Assuming imperfect channel estimation, we propose an improved detector for orthogonal frequency-division multiplexing (OFDM) systems over a frequency-selective fading channel. By adopting a Bayesian approach involving the statistics of the channel estimation errors, we formulate an improved maximum-likelihood (ML) detection metric taking into account the characteristics of the channel estimates. As a first step, we propose a modified iterative detector based on a maximum a posteriori receiver formulation which reduces the impact of channel uncertainty on the decoder performance, by an appropriate use of this metric. The results are compared to those obtained by using a detector based on a mismatched ML metric, which uses the channel estimate as if it was the perfect channel. In a second step, we calculate the information rates achieved by both the improved and mismatched ML detectors, in terms of achievable outage rates. These outage rates are compared to those provided by a theoretical (not practical) decoder defined as the best decoder in the presence of channel estimation errors. Numerical results over both uncorrelated Rayleigh fading channels and realistic ultra wideband channels show that the improved detector outperforms the classically-used mismatched approach in terms of bit error rate and achievable outage rates, without any increase in the receiver complexity.