How to implent 3D semi-supervised segmentation?

I need to train a 3D segmentation network, for example, 3D U-net. Each training volume contains 48 slices but I only have annotations of 1 slice for each volume. In this case how can I train the network in a semi-supervised manner?

I’m thinking using 1 slice annotation and just simply ignore the loss from rest 47 slices, it is correct?

Any suggestion is appreciated!