We introduce an adaptive time-frequency plane feature extraction and classification system for the classification of motor imagery EEG recordings in a Brain Computer Interface task. First the EEG is segmented in time axis with a merge/divide strategy. This is followed by a clustering procedure in the frequency domain in each selected time segment to choose the most discriminant frequency features. The resulting adaptively selected time-frequency features are processed by principal component analysis - PCA for dimension reduction and fed to a linear discriminant classifier. The algorithm was applied to all nine subjects of the 2002 BCI Competition. The classification performance of our proposed algorithm varied between 70% and 92.6% for each subject, which gives an average classification accuracy of 80.6%. The algorithm outperformed the reference standard Adaptive Autoregressive model based classification procedure for all subjects. This latter approach had an average error rate of %76.3 on the same subjects. We observed that the time-frequency tiling selected by the algorithm for EEG signal classification differs from subject to subject. Furthermore, the two hemispheres of the same subject are represented by distinct time-frequency segmentations and features. We argue that the method can adapt automatically to physio-anatomical differences and subject specific motor imagery patterns.