What is "Instead of predicting the class label of the center pixel"

[Bonsen]

I found u say "Submission for Multimodal Brain Tumor Segmentation Challenge 2017 (http://braintumorsegmentation.org/). A patch-based 3D U-Net model is used. Instead of predicting the class label of the center pixel, this model predicts the class label for the entire patch. A sliding-window method is used in deployment with overlaps between patches to average the predictions." What is the "Instead of predicting the class label of the center pixel"? For example,

In training, if I input a 48x48x48 train patch and a 48x48x48 label patch in one "predicting the class label of the center pixel" network, after some conv layers, it may output a 40x40x40 train patch, so I have to crop label patch to 40x40x40 surround center pixel for input?

and in testing, if I also need to add some background pixels surround the test image? Like the training, I need to add 8 pixel surround the the image? if add 1 pixel, like : 0 0 0 0 0 0 0 2 2 2 2 0 0 2 2 2 2 0 0 0 0 0 0 0

thanks.