In the field of medical imaging, content-based image retrieval (CBIR) techniques are employed to aid radiologists in the retrieval of images with similar contents. However, CBIR methods are usually developed based on specific features of images so that those methods are not readily inter-applicable among different kinds of medical images. This work proposes a general CBIR framework in attempt to alleviate this limitation. The framework is consisted of two parts: image analysis and image retrieval. In the image analysis part, normal and abnormal regions of interest (ROIs) in a number of images are selected to form a ROI dataset. These two groups of ROIs are used to analyze 11 textural features based on gray level co-occurrence matrices. The multivariate T test is then applied to identify the features with significant discriminating power for inclusion in a feature descriptor. In the image retrieval part, each feature of the descriptor is normalized by clipping the values of the largest 5% of the same feature component, and then projecting each normalized feature onto the unit sphere. The L2 norm is then employed to determine the similarity between the query image and each ROI in the dataset. This system works in the manner of query-by-example (QBE). Query images were selected from different classes of abnormal ROIs. A maximum precision of 51% and a maximum recall of 19% were obtained. The averages of precision and recall are 49% and 18% in this experiment.