How to change the mini-batch accuracy function using trainNetwork? (AKA How to change what the network is optimizing for)

I am using trainNetwork to train a network for semantic segmentation. I'd like to change the mini-batch accuracy function to be in terms of the boundary f1 score ( bfscore). It currently seems to be in terms of global accuracy, which is not optimal for my problem. How do I change this?

I have looked though the documentation for trainNetwork and trainingOptions and I did not see anything.

Thanks!