Hi Lana,
The problem is not with the software. The problem is that 50 particles just isn't a useful number, unless your particles exhibit a strongly preferred orientation. The whole point of single particle analysis is that you have many identical objects which are randomly oriented in solution. If you think about a particle in some orientation, then consider how a projection of that particle would change if you rotate it, say, 10 degrees out of the plane. The particle appearance changes quite a bit. If you were to average these two particles together you would be smearing out most of the features.
Even with this large angular step, there are over 200 different unique views of an asymmetric particle! If you reduce the step to 5 degrees, this number jumps to over 800, and 5 degrees is still a pretty significant change. So, if you have randomly oriented asymmetric particle and want to have on average 5 particles per class average, this would mean you would need to start with at least 1000 particles, and your class averages would still be pretty fuzzy.
Now, if you have a strongly preferred orientation, you may be able to do 2-D averages with far fewer particles, but in that case, you will not be able to do 3-D well, even if you incorporate tilting into your data collection.
If your particle has symmetry, say a 4-fold rotational symmetry, then the minimum number of required particles is reduced by this factor. So people looking, for example, at icosahedral viruses (60-fold symmetry) can make class-averages with a pretty small number of particles.
No, you should not get any better 2-D classification results in Relion. EMAN2's new 2-D classification works extremely well, though there are differences between the goals of EMAN2 and Relion in this respect, so results will not be identical. It's probably not worth getting into the details of that here.
Aggregation is a very common problem in TEM, but it is largely a problem to solve experimentally rather than computationally. ie - come up with a better buffer, make sure you are diluting the specimen following methods which limit aggregation, alter grid conditions, etc. In the end it is unlikely that you are going to be able to answer useful biological questions with fewer than a few thousand particles. (there are exceptions to this rule, of course)