Learning to re-rank medical images using a Bayesian network-based thesaurus

Abstract

In this paper, we believe that representing query and images with specific medical features allows to bridge the gap between the user information need and the searched images. Queries could be classified into three categories: textual, visual and combined. We present, in this work, the list of specific medical features such as image modality and image dimensionality. We exploit these specific features in a new medical image re-ranking method based on Bayesian network. Indeed, using a learning algorithm, we construct a Bayesian network that represents the relationships among these specific features appearing in a given image collection; this network is then considered as a thesaurus (specific for that collection). The relevance of an image to a given query is obtained by means of an inference process through the Bayesian network. Finally, the images are re-ranked based on combining their initial scores and the new scores. Experiments are performed on Medical ImageCLEF datasets from 2009 to 2012 and results show that our proposed model enhances significantly the image retrieval performance compared with BM25 model.