Accurate diagnosis is important for successful treatment of brain tumor. Content based medical image retrieval (CBMIR) can assist the radiologist in diagnosis of brain tumor by retrieving similar images from medical image database. Magnetic resonance imaging (MRI) is the most commonly used modality for imaging brain tumors. During image acquisition there can be misalignment of magnetic resonance (MR) image slices due to movement of patient and also the low level features extracted from MR image may not correspond with the high level semantics of brain tumor. These problems create a semantic gap and limit the application of automated image analysis tools on MR images. In order to address these problems, this paper proposes a two-level hierarchical CBMIR system which first classifies the query image of brain tumor as benign or malign and then searches for the most similar images within the identified class. Separate set of rotation invariant shape and texture features are used to discriminate between brain tumors at each level. Experiments have been conducted on medical image database consisting of 820 brain MR images. The proposed approach gives promising retrieval results by improving precision, recall and retrieval time.