A neural network model to index medical images

With the development of computer capabilities, memories and network abilities, we need more efficient and robust algorithms to manage databases and to store and retrieve the relevant information for the user. The aim of this work is to automate the construction of a neural network Information Retrieval System (IRS) adapted to a medical image database. The user builds queries and the system must retrieve the relevant documents or images. Queries are groups of keywords or items associated with relevant images. In our approach, the set of queries and the binary relevance judgments on the documents constitute complex learning data associations. There are two phases in the automatic construction of the IRS. The indexing phase builds the learning data base and then a specific learning algorithm builds the neural network. For the system to be able to immediately learn these complex data, we have developed a new specific algorithm. It allows a perfect learning of a binary logical table in a stepwise fashion without forgetting the previously learnt logical combinations. Furthermore, this algorithm works very quickly and leads to a parallel implementation for large databases.