Motor imagery (MI) is one of the brain-computer interface (BCI) strategies for stroke rehabilitation. Accurate decoding and meaningful interpretation of the decoding model are important. In this talk, we will discuss deep learning methods in decoding motor imagery from EEG, in subject-specific, subject-independent, and subject adaptive scenarios. We will present results on data collected from both healthy subjects and stroke patients. We will then analyze the motor imagery data using interpretation approaches and try to shed some light on the correlation between the interpretation patterns and clinical outcomes in stroke rehabilitation.