXMXN, and perform 3D convolutions on it (as shown in the figure with N=2)?

I tried concatenateLayer to merge along dim=4, but it gave an error saying that the expected input size did not match!