A general framework for content-based medical image retrieval with its application to mammograms

In the medical field, content-based image retrieval (CBIR) is used to aid radiologists in the retrieval of images with similar contents. CBIR methods are usually developed for specific features of images, so that those methods are not readily applicable across different kinds of medical images. This study proposes a sound methodology for CBIR of mammograms, which is applicable to various formats of medical image. The methodology is divided into two parts-image analysis and image retrieval. In the image analysis part, 19 abnormal regions of interest (ROI) and 20 normal ROIs are selected as samples for the whole 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 examine the significance of the differences for these 11 textural features from normal and abnormal ROIs. The discriminating features are incorporated into a feature descriptor for the ROI. This descriptor is embedded into the CBIR system. In the image retrieval part, a CBIR system for mammograms is developed. For normalization of feature vectors, a novel technique is proposed to clip the values of feature elements of the top 5%, and then project each image feature onto the unit sphere. To determine the similarity between query image and each ROI in the dataset, the L₂ norm is used to measure the similarity between two images. This system was designed by query-by-example (QBE). Query images were selected from different classes of abnormal ROIs. To evaluate the performance of the CBIR system, the precision and recall were measured. A maximum precision of 51% and recall of 19% were obtained using the gray level co-occurrence matrices and a distance of 5. The averages of precision and recall are 49% and 18 % in this experiment.