Multimodal information retrieval based on DSmT . Application to computer - aided medical diagnosis

We propose in this chapter a content-based informationretrieval framework to select documents in a database, consisting ofseveral images with semantic information. Information in these doc-uments is not only heterogeneous, but also often incomplete. So,themethod we propose may cover a wide range of applications. To se-lect the most relevant cases in the database, for a query, a degree ofmatch between two cases is defined for each case feature, and thesedegrees of match are fused. Two different fusion models are pro-posed: a Shafer’s model consisting of two hypotheses and a hybridDSm model consisting of N hypotheses, where N is the number ofcases in the database. They allow us to model our confidence ineach feature, and take it into account in the fusion process, to im-prove the retrieval performance. To include images in such a system,we characterize them by their digital content. The proposed meth-ods are applied to two multimodal medical databases for computeraided diagnosis; a comparison with other retrieval methods we pro-posed recently is provided. A mean precision at five of 81.8% and84.8% was obtained on a diabetic retinopathy and a mammographydatabase, respectively: the methods are precise enough to be used ina diagnosis aid system.