A semi-blind Expectation-Maximization (EM) channel estimation algorithm is proposed for 50 Gb/s quadrature phase shift keying (QPSK)-Discrete MultiTone (DMT) signal transmission systems using intensity modulation/direct detection (IM/DD) over 100 km standard single mode fiber (SSMF). The reported channel estimation methods for DMT systems can be roughly divided into two categories: semi-blind channel estimation and blind channel estimation. Due to the low accuracy of traditional blind channel estimation algorithm and lower spectral efficiency of the semi-blind channel estimation algorithm relying on more training sequences (TS), EM algorithm is proposed that is a two-step iterative procedure to maximize the likelihood function for achieving channel estimation instead of classical semi-blind channel estimation methods with more TS. Also, we assume the channel of the IM/DD DMT system as an additive white gaussian noise (AWGN) channel. Simulation results show that using EM algorithm yields about 2 dB optical signal noise ratio (OSNR) improvement at a bit error ratio (BER) of 3.8×10-3 compared to classical channel estimation based on TS under the same number of TS and has the similar performance compared to classical channel estimation relying on more TS. In addition, it is shown that at high OSNR (>19 dB), the performance of EM algorithm outperforms that of LMS algorithm. On the contrary, the performance of least mean square (LMS) algorithm outperforms that of EM algorithm at low OSNR (<19 dB).