Most Information retrieval systems are based on exact term matching. Though many researches address the "term mismatch" problem. This problem arises when different terms express the same meaning, in multilingual formulation of the query/documents, or when using expert terms either in the document or in the query. All these problems need a particular analysis that fills the gap between the document information and the user. In this work, we propose a solution by the enrichment of a meta-thesaurus. We propose to exploit the thesaurus relations between concepts but also to enrich them through the analysis of the terms associated to concepts. The matching is performed using concept derivation through a Bayesian network. A validation of our proposal is made on the test collection ImageCLEFMed 2005 and the resource UMLS 2005.