Well there are tons of things you could measure. Area fraction, nearest neighbor distances, area distribution, blob intensities, etc. Literally dozens upon dozens of things.
One approach is to just measure everything you can think of, then use principal components analysis, pca(), to make a model. You could threshold on the first principal component if there are two classes.
Or you could try kmeans() if you know there are two classes.
Or you should try the Classification Learner app to find out which classification/prediction scheme works best given the measurements you have made. Often random forest, though very ad hoc, turns out to be best. Or try several of them and have them "vote" like "adaboost" methods.
But don't discount deep learning - it may end up being the best of them all.
Best Answer