Feature Selection Guided by Perception in Medical CBIR Systems

This work aims at developing an efficient support to improve the precision of content-based medical image retrieval systems and also accelerate such retrieval, introducing a novel retrieval approach that integrates techniques of feature selection and relevance feedback to perform feature selection guided by perceptual similarity. Low-level features are commonly employed to represent the images by content. Feature selection is performed employing statistical association rules integrated with a relevance feedback process, tuning the mining process on the fly, according to the user´s perception. This integration not only improves the feature selection accuracy, but also allows personalising such process. The experiments performed show that the method improves up to 30% the query precision and decreases up to 11.6 times the number of features employed to compute the similarity in the content-based query, also decreasing the processing costs and memory requirements of the query execution.