Muscle T2 relaxometry can be used to monitor disease activity in neuromuscular disorders. Dictionary matching of multi-echo-spin-echo (MESE) data is the gold-standard method to estimate the T2 of the myocitic component (T2-water) because of its ability to correct for multiple confounding factors, but suffers from a high computational burden. This work proposes a neural network (NN) approach for fast muscle T2-water mapping with subject-specific T2-fat calibration to overcome computational limitations of the dictionary method. The method was validated in-vivo against the standard dictionary approach. The NN application outperformed the dictionary approach in computational resources (x140 faster) while retaining quantitative accuracy.