How to make the layer automatically ignores a given pixel label during network training?

I would like to train a semantic segmentation network without considering a given class label during training. I want that pixels from that class do not propagate the loss.

I want to implement a cross-entropy loss function similar to this Python example. The function ignores the last label during loss computation so the network does not "learn" to classify a given class, although this class is present in the training patch.

I know about the dicePixelClassificationLayer function that says "The layer automatically ignores undefined pixel labels during training.", but how to set a given class label to "undefined" so that the layer automatically ignores it?

Please help!

Answer 1

[Matheus Ferreira]

Already solved: Just use the standardizeMissing function to substitut the ID of a given class by NaN. For example, if the class label to ignore is 99 run the following:

img = standardizeMissing(img,99);

Save variable img to the TIFF format and create a pixelLabelDatastore with the class names and label IDs without the class 99, for example. The pixels os the class 99 will be converted to <undefined> and the pixelClassificationLayer ignores undefined pixel labels during training.