Multiclass classifiers are very similar to binary classifier, you may need to change the last layer of your model to make the multiclass classifier output compatible with your model. There is a function available in MATLAB "pixelLabelDatstore", which can generate the pixel label images that in turn may be used as a label data target in your network for semantic segmentation.
Also, there can be many reasons to get a constant loss function, Data imbalance could be one. Try using a weighted multiclass Dice loss function instead of “crossentropy”.
If that does not help, try using an adaptive learning rate for your network. Also check the target images before feeding it to your network, sometimes the target and predictive images comes out to be transpose of each other because of how the MATLAB handles the data.
May be 3D tumor segmentation example can help you set up your model.
