Question about training the model

[mpegia]

I read your paper and the original of UNet++ and everything is clear to me apart from how I can I train the network using a new dataset.
There is the option of deep supervision in which the nework takes the co-registered image pairs (each has size 256x256x6) concatenated and as
output the [output1, output2, output3, output4], with each output dimension of 256x256x1.
I would like to reproduce your experiments with deep supervision.
It is clear to me what the input should be, but I fail to understand what the output should be. Specifically, how can I produce these 4 output matrices.
In the dataset from Lebedev only one grayscale image was used as output, and each pixel of this image was produced from subtracting the two input images.