yeah, that was my suggestion, but after your last post I can see that you are doing something more elaborate than that.
Comment to your approach 1): As you state yourself, this doesn’t do much in means of machine learning. It would be the same to go to Fiji, apply a threshold of 0.5 and then do the quantification using this mask.
Comment to your approach 2): This can make a lot of sense if the background in smaller objects could be misleading, or if, in fact, pixel classification output is simply not good enough for the background.
Im pretty sure what you’re seeing here is the effect of the feature computation. Duration here is dependent on the size of the objects (in number of pixels that are included in the analysis). The classification step itself is cheap compared to this.
It might help to decrease the number of features you compute. On the other hand it might be worth investigating, what background mean values you get using e.g. Fiji with threshold of 0.5 on the background channel vs your approach 2), to at least quantify the improvement vs computation time.