Pathology-based vertebral image retrieval

Searching for vertebrae in a large collection of spine X-ray images that are relevant to pathology is potentially important for providing assistance to radiologists and bone morphometrists. Developing appropriate methods for such searching tasks is very challenging due to the high similarities among vertebral shapes in contrast to the subtle dissimilarities that characterize the pathology. In this paper, we target two aspects of this problem: first, we develop mathematical features that can effectively represent the biomedical characteristics of interest; second, we exploit similarity learning to enhance and try to optimize the retrieval performance. We evaluate our proposed method on an expert-annotated dataset of 856 vertebrae and demonstrate its retrieval performance by precision-recall and average-precision graphs. We also demonstrate how we have integrated our method into our Web-accessible spine X-ray image retrieval system.