Multicomponent T2 analysis (mcT2) yields a voxel-wise distribution of T2 values, which can be used to estimate sub-voxel information such as myelin content. Producing such data, however, remains challenging due to the large ambiguity in the T2 space. We present a data-driven approach for mcT2 analysis, which learns the anatomy in question and identifies microscopic tissue-specific features as a preprocessing step. It then utilizes them for analyzing each voxel locally using a designated optimization scheme. Experiments in human brain data show reproducible myelin content estimations at clinical settings without any prior assumptions.