Hey @cvanitnelav, sorry for the late update 
I’ve been processing the data you sent me, and I think I’ve been able to replicate what you saw. Here’s my initial 2D classification, and then the 2D classification on the featureless classes:
I turned off class sorting by # of particles so that I can better follow the classes as they change through the O-EM iterations, and what I saw is that some of the featureless classes have features in earlier iterations, but fade into noise-like blobs by the end. Do you see something similar in your own job?
If so, others who have run into issues like this (see this post and this post) have found success with switching off the noise model, increasing the number of iterations and/or the batch size, or increasing the initial classification uncertainty factor.