Content-based retrieval of medical images with relative entropy

Medical image databases are growing at a rapid rate because of the increase in digital medical imaging modalities and the deployment of Picture Archiving and Communication Systems (PACS), Electronic Medical Records (EMR) and telemedicine applications. There is growing research interest in Content-Based Image Retrieval (CBIR) of medical images from such digital archives. A new distance function for CBIR is presented for measuring the similarity between two images. The distance function is a variant of relative entropy, or the Kullback-Liebler distance. The new distance is the sum of the relative entropy of the two images to each other. The latter is a symmetric non-negative function and is only zero when the two images have identical probability distributions. This method has been implemented in a prototype system and has been applied to a database of medical images. Initial results demonstrate improvements over L-1-norm and L-2-norm histogram matching. The method is computationally simple since it does not require image segmentation. It is invariant to translation, rotation and scaling. The method has also been extended to support retrieval based on Region-Of-Interest (ROI) queries.