When working with sets of probabilities, basic information fusion operators quickly reach their limits: intersection becomes empty, while union results in a poorly informative model. An attractive means to overcome these limitations is to use maximal coherent subsets (MCS). However, identifying the maximal coherent subsets is generally NP-hard. Previous proposals advocating the use of MCS to merge probability sets have not provided efficient ways to perform this task. In this paper, we propose an efficient approach to do such a merging between imprecise probability masses, a popular model of probability sets, and test it on an ensemble classification problem.