Over the last few decades, the use of electroencephalography (EEG) signals for motor imagery based brain-computer interface (MI-BCI) has gained widespread attention. Deep learning have also gained widespread attention and used in various application such as natural language processing, computer vision and speech processing. However, deep learning has been rarely used for MI EEG signal classification. In this paper, we present a deep learning approach for classification of MI-BCI that uses adaptive method to determine the threshold. The widely used common spatial pattern (CSP) method is used to extract the variance based CSP features, which is then fed to the deep neural network for classification. Use of deep neural network (DNN) has been extensively explored for MI-BCI classification and the best framework obtained is presented. The effectiveness of the proposed framework has been evaluated using dataset IVa of the BCI Competition III. It is found that the proposed framework outperforms all other competing methods in terms of reducing the maximum error. The framework can be used for developing BCI systems using wearable devices as it is computationally less expensive and more reliable compared to the best competing methods.