The Unified Medical Language System (UMLS) 1 offers the possibility to use annotated medical terms for Computer Aided Diagnoses System (CADS). We present a new semantic fusion system, based on UMLS. This fusion system has applications on a CADS that diagnoses neurodegenerative diseases. Since the UMLS Metathesaurus contains a huge amount of data, classification and extraction of the data we use is necessary. For this purpose, we use a feedforward neural network which is capable of training the negative patterns as well as the positive ones. At the semantic level we generate a three-layered network structure, which gives us the possibility of adding medical knowledge in order to cluster the data and prepare it for the fusion process.