UNet with 3D input and 2D output?

[dzenanz]

I am creating a NN for segmenting ultrasound videos. Currently, I am using 3D approach, inputting the last 4 frames (512x512x4), getting the segmentation in the same shape, then discarding the first 3 frames, and only using the current frame. However, this feels wasteful.

Is it possible to output a single frame segmentation (512x512x1)? If yes, is it only the question of getting the strides parameter (and maybe upsample_kernel_size) to DynUnet correct, or is there more to it?

[KumoLiu]

Hi @dzenanz, you appear to be using 4 as a channel; perhaps you could make it a channel in the first dimension and then simply change the output_channels of the network.
Thanks!

[dzenanz]

Thank you, it works. It is retroactively obvious 😄